diff --git a/GEMINI_PROMPTS.md b/GEMINI_PROMPTS.md
new file mode 100644
index 0000000000000000000000000000000000000000..674f7deff8f6e612dc7085f36c178791c7c5e3c7
--- /dev/null
+++ b/GEMINI_PROMPTS.md
@@ -0,0 +1,287 @@
+# Gemini prompts used for the failure labels
+
+These are the prompt templates exactly as they are stored in `labeler.prompts` of every `labels.json` (identical in all 399 files). `{...}` are Python `str.format` fields filled per episode. Model: `gemini-3.5-flash`; the video is the base-camera `rollout.mp4`.
+
+Notes on how they were applied to this real-robot run:
+
+- The templates were written for a simulator benchmark (LIBERO-X) and reused unchanged, which is why pass 1 says "in LIBERO-X" and mentions a simulator goal predicate. A real run has no simulator predicate: `{failing_predicate}` was filled with the task instruction itself and no simulator detail was available. The model is told the instruction and sees the video; it is not given any ground truth about why the episode failed beyond the human success label (failed episodes only are sent to passes 1-3).
+- The onset is the earliest point at which the mistake responsible for the failure becomes apparent. Pass 1 picks one onset type: `obvious_mistake` (an identifiable task-relevant mistake), `operator_intervention` (no single mistake, but a human operator would step in, e.g. stuck or ineffective repetition) or `timeout` (no earlier identifiable point). It is not the time at which the task becomes unrecoverable.
+- Pipeline: `scripts/semantic_failure` of the labelling project (two-pass onset localiser adapted from dtl184/liberox-evals). With the interactive backend the passes and the keyword phrases run in one shared video session per episode; with the Batch backend each is a separate request on the same uploaded video, with the same prompt text.
+
+## Pass 1 - failure localisation on the full video
+
+_one Gemini chat per episode; the whole rollout video is attached; failed episodes only._  Key in `labels.json`: `labeler.prompts.failure_coarse`.
+
+```text
+You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.
+
+The video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.
+
+Task instruction:
+"{task_desc}"
+
+The simulator reports that this goal predicate was NOT satisfied at the end of the episode:
+{failing_predicate}
+
+Additional simulator information:
+{detail}
+
+The simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.
+
+Your job is to determine:
+
+1. the primary visible failure mode,
+2. when the failure behavior begins,
+3. why the episode failed (a long reason AND a short keyword summary),
+4. a recovery action for after the failure has already happened (long AND short),
+5. a prevention action for before the failure has happened (long AND short).
+
+FAILURE ONSET DEFINITION
+
+Use exactly one of the following onset types:
+
+obvious_mistake:
+ Use this when there is an identifiable task-relevant mistake such as
+ dropping an object, knocking an object away, manipulating the wrong
+ object, making a failed placement, or otherwise performing an action
+ that contributes to the eventual failure.
+
+ The onset should be the EARLIEST point where the mistake responsible
+ for the eventual failure becomes apparent.
+
+operator_intervention:
+ Use this when there is not one discrete mistake, but the robot reaches
+ a point where a reasonable human operator would decide that the policy
+ needs assistance.
+
+ Examples include repeatedly attempting an ineffective action, becoming
+ stuck, moving erratically, or clearly ceasing to make useful progress.
+
+timeout:
+ Use this when there is no defensible earlier failure point. If the robot
+ continues behaving plausibly but simply does not finish before the time
+ limit, the failure onset is the END of the rollout.
+
+IMPORTANT RULES
+
+- Do NOT treat the goal predicate being false early in the rollout as failure.
+ Most goal predicates are naturally false until the robot completes the task.
+
+- Do NOT label a temporary mistake as failure onset if the robot later
+ successfully recovers from that mistake.
+
+- Use the earliest mistake or intervention point that actually explains the
+ eventual failed outcome.
+
+- Do not use hindsight to label normal task execution as failure merely because
+ you know the episode eventually fails.
+
+- If there is no identifiable earlier point, use onset_type="timeout" and set
+ failure_onset_seconds to approximately {duration_seconds:.2f}.
+
+FAILURE TAXONOMY
+
+Choose exactly one:
+
+{taxonomy}
+
+REASON, RECOVERY, PREVENTION — TWO VERSIONS EACH
+
+Every language field has a LONG form and a SHORT keyword summary.
+
+LONG: one or two sentences. Specific enough for a human to audit.
+
+SHORT: 2–8 words, semicolon-separated clauses, NOT a full sentence.
+ Same style as the 3-second action phrases. Good for later alignment.
+ Examples: "collision with object; dropping gripped object"
+
+failure_reason / failure_reason_summary:
+ Why this episode failed. Example long: the robot hit the basket while
+ carrying the moka pot, dropped it, and knocked the basket over.
+ Example short: "collision with object; dropping gripped object"
+
+RECOVERY (failure HAS already happened):
+ What the robot should do at or immediately after onset, now that the
+ scene is already broken. Identify object, action, and the condition to
+ restore before continuing.
+ Example long: move the basket away from the hot plate, then place the
+ moka pot on the hot-plate.
+ Example short: "clear basket; place moka on stove"
+
+PREVENTION (failure has NOT yet happened):
+ The counterfactual at or before onset: what the robot should have done
+ so the mistake never occurs. Do not describe post-drop cleanup.
+ Example long: place the moka pot on the stove; do not approach the box.
+ Example short: "place moka on stove; avoid box"
+
+Do not copy taxonomy names into these strings.
+
+Return ONLY a JSON object with this structure:
+
+{{
+ "failure_mode": " ",
+ "onset_type": "obvious_mistake|operator_intervention|timeout",
+ "failure_onset_seconds":,
+ "failure_window_start_seconds":,
+ "failure_window_end_seconds":,
+ "confidence": "high|medium|low",
+ "failure_reason": " ",
+ "failure_reason_summary": " ",
+ "recovery_action": " ",
+ "recovery_action_summary": " ",
+ "prevention_action": " ",
+ "prevention_action_summary": " ",
+ "justification": " "
+}}
+```
+
+## Pass 2 - temporal refinement
+
+_a magnified clip around the coarse onset (frames sampled at 1 FPS, +-3 s) is attached; failed episodes only._  Key in `labels.json`: `labeler.prompts.failure_refine`.
+
+```text
+You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.
+
+The complete rollout was already reviewed by another pass of the same VLM.
+
+The coarse analysis identified:
+
+Failure mode:
+{failure_mode}
+
+Failure onset type:
+{onset_type}
+
+Failure reason:
+{failure_reason}
+
+Coarse onset estimate in the original rollout:
+{coarse_seconds:.2f} seconds
+
+Task instruction:
+"{task_desc}"
+
+Failed simulator predicate:
+{failing_predicate}
+
+TEMPORAL MAPPING
+
+This clip has been deliberately slowed down.
+
+Each SECOND of this clip corresponds to exactly ONE FRAME from the original
+rollout.
+
+Therefore:
+
+ refined clip second 0 = original rollout frame {start_frame}
+ refined clip second 1 = original rollout frame {start_frame_plus_one}
+ refined clip second 2 = original rollout frame {start_frame_plus_two}
+ ...
+
+The clip contains {num_frames} original rollout frames.
+
+Your job is ONLY to refine the temporal location of the failure.
+
+Find the EARLIEST frame in this clip where the failure identified above becomes
+visibly apparent or where operator intervention becomes justified.
+
+Remember:
+
+- Do not select a temporary mistake if the robot recovers from it.
+- Do not select a frame merely because the task is not complete yet.
+- Select the first frame associated with the behavior responsible for the
+ eventual failure.
+- If the coarse event is not actually visible in this clip, return null.
+
+Return ONLY:
+
+{{
+ "refined_second":,
+ "confidence": "high|medium|low",
+ "justification": " "
+}}
+```
+
+## Pass 3 - failure language
+
+_same chat; onset, mode and taxonomy are fixed, the model only writes the reason / recovery / prevention text; failed episodes only._  Key in `labels.json`: `labeler.prompts.failure_language`.
+
+```text
+You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.
+
+Task instruction:
+"{task_desc}"
+
+Existing labels:
+- failure_mode: {failure_mode}
+- onset_type: {onset_type}
+- failure_onset_seconds: {onset_seconds}
+- failure_reason: {failure_reason}
+- recovery_action: {recovery_action}
+
+Fill the missing language fields.
+
+SHORT STYLE
+2–8 words, semicolon-separated clauses, not a full sentence.
+Example: "collision with object; dropping gripped object"
+
+RECOVERY vs PREVENTION
+- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).
+- prevention_action: long, what the robot should have done at or before onset so the failure never happens.
+- prevention_action_summary: keyword form of that prevention.
+
+Return ONLY:
+
+{{
+ "failure_reason_summary": " ",
+ "recovery_action_summary": " ",
+ "prevention_action": " ",
+ "prevention_action_summary": " "
+}}
+```
+
+## Semantic timeline - 3 second keyword phrases
+
+_same chat; every episode (successes and failures)._  Key in `labels.json`: `labeler.prompts.keyword_phrases`.
+
+```text
+You are reviewing the SAME complete robot manipulation rollout.
+
+The video is {hz:.0f} FPS. Each frame is exactly one executed robot action.
+Episode duration is {duration:.1f} seconds ({n_control} control steps).
+
+Task instruction:
+"{instruction}"
+{failure_context}
+For EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2–6 words)
+for what the manipulator is doing, like "reaching for mug" or "closing gripper"
+or "idle / no progress". Not a full sentence. Only what is visible in THAT window.
+
+Do not copy a failure mode, failure reason, recovery, or prevention into these phrases.
+Those are stored in a separate array. If a window happens to show a mistake,
+describe the visible motion (e.g. "dropping can"), not the taxonomy name.
+
+Windows:
+{windows}
+
+Return ONLY a JSON object:
+{{"segments": [{{"segment_index": 0, "phrase": "..."}}, ...]}}
+Include every segment_index listed above.
+```
+
+## Failure-mode taxonomy (inserted into the pass-1 prompt as {taxonomy})
+
+_the allowed values of `failure_mode`._  Key in `labels.json`: `labeler.prompts.failure_taxonomy`.
+
+```text
+- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.
+- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.
+- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).
+- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.
+- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).
+- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.
+- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).
+- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.
+- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.
+- other: the failure fits none of the labels above. Briefly explain what went wrong.
+```
diff --git a/LABELS.md b/LABELS.md
new file mode 100644
index 0000000000000000000000000000000000000000..f6fcf5e88b4c7a4e2538a3fb194f87beb108308f
--- /dev/null
+++ b/LABELS.md
@@ -0,0 +1,24 @@
+# Failure labels for the 399 rollouts
+
+Per-episode files next to `rollout.npz` / `rollout.mp4` in each `<task>/<ep>/` folder, plus `manifest.csv` (one row per episode with outcome and onset).
+
+| file | content |
+|---|---|
+| `labels.json` | `success` (human annotation), `semantic_timeline` (3 s keyword phrases per segment with control / policy step ranges), `vlm_failure` (failed episodes: mode, onset time and frame, reason / recovery / prevention text with short summaries, confidence, raw Gemini responses, token usage), `failure_annotation` (onset mapped to `control_step`, `policy_step`, `chunk_index`), `labeler` (model, backend, date, the prompts) |
+| `labels.npz` | the same fields as arrays: `fail_onset_frame`, `fail_onset_seconds`, `fail_mode`, `fail_reason`, `fail_recovery`, `fail_prevention` (+ `_summary`), `sem_t_start`, `sem_t_end`, `sem_control_start`, `sem_control_end`, `sem_phrase` |
+| `example.md` | a readable rendering of the above for that episode |
+
+## Contents
+
+- 399 episodes: 145 successes (no onset), 254 failures with a failure onset.
+- Failure modes (Gemini, one per failed episode): grasp_failure 144, stuck_or_no_progress 31, placement_or_insertion_failure 29, object_displacement 24, open_close_failure 19, timeout_or_insufficient_progress 4, wrong_object 3.
+- Onset types: obvious_mistake 223, operator_intervention 27, timeout 4.
+- Per task (episodes / failures): basket 50/42, blocks 50/44, closer_drawer 50/22, eggplant 50/25, mushroom 49/36, pineapple_out_of_drawer 50/38, push_pineapple 50/23, tomato 50/24.
+
+## Provenance and caveats
+
+- Labeller: Gemini (`gemini-3.5-flash`), prompts in `GEMINI_PROMPTS.md`. 399 episode(s) via `GeminiBackend` (labelled 2026-10-20 to 2026-10-22).
+- The labels are model outputs and have not been individually verified by a person; the success flag comes from the operator. Onset times are the first visible mistake.
+- `vlm_failure_onset_frame` is a control-step index at the recorded control rate (about 50 Hz); the policy step is stored in `failure_annotation.failure_policy_step` and should be used instead of dividing the frame by `replan_steps` (the robot executes about 5.8 control steps per policy call on average, not the nominal 5 or 6).
+- A few onsets are at the very start or very end of a video (`operator_intervention` at the first frames, `timeout` in the last seconds); these are worth a manual look before using them as supervision.
+- Successes carry a keyword-phrase timeline but no failure fields.
diff --git a/basket/ep000/example.md b/basket/ep000/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..3f4ec5c4761d11428bd49ef9f401b76ad0089e46
--- /dev/null
+++ b/basket/ep000/example.md
@@ -0,0 +1,33 @@
+# t002__ep000__basket
+
+- **suite / variant:** `` / ``
+- **instruction:** Flip the basket upright
+- **success:** False
+- **control steps:** 430  (8.4s at 51 Hz)
+- **policy inferences:** 74  (replan_steps=6)
+- **video:** `basket/ep000/rollout.mp4`
+- **features:** `74` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–152 | 0–25 | approaching and grasping basket handle |
+| 3.0–6.0 | 153–305 | 25–50 | pulling handle and losing grip |
+| 6.0–8.4 | 306–429 | 51–71 | retracting gripper and remaining idle |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 3.1434184675834973 s (00:03.1)
+- **frame / control_step:** 160 (policy_step=26, chunk_index=4)
+- **coarse onset:** 3.8 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to flip the basket by grabbing its wire handle, but lost its grip while lifting, causing the basket to fall back down upside down.
+  Summary: lost grasp on handle; basket remained upside down
+- **recovery:** Re-align the end-effector over the basket handle, establish a firm grip, and lift upward while rotating to flip the basket upright.
+  Summary: re-grasp handle securely; flip basket upright
+- **prevention:** Clamp down tightly on the basket handle and maintain a firm hold while pulling up and over to ensure the basket flips completely.
+  Summary: secure firm grasp; rotate handle to flip basket
+- **justification:** None
diff --git a/basket/ep000/labels.json b/basket/ep000/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..d4cfd2ecf8a949c2470bb05d3e3eaad783160b08
--- /dev/null
+++ b/basket/ep000/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t002__ep000__basket",
+  "scene_variant": "",
+  "task_id": 2,
+  "task_file": "",
+  "instruction": "Flip the basket upright",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-20T21:43:22+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.005893909626719,
+      "control_step_start": 0,
+      "control_step_end": 152,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "approaching and grasping basket handle",
+      "description": "approaching and grasping basket handle"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.005893909626719,
+      "t_end_sec": 6.011787819253438,
+      "control_step_start": 153,
+      "control_step_end": 305,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "pulling handle and losing grip",
+      "description": "pulling handle and losing grip"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.011787819253438,
+      "t_end_sec": 8.447937131630649,
+      "control_step_start": 306,
+      "control_step_end": 429,
+      "policy_step_start": 51,
+      "policy_step_end": 71,
+      "phrase": "retracting gripper and remaining idle",
+      "description": "retracting gripper and remaining idle"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 3.1434184675834973,
+    "vlm_failure_onset_timestamp": "00:03.1",
+    "vlm_failure_onset_frame": 160,
+    "vlm_failure_onset_step": 160,
+    "vlm_coarse_failure_onset_seconds": 3.8,
+    "vlm_failure_window_start_seconds": 3.2,
+    "vlm_failure_window_end_seconds": 4.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to flip the basket by grabbing its wire handle, but lost its grip while lifting, causing the basket to fall back down upside down.",
+    "vlm_failure_reason_summary": "lost grasp on handle; basket remained upside down",
+    "vlm_recovery_action": "Re-align the end-effector over the basket handle, establish a firm grip, and lift upward while rotating to flip the basket upright.",
+    "vlm_recovery_action_summary": "re-grasp handle securely; flip basket upright",
+    "vlm_prevention_action": "Clamp down tightly on the basket handle and maintain a firm hold while pulling up and over to ensure the basket flips completely.",
+    "vlm_prevention_action_summary": "secure firm grasp; rotate handle to flip basket",
+    "vlm_justification": null,
+    "vlm_temporal_justification": "At timestamp 02:00 (120 seconds), as the robot arm lifts and rotates to flip the basket, the wire handle slips out of the gripper's fingers, causing the basket to fall back onto the table upside down.",
+    "vlm_video_fps": 50.9,
+    "vlm_video_num_frames": 430,
+    "vlm_video_duration_seconds": 8.428290766208251,
+    "vlm_usage": {
+      "input_tokens": 21692,
+      "output_tokens": 325,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 3.8,\n \"failure_window_start_seconds\": 3.2,\n \"failure_window_end_seconds\": 4.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to flip the basket by grabbing its wire handle, but lost its grip while lifting, causing the basket to fall back down upside down.\",\n \"failure_reason_summary\": \"lost grasp on handle; basket remained upside down\",\n \"recovery_action\": \"Re-align the end-effector over the basket handle, establish a firm grip, and lift upward while rotating to flip the basket upright.\",\n \"recovery_action_summary\": \"re-grasp handle securely; flip basket upright\",\n \"prevention_action\": \"Clamp down tightly on the basket handle and maintain a firm hold while pulling up and over to ensure the basket flips completely.\",\n \"prevention_action_summary\": \"secure firm grasp; rotate handle to flip basket\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 160,
+    "failure_sim_step": 160,
+    "failure_policy_step": 26,
+    "failure_chunk_index": 4,
+    "first_post_failure_policy_step": 27,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-align the end-effector over the basket handle, establish a firm grip, and lift upward while rotating to flip the basket upright."
+  }
+}
\ No newline at end of file
diff --git a/basket/ep000/labels.npz b/basket/ep000/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..24c9fe3f251340fdddf5e414b98a07fc1083aac4
--- /dev/null
+++ b/basket/ep000/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:df71c0f780462ee77b9d4ed9e86399d431d3a2b07da16c4406a4aa0f9767c00b
+size 3839
diff --git a/basket/ep001/example.md b/basket/ep001/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..a5131e6dcf1dc2ec2707e5c2c7ec0482a2fb40fc
--- /dev/null
+++ b/basket/ep001/example.md
@@ -0,0 +1,33 @@
+# t002__ep001__basket
+
+- **suite / variant:** `` / ``
+- **instruction:** Flip the basket upright
+- **success:** False
+- **control steps:** 388  (7.8s at 50 Hz)
+- **policy inferences:** 66  (replan_steps=6)
+- **video:** `basket/ep001/rollout.mp4`
+- **features:** `66` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–148 | 0–24 | approaching basket handle |
+| 3.0–6.0 | 149–297 | 24–49 | grasping handle; attempting to flip |
+| 6.0–7.8 | 298–387 | 49–64 | losing grip; retracting upward |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 5.70281124497992 s (00:05.7)
+- **frame / control_step:** 284 (policy_step=47, chunk_index=2)
+- **coarse onset:** 5.2 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the basket handle to flip it, but the gripper failed to establish a secure hold and slipped off, leaving the basket in its initial position as the arm retracted.
+  Summary: failed handle grasp; gripper slipped off basket
+- **recovery:** Re-align the gripper over the basket handle, establish a secure grip, and lift while rotating to flip the basket upright.
+  Summary: re-align gripper; grasp handle firmly; flip basket upright
+- **prevention:** Position the gripper precisely over the basket handle before closing to ensure a secure hold prior to lifting and flipping.
+  Summary: center gripper on handle; secure grip before flipping
+- **justification:** At around 5.2 seconds, the robot closes its gripper on the basket handle but fails to achieve a stable grip. At 6.0 seconds, the arm lifts up empty while the basket remains unmoved on the table.
diff --git a/basket/ep001/labels.json b/basket/ep001/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..1e1f6acc741867259ba2d52e8a6c8b7d66148d28
--- /dev/null
+++ b/basket/ep001/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t002__ep001__basket",
+  "scene_variant": "",
+  "task_id": 2,
+  "task_file": "",
+  "instruction": "Flip the basket upright",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-20T21:44:53+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.991967871485944,
+      "control_step_start": 0,
+      "control_step_end": 148,
+      "policy_step_start": 0,
+      "policy_step_end": 24,
+      "phrase": "approaching basket handle",
+      "description": "approaching basket handle"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.991967871485944,
+      "t_end_sec": 5.983935742971888,
+      "control_step_start": 149,
+      "control_step_end": 297,
+      "policy_step_start": 24,
+      "policy_step_end": 49,
+      "phrase": "grasping handle; attempting to flip",
+      "description": "grasping handle; attempting to flip"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.983935742971888,
+      "t_end_sec": 7.791164658634538,
+      "control_step_start": 298,
+      "control_step_end": 387,
+      "policy_step_start": 49,
+      "policy_step_end": 64,
+      "phrase": "losing grip; retracting upward",
+      "description": "losing grip; retracting upward"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 5.70281124497992,
+    "vlm_failure_onset_timestamp": "00:05.7",
+    "vlm_failure_onset_frame": 284,
+    "vlm_failure_onset_step": 284,
+    "vlm_coarse_failure_onset_seconds": 5.2,
+    "vlm_failure_window_start_seconds": 5.2,
+    "vlm_failure_window_end_seconds": 6.2,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the basket handle to flip it, but the gripper failed to establish a secure hold and slipped off, leaving the basket in its initial position as the arm retracted.",
+    "vlm_failure_reason_summary": "failed handle grasp; gripper slipped off basket",
+    "vlm_recovery_action": "Re-align the gripper over the basket handle, establish a secure grip, and lift while rotating to flip the basket upright.",
+    "vlm_recovery_action_summary": "re-align gripper; grasp handle firmly; flip basket upright",
+    "vlm_prevention_action": "Position the gripper precisely over the basket handle before closing to ensure a secure hold prior to lifting and flipping.",
+    "vlm_prevention_action_summary": "center gripper on handle; secure grip before flipping",
+    "vlm_justification": "At around 5.2 seconds, the robot closes its gripper on the basket handle but fails to achieve a stable grip. At 6.0 seconds, the arm lifts up empty while the basket remains unmoved on the table.",
+    "vlm_temporal_justification": "The robot closes its gripper at the handle and begins lifting upward around 02:50. By 02:54 (frame 284 of the rollout), it is visibly apparent that the gripper failed to establish a hold on the handle, lifting away while leaving the basket completely stationary.",
+    "vlm_video_fps": 49.8,
+    "vlm_video_num_frames": 388,
+    "vlm_video_duration_seconds": 7.771084337349398,
+    "vlm_usage": {
+      "input_tokens": 19875,
+      "output_tokens": 394,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 5.2,\n \"failure_window_start_seconds\": 5.2,\n \"failure_window_end_seconds\": 6.2,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the basket handle to flip it, but the gripper failed to establish a secure hold and slipped off, leaving the basket in its initial position as the arm retracted.\",\n \"failure_reason_summary\": \"failed handle grasp; gripper slipped off basket\",\n \"recovery_action\": \"Re-align the gripper over the basket handle, establish a secure grip, and lift while rotating to flip the basket upright.\",\n \"recovery_action_summary\": \"re-align gripper; grasp handle firmly; flip basket upright\",\n \"prevention_action\": \"Position the gripper precisely over the basket handle before closing to ensure a secure hold prior to lifting and flipping.\",\n \"prevention_action_summary\": \"center gripper on handle; secure grip before flipping\",\n \"justification\": \"At around 5.2 seconds, the robot closes its gripper on the basket handle but fails to achieve a stable grip. At 6.0 seconds, the arm lifts up empty while the basket remains unmoved on the table.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 284,
+    "failure_sim_step": 284,
+    "failure_policy_step": 47,
+    "failure_chunk_index": 2,
+    "first_post_failure_policy_step": 48,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-align the gripper over the basket handle, establish a secure grip, and lift while rotating to flip the basket upright."
+  }
+}
\ No newline at end of file
diff --git a/basket/ep001/labels.npz b/basket/ep001/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..50a4ec7b68095576562be577f550f2a0449fa48a
--- /dev/null
+++ b/basket/ep001/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:72be09e4401c14b731b30b53a72ac066426fa90460f605a1b61f0da80789b25e
+size 3841
diff --git a/basket/ep002/example.md b/basket/ep002/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..caa9b6ee51831c66c6dfb7ea58d214f2848d2547
--- /dev/null
+++ b/basket/ep002/example.md
@@ -0,0 +1,33 @@
+# t002__ep002__basket
+
+- **suite / variant:** `` / ``
+- **instruction:** Flip the basket upright
+- **success:** False
+- **control steps:** 378  (7.4s at 51 Hz)
+- **policy inferences:** 65  (replan_steps=6)
+- **video:** `basket/ep002/rollout.mp4`
+- **features:** `65` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–152 | 0–25 | reaching toward red basket |
+| 3.0–6.0 | 153–305 | 25–50 | grasping basket; slipping off handle |
+| 6.0–7.4 | 306–377 | 51–62 | hovering arm above table |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 4.52054794520548 s (00:04.5)
+- **frame / control_step:** 231 (policy_step=38, chunk_index=3)
+- **coarse onset:** 4.0 s (refined=True)
+- **mode:** object_displacement
+- **reason:** The robot attempted to flip the basket by grabbing its handle, but lost its grip mid-motion, dropping the basket onto its side without orienting it fully upright.
+  Summary: losing grip on basket rim; leaving basket tilted
+- **recovery:** Re-grasp the edge or handle of the basket, lift it, and rotate it until the opening faces upward and the base rests flat on the table.
+  Summary: re-grasp basket handle; rotate until fully upright
+- **prevention:** Establish a deep and secure grasp on the basket handle before executing the upward rotation to avoid slipping.
+  Summary: secure firm grasp on rim; complete rotation fully
+- **justification:** At 4.0 seconds, the robot gripper slips off the basket handle during the flip attempt, dropping the basket onto its side. The robot then retracts its end effector, leaving the basket incompletely flipped for the remainder of the rollout.
diff --git a/basket/ep002/labels.json b/basket/ep002/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..a7d994ce695cd27f30972045b51e3f25f9a5d784
--- /dev/null
+++ b/basket/ep002/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t002__ep002__basket",
+  "scene_variant": "",
+  "task_id": 2,
+  "task_file": "",
+  "instruction": "Flip the basket upright",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-20T21:46:33+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.99412915851272,
+      "control_step_start": 0,
+      "control_step_end": 152,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching toward red basket",
+      "description": "reaching toward red basket"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.99412915851272,
+      "t_end_sec": 5.98825831702544,
+      "control_step_start": 153,
+      "control_step_end": 305,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "grasping basket; slipping off handle",
+      "description": "grasping basket; slipping off handle"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.98825831702544,
+      "t_end_sec": 7.397260273972602,
+      "control_step_start": 306,
+      "control_step_end": 377,
+      "policy_step_start": 51,
+      "policy_step_end": 62,
+      "phrase": "hovering arm above table",
+      "description": "hovering arm above table"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "object_displacement",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 4.52054794520548,
+    "vlm_failure_onset_timestamp": "00:04.5",
+    "vlm_failure_onset_frame": 231,
+    "vlm_failure_onset_step": 231,
+    "vlm_coarse_failure_onset_seconds": 4.0,
+    "vlm_failure_window_start_seconds": 4.0,
+    "vlm_failure_window_end_seconds": 7.38,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to flip the basket by grabbing its handle, but lost its grip mid-motion, dropping the basket onto its side without orienting it fully upright.",
+    "vlm_failure_reason_summary": "losing grip on basket rim; leaving basket tilted",
+    "vlm_recovery_action": "Re-grasp the edge or handle of the basket, lift it, and rotate it until the opening faces upward and the base rests flat on the table.",
+    "vlm_recovery_action_summary": "re-grasp basket handle; rotate until fully upright",
+    "vlm_prevention_action": "Establish a deep and secure grasp on the basket handle before executing the upward rotation to avoid slipping.",
+    "vlm_prevention_action_summary": "secure firm grasp on rim; complete rotation fully",
+    "vlm_justification": "At 4.0 seconds, the robot gripper slips off the basket handle during the flip attempt, dropping the basket onto its side. The robot then retracts its end effector, leaving the basket incompletely flipped for the remainder of the rollout.",
+    "vlm_temporal_justification": "At second 180 (03:00 in the video overlay), the robot gripper loses its grip on the basket's handle during the flip attempt, dropping the basket back down onto the surface without orienting it fully upright.",
+    "vlm_video_fps": 51.1,
+    "vlm_video_num_frames": 378,
+    "vlm_video_duration_seconds": 7.3776908023483365,
+    "vlm_usage": {
+      "input_tokens": 21632,
+      "output_tokens": 384,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"object_displacement\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 4.0,\n \"failure_window_start_seconds\": 4.0,\n \"failure_window_end_seconds\": 7.38,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to flip the basket by grabbing its handle, but lost its grip mid-motion, dropping the basket onto its side without orienting it fully upright.\",\n \"failure_reason_summary\": \"losing grip on basket rim; leaving basket tilted\",\n \"recovery_action\": \"Re-grasp the edge or handle of the basket, lift it, and rotate it until the opening faces upward and the base rests flat on the table.\",\n \"recovery_action_summary\": \"re-grasp basket handle; rotate until fully upright\",\n \"prevention_action\": \"Establish a deep and secure grasp on the basket handle before executing the upward rotation to avoid slipping.\",\n \"prevention_action_summary\": \"secure firm grasp on rim; complete rotation fully\",\n \"justification\": \"At 4.0 seconds, the robot gripper slips off the basket handle during the flip attempt, dropping the basket onto its side. The robot then retracts its end effector, leaving the basket incompletely flipped for the remainder of the rollout.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 231,
+    "failure_sim_step": 231,
+    "failure_policy_step": 38,
+    "failure_chunk_index": 3,
+    "first_post_failure_policy_step": 39,
+    "failure_type": "object_displacement",
+    "correction_action": "Re-grasp the edge or handle of the basket, lift it, and rotate it until the opening faces upward and the base rests flat on the table."
+  }
+}
\ No newline at end of file
diff --git a/basket/ep002/labels.npz b/basket/ep002/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..7a42f3e456bf964dbea8dfd219cec4943d180e6b
--- /dev/null
+++ b/basket/ep002/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ac136450fa6768ecf8a5ddd7c15336c388822e6ecd571d77dc7cc508b10237c9
+size 3826
diff --git a/basket/ep003/example.md b/basket/ep003/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..2c065b51c5b7c72dd8537644b7e85e1a49891a1a
--- /dev/null
+++ b/basket/ep003/example.md
@@ -0,0 +1,33 @@
+# t002__ep003__basket
+
+- **suite / variant:** `` / ``
+- **instruction:** Flip the basket upright
+- **success:** False
+- **control steps:** 311  (6.1s at 51 Hz)
+- **policy inferences:** 54  (replan_steps=6)
+- **video:** `basket/ep003/rollout.mp4`
+- **features:** `54` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | approaching and grasping basket edge |
+| 3.0–6.0 | 152–303 | 25–50 | slipping off basket and retracting |
+| 6.0–6.1 | 304–310 | 50–51 | idle hovering above basket |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 3.201581027667984 s (00:03.2)
+- **frame / control_step:** 162 (policy_step=27, chunk_index=0)
+- **coarse onset:** 3.0 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the rim of the red basket to flip it, but failed to secure a firm grip, slipping off and retreating without flipping the basket.
+  Summary: failed grasp on basket edge; slipping off object
+- **recovery:** Re-align the gripper with the edge of the red basket, establish a secure grasp on the rim, and pull upward to rotate the basket upright.
+  Summary: re-align gripper; secure basket edge; flip upright
+- **prevention:** Position the gripper precisely over the edge of the red basket and close the fingers tightly before initiating the lifting/flipping motion.
+  Summary: position gripper securely; grasp rim firmly; flip basket
+- **justification:** Around 3.0 seconds into the rollout, the robot closes its gripper on the top rim of the red basket, but slips off almost immediately due to a poor grasp angle. It then retracts upward and hovers for the rest of the episode, failing the task.
diff --git a/basket/ep003/labels.json b/basket/ep003/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..3c7c3c4ba83415429ad66ba4a6ced406f41c437a
--- /dev/null
+++ b/basket/ep003/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t002__ep003__basket",
+  "scene_variant": "",
+  "task_id": 2,
+  "task_file": "",
+  "instruction": "Flip the basket upright",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-20T21:48:42+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0039525691699605,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "approaching and grasping basket edge",
+      "description": "approaching and grasping basket edge"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0039525691699605,
+      "t_end_sec": 6.007905138339921,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "slipping off basket and retracting",
+      "description": "slipping off basket and retracting"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.007905138339921,
+      "t_end_sec": 6.146245059288537,
+      "control_step_start": 304,
+      "control_step_end": 310,
+      "policy_step_start": 50,
+      "policy_step_end": 51,
+      "phrase": "idle hovering above basket",
+      "description": "idle hovering above basket"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 3.201581027667984,
+    "vlm_failure_onset_timestamp": "00:03.2",
+    "vlm_failure_onset_frame": 162,
+    "vlm_failure_onset_step": 162,
+    "vlm_coarse_failure_onset_seconds": 3.0,
+    "vlm_failure_window_start_seconds": 2.5,
+    "vlm_failure_window_end_seconds": 3.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the rim of the red basket to flip it, but failed to secure a firm grip, slipping off and retreating without flipping the basket.",
+    "vlm_failure_reason_summary": "failed grasp on basket edge; slipping off object",
+    "vlm_recovery_action": "Re-align the gripper with the edge of the red basket, establish a secure grasp on the rim, and pull upward to rotate the basket upright.",
+    "vlm_recovery_action_summary": "re-align gripper; secure basket edge; flip upright",
+    "vlm_prevention_action": "Position the gripper precisely over the edge of the red basket and close the fingers tightly before initiating the lifting/flipping motion.",
+    "vlm_prevention_action_summary": "position gripper securely; grasp rim firmly; flip basket",
+    "vlm_justification": "Around 3.0 seconds into the rollout, the robot closes its gripper on the top rim of the red basket, but slips off almost immediately due to a poor grasp angle. It then retracts upward and hovers for the rest of the episode, failing the task.",
+    "vlm_temporal_justification": "At timestamp 02:42 (frame 162), the robot lowers its gripper onto the top surface of the basket and closes its fingers in an improper position, failing to clamp onto the rim or edge required to flip the basket.",
+    "vlm_video_fps": 50.6,
+    "vlm_video_num_frames": 311,
+    "vlm_video_duration_seconds": 6.126482213438735,
+    "vlm_usage": {
+      "input_tokens": 21439,
+      "output_tokens": 394,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 3.0,\n \"failure_window_start_seconds\": 2.5,\n \"failure_window_end_seconds\": 3.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the rim of the red basket to flip it, but failed to secure a firm grip, slipping off and retreating without flipping the basket.\",\n \"failure_reason_summary\": \"failed grasp on basket edge; slipping off object\",\n \"recovery_action\": \"Re-align the gripper with the edge of the red basket, establish a secure grasp on the rim, and pull upward to rotate the basket upright.\",\n \"recovery_action_summary\": \"re-align gripper; secure basket edge; flip upright\",\n \"prevention_action\": \"Position the gripper precisely over the edge of the red basket and close the fingers tightly before initiating the lifting/flipping motion.\",\n \"prevention_action_summary\": \"position gripper securely; grasp rim firmly; flip basket\",\n \"justification\": \"Around 3.0 seconds into the rollout, the robot closes its gripper on the top rim of the red basket, but slips off almost immediately due to a poor grasp angle. It then retracts upward and hovers for the rest of the episode, failing the task.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 162,
+    "failure_sim_step": 162,
+    "failure_policy_step": 27,
+    "failure_chunk_index": 0,
+    "first_post_failure_policy_step": 28,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-align the gripper with the edge of the red basket, establish a secure grasp on the rim, and pull upward to rotate the basket upright."
+  }
+}
\ No newline at end of file
diff --git a/basket/ep003/labels.npz b/basket/ep003/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..2d69708ec50d508ae287f3c1d9f8708400ab9741
--- /dev/null
+++ b/basket/ep003/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:07c248f6a33b5a882fcd8a812f016e274fb0f17a9241e02ef0cb32a84deb150d
+size 3832
diff --git a/basket/ep004/example.md b/basket/ep004/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..af4dbc6d1777e1e600448ae181b8bd6911a1d514
--- /dev/null
+++ b/basket/ep004/example.md
@@ -0,0 +1,34 @@
+# t002__ep004__basket
+
+- **suite / variant:** `` / ``
+- **instruction:** Flip the basket upright
+- **success:** False
+- **control steps:** 480  (9.4s at 51 Hz)
+- **policy inferences:** 82  (replan_steps=6)
+- **video:** `basket/ep004/rollout.mp4`
+- **features:** `82` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–152 | 0–25 | approaching and grasping basket |
+| 3.0–6.0 | 153–305 | 25–50 | lifting and placing basket down |
+| 6.0–9.0 | 306–458 | 51–76 | retracting gripper away |
+| 9.0–9.4 | 459–479 | 76–79 | remaining idle above table |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 3.9411764705882355 s (00:03.9)
+- **frame / control_step:** 201 (policy_step=33, chunk_index=3)
+- **coarse onset:** 3.8 s (refined=True)
+- **mode:** placement_or_insertion_failure
+- **reason:** The robot lifts the basket by its handle but fails to rotate it into an upright orientation before setting it back down on its side.
+  Summary: failed to rotate basket; incorrect placement orientation
+- **recovery:** Grasp the basket again, rotate it 90 degrees so its bottom rests flat on the table, and release it in the upright position.
+  Summary: re-grasp basket; rotate upright; place flat on table
+- **prevention:** Rotate the wrist joint sufficiently while lifting the basket to orientation-flip it upright before setting it down on the table surface.
+  Summary: rotate gripper to flip basket upright; verify orientation before releasing
+- **justification:** None
diff --git a/basket/ep004/labels.json b/basket/ep004/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..fcf6e035b43340f7368c8b66a3d3bae425111fbf
--- /dev/null
+++ b/basket/ep004/labels.json
@@ -0,0 +1,109 @@
+{
+  "rollout_id": "t002__ep004__basket",
+  "scene_variant": "",
+  "task_id": 2,
+  "task_file": "",
+  "instruction": "Flip the basket upright",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T19:31:20+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0,
+      "control_step_start": 0,
+      "control_step_end": 152,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "approaching and grasping basket",
+      "description": "approaching and grasping basket"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0,
+      "t_end_sec": 6.0,
+      "control_step_start": 153,
+      "control_step_end": 305,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "lifting and placing basket down",
+      "description": "lifting and placing basket down"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.0,
+      "t_end_sec": 9.0,
+      "control_step_start": 306,
+      "control_step_end": 458,
+      "policy_step_start": 51,
+      "policy_step_end": 76,
+      "phrase": "retracting gripper away",
+      "description": "retracting gripper away"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.0,
+      "t_end_sec": 9.411764705882353,
+      "control_step_start": 459,
+      "control_step_end": 479,
+      "policy_step_start": 76,
+      "policy_step_end": 79,
+      "phrase": "remaining idle above table",
+      "description": "remaining idle above table"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "placement_or_insertion_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 3.9411764705882355,
+    "vlm_failure_onset_timestamp": "00:03.9",
+    "vlm_failure_onset_frame": 201,
+    "vlm_failure_onset_step": 201,
+    "vlm_coarse_failure_onset_seconds": 3.8,
+    "vlm_failure_window_start_seconds": 3.0,
+    "vlm_failure_window_end_seconds": 4.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot lifts the basket by its handle but fails to rotate it into an upright orientation before setting it back down on its side.",
+    "vlm_failure_reason_summary": "failed to rotate basket; incorrect placement orientation",
+    "vlm_recovery_action": "Grasp the basket again, rotate it 90 degrees so its bottom rests flat on the table, and release it in the upright position.",
+    "vlm_recovery_action_summary": "re-grasp basket; rotate upright; place flat on table",
+    "vlm_prevention_action": "Rotate the wrist joint sufficiently while lifting the basket to orientation-flip it upright before setting it down on the table surface.",
+    "vlm_prevention_action_summary": "rotate gripper to flip basket upright; verify orientation before releasing",
+    "vlm_justification": null,
+    "vlm_temporal_justification": "The robot lifts the basket by its handle to peak height, but at second 160 it begins descending back towards the table surface without attempting to rotate or flip the basket into an upright orientation.",
+    "vlm_video_fps": 51.0,
+    "vlm_video_num_frames": 480,
+    "vlm_video_duration_seconds": 9.392156862745098,
+    "vlm_usage": {
+      "input_tokens": 21755,
+      "output_tokens": 321,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"placement_or_insertion_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 3.8,\n \"failure_window_start_seconds\": 3.0,\n \"failure_window_end_seconds\": 4.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot lifts the basket by its handle but fails to rotate it into an upright orientation before setting it back down on its side.\",\n \"failure_reason_summary\": \"failed to rotate basket; incorrect placement orientation\",\n \"recovery_action\": \"Grasp the basket again, rotate it 90 degrees so its bottom rests flat on the table, and release it in the upright position.\",\n \"recovery_action_summary\": \"re-grasp basket; rotate upright; place flat on table\",\n \"prevention_action\": \"Rotate the wrist joint sufficiently while lifting the basket to orientation-flip it upright before setting it down on the table surface.\",\n \"prevention_action_summary\": \"rotate gripper to flip basket upright; verify orientation before releasing\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 201,
+    "failure_sim_step": 201,
+    "failure_policy_step": 33,
+    "failure_chunk_index": 3,
+    "first_post_failure_policy_step": 34,
+    "failure_type": "placement_or_insertion_failure",
+    "correction_action": "Grasp the basket again, rotate it 90 degrees so its bottom rests flat on the table, and release it in the upright position."
+  }
+}
\ No newline at end of file
diff --git a/basket/ep004/labels.npz b/basket/ep004/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..dea9a19d4b66d13ba80b4a00b0de344923a445b4
--- /dev/null
+++ b/basket/ep004/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cf3cb55a2743d032de650f7be061ddc1957a1450fe4b5afe5ae5e55fa9f918de
+size 3911
diff --git a/basket/ep005/example.md b/basket/ep005/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..f11b5270b73d4bf7395bee7b9e46854747d8317b
--- /dev/null
+++ b/basket/ep005/example.md
@@ -0,0 +1,33 @@
+# t002__ep005__basket
+
+- **suite / variant:** `` / ``
+- **instruction:** Flip the basket upright
+- **success:** False
+- **control steps:** 334  (6.7s at 50 Hz)
+- **policy inferences:** 58  (replan_steps=6)
+- **video:** `basket/ep005/rollout.mp4`
+- **features:** `58` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | approaching basket and slipping |
+| 3.0–6.0 | 151–301 | 25–50 | attempting second grasp on basket |
+| 6.0–6.7 | 302–333 | 50–55 | retracting end effector upward |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 2.549800796812749 s (00:02.5)
+- **frame / control_step:** 128 (policy_step=21, chunk_index=2)
+- **coarse onset:** 2.0 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the edge or handle of the upside-down basket to flip it, but failed to secure a solid grip and slipped off without moving the basket upright.
+  Summary: failed grasp on basket; slipped off handle
+- **recovery:** Reposition the end-effector directly over the basket edge or handle, secure a tight grip, and rotate the wrist to flip the basket upright.
+  Summary: reposition gripper; secure grip on edge; flip basket upright
+- **prevention:** Align the gripper precisely with the basket's rim or handle before closing the fingers to establish a stable grasp prior to applying the flipping rotation.
+  Summary: align gripper with rim; secure firm grip; rotate to flip
+- **justification:** At approximately 2.0 seconds, the robot's gripper contacts the basket but fails to grasp or hook it securely, slipping away without flipping the basket. It repeats this unsuccessful maneuver around 5.0 seconds.
diff --git a/basket/ep005/labels.json b/basket/ep005/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..203421feb0416e25b109bbfe300d49d023a751ba
--- /dev/null
+++ b/basket/ep005/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t002__ep005__basket",
+  "scene_variant": "",
+  "task_id": 2,
+  "task_file": "",
+  "instruction": "Flip the basket upright",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T19:31:51+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00796812749004,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "approaching basket and slipping",
+      "description": "approaching basket and slipping"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00796812749004,
+      "t_end_sec": 6.01593625498008,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "attempting second grasp on basket",
+      "description": "attempting second grasp on basket"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01593625498008,
+      "t_end_sec": 6.653386454183266,
+      "control_step_start": 302,
+      "control_step_end": 333,
+      "policy_step_start": 50,
+      "policy_step_end": 55,
+      "phrase": "retracting end effector upward",
+      "description": "retracting end effector upward"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 2.549800796812749,
+    "vlm_failure_onset_timestamp": "00:02.5",
+    "vlm_failure_onset_frame": 128,
+    "vlm_failure_onset_step": 128,
+    "vlm_coarse_failure_onset_seconds": 2.0,
+    "vlm_failure_window_start_seconds": 1.8,
+    "vlm_failure_window_end_seconds": 2.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the edge or handle of the upside-down basket to flip it, but failed to secure a solid grip and slipped off without moving the basket upright.",
+    "vlm_failure_reason_summary": "failed grasp on basket; slipped off handle",
+    "vlm_recovery_action": "Reposition the end-effector directly over the basket edge or handle, secure a tight grip, and rotate the wrist to flip the basket upright.",
+    "vlm_recovery_action_summary": "reposition gripper; secure grip on edge; flip basket upright",
+    "vlm_prevention_action": "Align the gripper precisely with the basket's rim or handle before closing the fingers to establish a stable grasp prior to applying the flipping rotation.",
+    "vlm_prevention_action_summary": "align gripper with rim; secure firm grip; rotate to flip",
+    "vlm_justification": "At approximately 2.0 seconds, the robot's gripper contacts the basket but fails to grasp or hook it securely, slipping away without flipping the basket. It repeats this unsuccessful maneuver around 5.0 seconds.",
+    "vlm_temporal_justification": "At 02:08 (128 seconds), as the robotic arm pulls upward to flip the basket, the gripper fingers lose their hold on the handle and slip off completely, leaving the basket in its upside-down position.",
+    "vlm_video_fps": 50.2,
+    "vlm_video_num_frames": 334,
+    "vlm_video_duration_seconds": 6.633466135458167,
+    "vlm_usage": {
+      "input_tokens": 18166,
+      "output_tokens": 389,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 2.0,\n \"failure_window_start_seconds\": 1.8,\n \"failure_window_end_seconds\": 2.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the edge or handle of the upside-down basket to flip it, but failed to secure a solid grip and slipped off without moving the basket upright.\",\n \"failure_reason_summary\": \"failed grasp on basket; slipped off handle\",\n \"recovery_action\": \"Reposition the end-effector directly over the basket edge or handle, secure a tight grip, and rotate the wrist to flip the basket upright.\",\n \"recovery_action_summary\": \"reposition gripper; secure grip on edge; flip basket upright\",\n \"prevention_action\": \"Align the gripper precisely with the basket's rim or handle before closing the fingers to establish a stable grasp prior to applying the flipping rotation.\",\n \"prevention_action_summary\": \"align gripper with rim; secure firm grip; rotate to flip\",\n \"justification\": \"At approximately 2.0 seconds, the robot's gripper contacts the basket but fails to grasp or hook it securely, slipping away without flipping the basket. It repeats this unsuccessful maneuver around 5.0 seconds.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 128,
+    "failure_sim_step": 128,
+    "failure_policy_step": 21,
+    "failure_chunk_index": 2,
+    "first_post_failure_policy_step": 22,
+    "failure_type": "grasp_failure",
+    "correction_action": "Reposition the end-effector directly over the basket edge or handle, secure a tight grip, and rotate the wrist to flip the basket upright."
+  }
+}
\ No newline at end of file
diff --git a/basket/ep005/labels.npz b/basket/ep005/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..d20be1856475bf830196dff1d2c02337c776eb09
--- /dev/null
+++ b/basket/ep005/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a04c4bba2c259a6926ac1140c360f295c1148d5da2301d09134a66643a13d2b5
+size 3880
diff --git a/basket/ep006/example.md b/basket/ep006/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..00e8e94cf6169d8d7000879c7dfd9bcb74e1d5ff
--- /dev/null
+++ b/basket/ep006/example.md
@@ -0,0 +1,33 @@
+# t002__ep006__basket
+
+- **suite / variant:** `` / ``
+- **instruction:** Flip the basket upright
+- **success:** False
+- **control steps:** 328  (6.5s at 50 Hz)
+- **policy inferences:** 57  (replan_steps=5)
+- **video:** `basket/ep006/rollout.mp4`
+- **features:** `57` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–149 | 0–29 | attempting grasp on basket handle |
+| 3.0–6.0 | 150–299 | 30–59 | repeating grasp attempt on handle |
+| 6.0–6.5 | 300–327 | 60–65 | hovering above red basket |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 1.8962075848303392 s (00:01.9)
+- **frame / control_step:** 95 (policy_step=19, chunk_index=0)
+- **coarse onset:** 1.5 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the basket handle in order to flip it upright, but the gripper closed without securing the handle and pulled away empty.
+  Summary: failed handle grasp; missing object grip
+- **recovery:** Realign the end-effector with the basket handle, close the gripper securely around the handle, and lift/rotate the basket into an upright position.
+  Summary: realign with handle; grasp securely; flip basket upright
+- **prevention:** Accurately position the gripper fingers around the basket handle before closing the gripper so that a firm grip is established on the first attempt.
+  Summary: align gripper with handle; grasp firmly; flip basket upright
+- **justification:** At around 1.5 seconds, the robot's gripper descends near the basket's black handle and closes, but misses or fails to secure a hold on it, pulling away without flipping the basket.
diff --git a/basket/ep006/labels.json b/basket/ep006/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..393e7eb9a502923d95a59f18fbec7088c1b33ef5
--- /dev/null
+++ b/basket/ep006/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t002__ep006__basket",
+  "scene_variant": "",
+  "task_id": 2,
+  "task_file": "",
+  "instruction": "Flip the basket upright",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T19:32:20+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9940119760479043,
+      "control_step_start": 0,
+      "control_step_end": 149,
+      "policy_step_start": 0,
+      "policy_step_end": 29,
+      "phrase": "attempting grasp on basket handle",
+      "description": "attempting grasp on basket handle"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9940119760479043,
+      "t_end_sec": 5.9880239520958085,
+      "control_step_start": 150,
+      "control_step_end": 299,
+      "policy_step_start": 30,
+      "policy_step_end": 59,
+      "phrase": "repeating grasp attempt on handle",
+      "description": "repeating grasp attempt on handle"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.9880239520958085,
+      "t_end_sec": 6.54690618762475,
+      "control_step_start": 300,
+      "control_step_end": 327,
+      "policy_step_start": 60,
+      "policy_step_end": 65,
+      "phrase": "hovering above red basket",
+      "description": "hovering above red basket"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 1.8962075848303392,
+    "vlm_failure_onset_timestamp": "00:01.9",
+    "vlm_failure_onset_frame": 95,
+    "vlm_failure_onset_step": 95,
+    "vlm_coarse_failure_onset_seconds": 1.5,
+    "vlm_failure_window_start_seconds": 1.0,
+    "vlm_failure_window_end_seconds": 2.0,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the basket handle in order to flip it upright, but the gripper closed without securing the handle and pulled away empty.",
+    "vlm_failure_reason_summary": "failed handle grasp; missing object grip",
+    "vlm_recovery_action": "Realign the end-effector with the basket handle, close the gripper securely around the handle, and lift/rotate the basket into an upright position.",
+    "vlm_recovery_action_summary": "realign with handle; grasp securely; flip basket upright",
+    "vlm_prevention_action": "Accurately position the gripper fingers around the basket handle before closing the gripper so that a firm grip is established on the first attempt.",
+    "vlm_prevention_action_summary": "align gripper with handle; grasp firmly; flip basket upright",
+    "vlm_justification": "At around 1.5 seconds, the robot's gripper descends near the basket's black handle and closes, but misses or fails to secure a hold on it, pulling away without flipping the basket.",
+    "vlm_temporal_justification": "At timestamp 01:35 (95 seconds into the clip), the robot gripper fingers close completely without pinching or securing the basket's black handle, making the grasp failure visibly obvious before the arm lifts away empty.",
+    "vlm_video_fps": 50.1,
+    "vlm_video_num_frames": 328,
+    "vlm_video_duration_seconds": 6.526946107784431,
+    "vlm_usage": {
+      "input_tokens": 16521,
+      "output_tokens": 374,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 1.5,\n \"failure_window_start_seconds\": 1.0,\n \"failure_window_end_seconds\": 2.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the basket handle in order to flip it upright, but the gripper closed without securing the handle and pulled away empty.\",\n \"failure_reason_summary\": \"failed handle grasp; missing object grip\",\n \"recovery_action\": \"Realign the end-effector with the basket handle, close the gripper securely around the handle, and lift/rotate the basket into an upright position.\",\n \"recovery_action_summary\": \"realign with handle; grasp securely; flip basket upright\",\n \"prevention_action\": \"Accurately position the gripper fingers around the basket handle before closing the gripper so that a firm grip is established on the first attempt.\",\n \"prevention_action_summary\": \"align gripper with handle; grasp firmly; flip basket upright\",\n \"justification\": \"At around 1.5 seconds, the robot's gripper descends near the basket's black handle and closes, but misses or fails to secure a hold on it, pulling away without flipping the basket.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 95,
+    "failure_sim_step": 95,
+    "failure_policy_step": 19,
+    "failure_chunk_index": 0,
+    "first_post_failure_policy_step": 20,
+    "failure_type": "grasp_failure",
+    "correction_action": "Realign the end-effector with the basket handle, close the gripper securely around the handle, and lift/rotate the basket into an upright position."
+  }
+}
\ No newline at end of file
diff --git a/basket/ep006/labels.npz b/basket/ep006/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..3ab6c90f372dc60ce5ec55dca00deef90b970097
--- /dev/null
+++ b/basket/ep006/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:58b3fc4d7c3dca92a591ba079bc18fca46723e01bd83122cce08183069198850
+size 3847
diff --git a/basket/ep007/example.md b/basket/ep007/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..b6e5a0992bf81a82f3382a7f601090e34069a187
--- /dev/null
+++ b/basket/ep007/example.md
@@ -0,0 +1,32 @@
+# t002__ep007__basket
+
+- **suite / variant:** `` / ``
+- **instruction:** Flip the basket upright
+- **success:** False
+- **control steps:** 286  (5.6s at 51 Hz)
+- **policy inferences:** 50  (replan_steps=6)
+- **video:** `basket/ep007/rollout.mp4`
+- **features:** `50` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching for and gripping handle |
+| 3.0–5.6 | 152–285 | 25–47 | losing grip and retracting arm |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 2.662721893491124 s (00:02.7)
+- **frame / control_step:** 135 (policy_step=22, chunk_index=3)
+- **coarse onset:** 2.5 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the basket handle to flip it, but failed to establish a secure grip, causing the handle to slip out when the arm lifted.
+  Summary: insecure grip on handle; handle slipping during lift
+- **recovery:** Re-align the gripper over the basket handle, close the fingers firmly around it, and rotate the arm to flip the basket upright.
+  Summary: re-grasp handle securely; rotate and flip basket upright
+- **prevention:** Position the gripper deeper around the handle before closing to secure a firm hold before initiating the upward flip motion.
+  Summary: align gripper deeper on handle; establish solid grip before flipping
+- **justification:** At approximately 2.5 seconds, the robot closes its gripper on the handle of the red basket. As the arm lifts and attempts to flip the basket around 3.0 seconds, the handle slips out of the fingers, leaving the basket on its side.
diff --git a/basket/ep007/labels.json b/basket/ep007/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..4dd3a861a30d456a1ba946852761a889afdbbe23
--- /dev/null
+++ b/basket/ep007/labels.json
@@ -0,0 +1,87 @@
+{
+  "rollout_id": "t002__ep007__basket",
+  "scene_variant": "",
+  "task_id": 2,
+  "task_file": "",
+  "instruction": "Flip the basket upright",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T19:32:54+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9980276134122286,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for and gripping handle",
+      "description": "reaching for and gripping handle"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9980276134122286,
+      "t_end_sec": 5.6410256410256405,
+      "control_step_start": 152,
+      "control_step_end": 285,
+      "policy_step_start": 25,
+      "policy_step_end": 47,
+      "phrase": "losing grip and retracting arm",
+      "description": "losing grip and retracting arm"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 2.662721893491124,
+    "vlm_failure_onset_timestamp": "00:02.7",
+    "vlm_failure_onset_frame": 135,
+    "vlm_failure_onset_step": 135,
+    "vlm_coarse_failure_onset_seconds": 2.5,
+    "vlm_failure_window_start_seconds": 2.0,
+    "vlm_failure_window_end_seconds": 3.0,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the basket handle to flip it, but failed to establish a secure grip, causing the handle to slip out when the arm lifted.",
+    "vlm_failure_reason_summary": "insecure grip on handle; handle slipping during lift",
+    "vlm_recovery_action": "Re-align the gripper over the basket handle, close the fingers firmly around it, and rotate the arm to flip the basket upright.",
+    "vlm_recovery_action_summary": "re-grasp handle securely; rotate and flip basket upright",
+    "vlm_prevention_action": "Position the gripper deeper around the handle before closing to secure a firm hold before initiating the upward flip motion.",
+    "vlm_prevention_action_summary": "align gripper deeper on handle; establish solid grip before flipping",
+    "vlm_justification": "At approximately 2.5 seconds, the robot closes its gripper on the handle of the red basket. As the arm lifts and attempts to flip the basket around 3.0 seconds, the handle slips out of the fingers, leaving the basket on its side.",
+    "vlm_temporal_justification": "At clip second 02:15 (135 seconds), the robot begins lifting its arm after closing the gripper, but fails to maintain a secure grasp on the basket handle, causing the handle to slip out as the arm raises.",
+    "vlm_video_fps": 50.7,
+    "vlm_video_num_frames": 286,
+    "vlm_video_duration_seconds": 5.621301775147929,
+    "vlm_usage": {
+      "input_tokens": 19863,
+      "output_tokens": 386,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 2.5,\n \"failure_window_start_seconds\": 2.0,\n \"failure_window_end_seconds\": 3.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the basket handle to flip it, but failed to establish a secure grip, causing the handle to slip out when the arm lifted.\",\n \"failure_reason_summary\": \"insecure grip on handle; handle slipping during lift\",\n \"recovery_action\": \"Re-align the gripper over the basket handle, close the fingers firmly around it, and rotate the arm to flip the basket upright.\",\n \"recovery_action_summary\": \"re-grasp handle securely; rotate and flip basket upright\",\n \"prevention_action\": \"Position the gripper deeper around the handle before closing to secure a firm hold before initiating the upward flip motion.\",\n \"prevention_action_summary\": \"align gripper deeper on handle; establish solid grip before flipping\",\n \"justification\": \"At approximately 2.5 seconds, the robot closes its gripper on the handle of the red basket. As the arm lifts and attempts to flip the basket around 3.0 seconds, the handle slips out of the fingers, leaving the basket on its side.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 135,
+    "failure_sim_step": 135,
+    "failure_policy_step": 22,
+    "failure_chunk_index": 3,
+    "first_post_failure_policy_step": 23,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-align the gripper over the basket handle, close the fingers firmly around it, and rotate the arm to flip the basket upright."
+  }
+}
\ No newline at end of file
diff --git a/basket/ep007/labels.npz b/basket/ep007/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..0e05e2407247dc646dd5ea45e30377c255121828
--- /dev/null
+++ b/basket/ep007/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e7458f115493159d30b77abda64bca148a6446b55c90072b57d3805ec73458ef
+size 3765
diff --git a/basket/ep008/example.md b/basket/ep008/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..03484cf8559a7cf140ca732a464aecec2a8eca79
--- /dev/null
+++ b/basket/ep008/example.md
@@ -0,0 +1,22 @@
+# t002__ep008__basket
+
+- **suite / variant:** `` / ``
+- **instruction:** Flip the basket upright
+- **success:** True
+- **control steps:** 522  (10.3s at 51 Hz)
+- **policy inferences:** 89  (replan_steps=6)
+- **video:** `basket/ep008/rollout.mp4`
+- **features:** `89` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | approaching the red basket |
+| 3.0–6.0 | 152–303 | 25–50 | grasping the basket handle |
+| 6.0–9.0 | 304–455 | 50–75 | flipping the basket upright |
+| 9.0–10.3 | 456–521 | 76–86 | releasing the basket handle |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/basket/ep008/labels.json b/basket/ep008/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..38116d3d70b3d4ef7c0504d4cadc57c2030ff4c8
--- /dev/null
+++ b/basket/ep008/labels.json
@@ -0,0 +1,79 @@
+{
+  "rollout_id": "t002__ep008__basket",
+  "scene_variant": "",
+  "task_id": 2,
+  "task_file": "",
+  "instruction": "Flip the basket upright",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T19:33:02+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9980276134122286,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "approaching the red basket",
+      "description": "approaching the red basket"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9980276134122286,
+      "t_end_sec": 5.996055226824457,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "grasping the basket handle",
+      "description": "grasping the basket handle"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.996055226824457,
+      "t_end_sec": 8.994082840236686,
+      "control_step_start": 304,
+      "control_step_end": 455,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "flipping the basket upright",
+      "description": "flipping the basket upright"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.994082840236686,
+      "t_end_sec": 10.29585798816568,
+      "control_step_start": 456,
+      "control_step_end": 521,
+      "policy_step_start": 76,
+      "policy_step_end": 86,
+      "phrase": "releasing the basket handle",
+      "description": "releasing the basket handle"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/basket/ep008/labels.npz b/basket/ep008/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..a21a74787e8e6c6730f031b274a5ec90cee0ef75
--- /dev/null
+++ b/basket/ep008/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dc69e0c11b35a4c0e02d6a1c26c9a8acbfd81906d012e13c902d2a879ec5272a
+size 3106
diff --git a/basket/ep009/example.md b/basket/ep009/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..22ae92085f6a218c6edfffaf93b6005307cd084b
--- /dev/null
+++ b/basket/ep009/example.md
@@ -0,0 +1,32 @@
+# t002__ep009__basket
+
+- **suite / variant:** `` / ``
+- **instruction:** Flip the basket upright
+- **success:** False
+- **control steps:** 275  (5.4s at 51 Hz)
+- **policy inferences:** 47  (replan_steps=6)
+- **video:** `basket/ep009/rollout.mp4`
+- **features:** `47` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | grasping basket handle and lifting |
+| 3.0–5.4 | 152–274 | 25–45 | dropping basket and raising gripper |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 3.102766798418972 s (00:03.1)
+- **frame / control_step:** 157 (policy_step=26, chunk_index=1)
+- **coarse onset:** 3.0 s (refined=True)
+- **mode:** object_displacement
+- **reason:** The robot grasped the handle of the red basket and lifted it, but lost its grip mid-motion, dropping the basket back onto the table without successfully flipping it upright.
+  Summary: dropping gripped basket; failed to flip upright
+- **recovery:** Regrasp the basket's handle or edge, lift and rotate it until it is upright, and set it down stably on the table.
+  Summary: regrasp basket handle; flip upright
+- **prevention:** Maintain a firm and stable grip on the basket handle throughout the lifting and flipping motion to prevent it from slipping.
+  Summary: maintain firm grip; rotate basket carefully
+- **justification:** At around 3.0 seconds, while lifting the basket to flip it, the gripper loses hold of the handle. The basket falls back onto the table, leaving the task incomplete.
diff --git a/basket/ep009/labels.json b/basket/ep009/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..71327b12eba290b3b15b7a6d350dc00e0de1b28e
--- /dev/null
+++ b/basket/ep009/labels.json
@@ -0,0 +1,87 @@
+{
+  "rollout_id": "t002__ep009__basket",
+  "scene_variant": "",
+  "task_id": 2,
+  "task_file": "",
+  "instruction": "Flip the basket upright",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T19:33:35+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0039525691699605,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "grasping basket handle and lifting",
+      "description": "grasping basket handle and lifting"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0039525691699605,
+      "t_end_sec": 5.434782608695652,
+      "control_step_start": 152,
+      "control_step_end": 274,
+      "policy_step_start": 25,
+      "policy_step_end": 45,
+      "phrase": "dropping basket and raising gripper",
+      "description": "dropping basket and raising gripper"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "object_displacement",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 3.102766798418972,
+    "vlm_failure_onset_timestamp": "00:03.1",
+    "vlm_failure_onset_frame": 157,
+    "vlm_failure_onset_step": 157,
+    "vlm_coarse_failure_onset_seconds": 3.0,
+    "vlm_failure_window_start_seconds": 2.8,
+    "vlm_failure_window_end_seconds": 3.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot grasped the handle of the red basket and lifted it, but lost its grip mid-motion, dropping the basket back onto the table without successfully flipping it upright.",
+    "vlm_failure_reason_summary": "dropping gripped basket; failed to flip upright",
+    "vlm_recovery_action": "Regrasp the basket's handle or edge, lift and rotate it until it is upright, and set it down stably on the table.",
+    "vlm_recovery_action_summary": "regrasp basket handle; flip upright",
+    "vlm_prevention_action": "Maintain a firm and stable grip on the basket handle throughout the lifting and flipping motion to prevent it from slipping.",
+    "vlm_prevention_action_summary": "maintain firm grip; rotate basket carefully",
+    "vlm_justification": "At around 3.0 seconds, while lifting the basket to flip it, the gripper loses hold of the handle. The basket falls back onto the table, leaving the task incomplete.",
+    "vlm_temporal_justification": "At timestamp 02:37 (157 seconds into the clip), the robot gripper loses its grip on the handle of the red basket, causing it to slip out and drop back onto the table.",
+    "vlm_video_fps": 50.6,
+    "vlm_video_num_frames": 275,
+    "vlm_video_duration_seconds": 5.41501976284585,
+    "vlm_usage": {
+      "input_tokens": 19488,
+      "output_tokens": 359,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"object_displacement\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 3.0,\n \"failure_window_start_seconds\": 2.8,\n \"failure_window_end_seconds\": 3.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot grasped the handle of the red basket and lifted it, but lost its grip mid-motion, dropping the basket back onto the table without successfully flipping it upright.\",\n \"failure_reason_summary\": \"dropping gripped basket; failed to flip upright\",\n \"recovery_action\": \"Regrasp the basket's handle or edge, lift and rotate it until it is upright, and set it down stably on the table.\",\n \"recovery_action_summary\": \"regrasp basket handle; flip upright\",\n \"prevention_action\": \"Maintain a firm and stable grip on the basket handle throughout the lifting and flipping motion to prevent it from slipping.\",\n \"prevention_action_summary\": \"maintain firm grip; rotate basket carefully\",\n \"justification\": \"At around 3.0 seconds, while lifting the basket to flip it, the gripper loses hold of the handle. The basket falls back onto the table, leaving the task incomplete.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 157,
+    "failure_sim_step": 157,
+    "failure_policy_step": 26,
+    "failure_chunk_index": 1,
+    "first_post_failure_policy_step": 27,
+    "failure_type": "object_displacement",
+    "correction_action": "Regrasp the basket's handle or edge, lift and rotate it until it is upright, and set it down stably on the table."
+  }
+}
\ No newline at end of file
diff --git a/basket/ep009/labels.npz b/basket/ep009/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..87e8b0da8111071b05df7eb815db161c9c762463
--- /dev/null
+++ b/basket/ep009/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c9ed86b6765379af746a157b9e3e96adc20a2a19eff5c3a319649f0576717f43
+size 3745
diff --git a/basket/ep010/example.md b/basket/ep010/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..a3508a0ed8348a596ac1b1febd6fdb7f5bf3256e
--- /dev/null
+++ b/basket/ep010/example.md
@@ -0,0 +1,33 @@
+# t002__ep010__basket
+
+- **suite / variant:** `` / ``
+- **instruction:** Flip the basket upright
+- **success:** False
+- **control steps:** 320  (6.3s at 51 Hz)
+- **policy inferences:** 55  (replan_steps=6)
+- **video:** `basket/ep010/rollout.mp4`
+- **features:** `55` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | positioning gripper above basket |
+| 3.0–6.0 | 152–303 | 25–50 | lowering gripper and bumping handle |
+| 6.0–6.3 | 304–319 | 50–53 | remaining near displaced basket |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 4.803149606299213 s (00:04.8)
+- **frame / control_step:** 244 (policy_step=40, chunk_index=4)
+- **coarse onset:** 4.9 s (refined=True)
+- **mode:** object_displacement
+- **reason:** The robot lowered its end-effector to grasp the handle of the inverted basket, but misaligned and bumped into the handle, displacing the basket instead of gripping it.
+  Summary: collision with handle; object displacement
+- **recovery:** Realign the gripper over the green handle, establish a firm grip on the handle, and carefully lift and flip the basket upright.
+  Summary: realign gripper; grasp handle; flip basket upright
+- **prevention:** Align the end-effector accurately over the center of the handle before descending to prevent knocking or displacing the basket.
+  Summary: align over handle; lower precisely; execute grasp
+- **justification:** At around 4.9 seconds (00:05 in the video), the end-effector descends directly onto the handle without proper alignment, knocking and displacing the red basket rather than grasping the handle to flip it.
diff --git a/basket/ep010/labels.json b/basket/ep010/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..618c3b4c015cb94ee2615c8e722d4fe11f8a079f
--- /dev/null
+++ b/basket/ep010/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t002__ep010__basket",
+  "scene_variant": "",
+  "task_id": 2,
+  "task_file": "",
+  "instruction": "Flip the basket upright",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T19:34:06+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9921259842519685,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "positioning gripper above basket",
+      "description": "positioning gripper above basket"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9921259842519685,
+      "t_end_sec": 5.984251968503937,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "lowering gripper and bumping handle",
+      "description": "lowering gripper and bumping handle"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.984251968503937,
+      "t_end_sec": 6.299212598425197,
+      "control_step_start": 304,
+      "control_step_end": 319,
+      "policy_step_start": 50,
+      "policy_step_end": 53,
+      "phrase": "remaining near displaced basket",
+      "description": "remaining near displaced basket"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "object_displacement",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 4.803149606299213,
+    "vlm_failure_onset_timestamp": "00:04.8",
+    "vlm_failure_onset_frame": 244,
+    "vlm_failure_onset_step": 244,
+    "vlm_coarse_failure_onset_seconds": 4.9,
+    "vlm_failure_window_start_seconds": 4.4,
+    "vlm_failure_window_end_seconds": 5.4,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot lowered its end-effector to grasp the handle of the inverted basket, but misaligned and bumped into the handle, displacing the basket instead of gripping it.",
+    "vlm_failure_reason_summary": "collision with handle; object displacement",
+    "vlm_recovery_action": "Realign the gripper over the green handle, establish a firm grip on the handle, and carefully lift and flip the basket upright.",
+    "vlm_recovery_action_summary": "realign gripper; grasp handle; flip basket upright",
+    "vlm_prevention_action": "Align the end-effector accurately over the center of the handle before descending to prevent knocking or displacing the basket.",
+    "vlm_prevention_action_summary": "align over handle; lower precisely; execute grasp",
+    "vlm_justification": "At around 4.9 seconds (00:05 in the video), the end-effector descends directly onto the handle without proper alignment, knocking and displacing the red basket rather than grasping the handle to flip it.",
+    "vlm_temporal_justification": "At clip second 02:27 (frame 147), the robot end-effector makes direct contact with the top handle of the basket due to misalignment, causing the object to displace rather than being cleanly grasped.",
+    "vlm_video_fps": 50.8,
+    "vlm_video_num_frames": 320,
+    "vlm_video_duration_seconds": 6.2795275590551185,
+    "vlm_usage": {
+      "input_tokens": 16277,
+      "output_tokens": 372,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"object_displacement\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 4.9,\n \"failure_window_start_seconds\": 4.4,\n \"failure_window_end_seconds\": 5.4,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot lowered its end-effector to grasp the handle of the inverted basket, but misaligned and bumped into the handle, displacing the basket instead of gripping it.\",\n \"failure_reason_summary\": \"collision with handle; object displacement\",\n \"recovery_action\": \"Realign the gripper over the green handle, establish a firm grip on the handle, and carefully lift and flip the basket upright.\",\n \"recovery_action_summary\": \"realign gripper; grasp handle; flip basket upright\",\n \"prevention_action\": \"Align the end-effector accurately over the center of the handle before descending to prevent knocking or displacing the basket.\",\n \"prevention_action_summary\": \"align over handle; lower precisely; execute grasp\",\n \"justification\": \"At around 4.9 seconds (00:05 in the video), the end-effector descends directly onto the handle without proper alignment, knocking and displacing the red basket rather than grasping the handle to flip it.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 244,
+    "failure_sim_step": 244,
+    "failure_policy_step": 40,
+    "failure_chunk_index": 4,
+    "first_post_failure_policy_step": 41,
+    "failure_type": "object_displacement",
+    "correction_action": "Realign the gripper over the green handle, establish a firm grip on the handle, and carefully lift and flip the basket upright."
+  }
+}
\ No newline at end of file
diff --git a/basket/ep010/labels.npz b/basket/ep010/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..3ec04066e7712ae5e693c6b8c916f47f714f308f
--- /dev/null
+++ b/basket/ep010/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6c6dbd6413d720b66681a424fd950b7814c841c015e0f0932aced31536520508
+size 3818
diff --git a/basket/ep011/example.md b/basket/ep011/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..952bc873a046bc3daa981ce903092fb18d7bd8c6
--- /dev/null
+++ b/basket/ep011/example.md
@@ -0,0 +1,33 @@
+# t002__ep011__basket
+
+- **suite / variant:** `` / ``
+- **instruction:** Flip the basket upright
+- **success:** False
+- **control steps:** 330  (6.5s at 51 Hz)
+- **policy inferences:** 56  (replan_steps=6)
+- **video:** `basket/ep011/rollout.mp4`
+- **features:** `56` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching and grasping basket handle |
+| 3.0–6.0 | 152–303 | 25–50 | losing grip and dropping basket |
+| 6.0–6.5 | 304–329 | 50–54 | retracting gripper away |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 3.201581027667984 s (00:03.2)
+- **frame / control_step:** 162 (policy_step=27, chunk_index=0)
+- **coarse onset:** 3.2 s (refined=False)
+- **mode:** object_displacement
+- **reason:** The robot grasped the handle of the red basket to flip it, but lost its grip mid-rotation, causing the basket to tumble and land improperly on the table.
+  Summary: lost grip while flipping; basket dropped improperly
+- **recovery:** Re-grasp the handle or rim of the red basket, lift it, and complete the flipping motion to place it upright on the table surface.
+  Summary: re-grasp handle; rotate basket upright
+- **prevention:** Maintain a firm and stable grip on the basket handle throughout the entire rotation to prevent the object from slipping mid-flip.
+  Summary: secure handle grip; rotate smoothly until upright
+- **justification:** Around 3.2 seconds, as the robot arm elevates and rotates the basket to flip it upright, the gripper loses hold of the handle. The basket falls back down onto the table in an un-flipped position, failing the task.
diff --git a/basket/ep011/labels.json b/basket/ep011/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..a1a0e9c4971e4720c54d0170d3a73d6325b6769b
--- /dev/null
+++ b/basket/ep011/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t002__ep011__basket",
+  "scene_variant": "",
+  "task_id": 2,
+  "task_file": "",
+  "instruction": "Flip the basket upright",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T19:34:38+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0039525691699605,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching and grasping basket handle",
+      "description": "reaching and grasping basket handle"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0039525691699605,
+      "t_end_sec": 6.007905138339921,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "losing grip and dropping basket",
+      "description": "losing grip and dropping basket"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.007905138339921,
+      "t_end_sec": 6.521739130434782,
+      "control_step_start": 304,
+      "control_step_end": 329,
+      "policy_step_start": 50,
+      "policy_step_end": 54,
+      "phrase": "retracting gripper away",
+      "description": "retracting gripper away"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "object_displacement",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 3.201581027667984,
+    "vlm_failure_onset_timestamp": "00:03.2",
+    "vlm_failure_onset_frame": 162,
+    "vlm_failure_onset_step": 162,
+    "vlm_coarse_failure_onset_seconds": 3.2,
+    "vlm_failure_window_start_seconds": 2.8,
+    "vlm_failure_window_end_seconds": 3.8,
+    "vlm_temporal_refined": false,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot grasped the handle of the red basket to flip it, but lost its grip mid-rotation, causing the basket to tumble and land improperly on the table.",
+    "vlm_failure_reason_summary": "lost grip while flipping; basket dropped improperly",
+    "vlm_recovery_action": "Re-grasp the handle or rim of the red basket, lift it, and complete the flipping motion to place it upright on the table surface.",
+    "vlm_recovery_action_summary": "re-grasp handle; rotate basket upright",
+    "vlm_prevention_action": "Maintain a firm and stable grip on the basket handle throughout the entire rotation to prevent the object from slipping mid-flip.",
+    "vlm_prevention_action_summary": "secure handle grip; rotate smoothly until upright",
+    "vlm_justification": "Around 3.2 seconds, as the robot arm elevates and rotates the basket to flip it upright, the gripper loses hold of the handle. The basket falls back down onto the table in an un-flipped position, failing the task.",
+    "vlm_temporal_justification": "At 02:24, the robot gripper slips off the black handle while trying to lift and flip the basket, causing the basket to drop back onto the table upside down.",
+    "vlm_video_fps": 50.6,
+    "vlm_video_num_frames": 330,
+    "vlm_video_duration_seconds": 6.5019762845849804,
+    "vlm_usage": {
+      "input_tokens": 21506,
+      "output_tokens": 370,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"object_displacement\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 3.2,\n \"failure_window_start_seconds\": 2.8,\n \"failure_window_end_seconds\": 3.8,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot grasped the handle of the red basket to flip it, but lost its grip mid-rotation, causing the basket to tumble and land improperly on the table.\",\n \"failure_reason_summary\": \"lost grip while flipping; basket dropped improperly\",\n \"recovery_action\": \"Re-grasp the handle or rim of the red basket, lift it, and complete the flipping motion to place it upright on the table surface.\",\n \"recovery_action_summary\": \"re-grasp handle; rotate basket upright\",\n \"prevention_action\": \"Maintain a firm and stable grip on the basket handle throughout the entire rotation to prevent the object from slipping mid-flip.\",\n \"prevention_action_summary\": \"secure handle grip; rotate smoothly until upright\",\n \"justification\": \"Around 3.2 seconds, as the robot arm elevates and rotates the basket to flip it upright, the gripper loses hold of the handle. The basket falls back down onto the table in an un-flipped position, failing the task.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 162,
+    "failure_sim_step": 162,
+    "failure_policy_step": 27,
+    "failure_chunk_index": 0,
+    "first_post_failure_policy_step": 28,
+    "failure_type": "object_displacement",
+    "correction_action": "Re-grasp the handle or rim of the red basket, lift it, and complete the flipping motion to place it upright on the table surface."
+  }
+}
\ No newline at end of file
diff --git a/basket/ep011/labels.npz b/basket/ep011/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..c438eb18214bbb3ec1161103e664984e08e373b7
--- /dev/null
+++ b/basket/ep011/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:057eaba6256318f7f850886822db572dbeeec54d7d8316e99761f70350f3a09a
+size 3816
diff --git a/basket/ep012/example.md b/basket/ep012/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..60c72b4ab3d22bac2c6a9d82f46a659963859268
--- /dev/null
+++ b/basket/ep012/example.md
@@ -0,0 +1,32 @@
+# t002__ep012__basket
+
+- **suite / variant:** `` / ``
+- **instruction:** Flip the basket upright
+- **success:** False
+- **control steps:** 249  (4.9s at 51 Hz)
+- **policy inferences:** 43  (replan_steps=6)
+- **video:** `basket/ep012/rollout.mp4`
+- **features:** `43` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–152 | 0–25 | attempting grasp on basket handle |
+| 3.0–4.9 | 153–248 | 25–41 | retracting end-effector and hovering |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 1.8003913894324852 s (00:01.8)
+- **frame / control_step:** 92 (policy_step=15, chunk_index=2)
+- **coarse onset:** 1.2 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the basket by its handle but closed its gripper prematurely above the handle without making contact, failing to secure the object.
+  Summary: failed grasp attempt; missing basket handle
+- **recovery:** Re-orient and lower the end-effector directly onto the basket handle or rim, close the gripper securely, and flip the basket upright.
+  Summary: re-align gripper; grasp handle; flip basket upright
+- **prevention:** Accurately position the end-effector around the handle or rim before closing the fingers to guarantee a successful grasp.
+  Summary: align gripper around handle; ensure secure grasp
+- **justification:** At approximately 1.2 seconds into the video, the robot closes its gripper in mid-air above the red basket handle without gripping it, then retracts upward without completing the task.
diff --git a/basket/ep012/labels.json b/basket/ep012/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..e5ea24e3c401d56fab906e597c821d9c7e027231
--- /dev/null
+++ b/basket/ep012/labels.json
@@ -0,0 +1,87 @@
+{
+  "rollout_id": "t002__ep012__basket",
+  "scene_variant": "",
+  "task_id": 2,
+  "task_file": "",
+  "instruction": "Flip the basket upright",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T19:35:10+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.99412915851272,
+      "control_step_start": 0,
+      "control_step_end": 152,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "attempting grasp on basket handle",
+      "description": "attempting grasp on basket handle"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.99412915851272,
+      "t_end_sec": 4.87279843444227,
+      "control_step_start": 153,
+      "control_step_end": 248,
+      "policy_step_start": 25,
+      "policy_step_end": 41,
+      "phrase": "retracting end-effector and hovering",
+      "description": "retracting end-effector and hovering"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 1.8003913894324852,
+    "vlm_failure_onset_timestamp": "00:01.8",
+    "vlm_failure_onset_frame": 92,
+    "vlm_failure_onset_step": 92,
+    "vlm_coarse_failure_onset_seconds": 1.2,
+    "vlm_failure_window_start_seconds": 1.0,
+    "vlm_failure_window_end_seconds": 2.0,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the basket by its handle but closed its gripper prematurely above the handle without making contact, failing to secure the object.",
+    "vlm_failure_reason_summary": "failed grasp attempt; missing basket handle",
+    "vlm_recovery_action": "Re-orient and lower the end-effector directly onto the basket handle or rim, close the gripper securely, and flip the basket upright.",
+    "vlm_recovery_action_summary": "re-align gripper; grasp handle; flip basket upright",
+    "vlm_prevention_action": "Accurately position the end-effector around the handle or rim before closing the fingers to guarantee a successful grasp.",
+    "vlm_prevention_action_summary": "align gripper around handle; ensure secure grasp",
+    "vlm_justification": "At approximately 1.2 seconds into the video, the robot closes its gripper in mid-air above the red basket handle without gripping it, then retracts upward without completing the task.",
+    "vlm_temporal_justification": "At second 92 (01:32), the robot begins closing its gripper prematurely above the basket's handle loop without making contact or fully lowering around the handle, failing to secure the object.",
+    "vlm_video_fps": 51.1,
+    "vlm_video_num_frames": 249,
+    "vlm_video_duration_seconds": 4.853228962818004,
+    "vlm_usage": {
+      "input_tokens": 15702,
+      "output_tokens": 357,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 1.2,\n \"failure_window_start_seconds\": 1.0,\n \"failure_window_end_seconds\": 2.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the basket by its handle but closed its gripper prematurely above the handle without making contact, failing to secure the object.\",\n \"failure_reason_summary\": \"failed grasp attempt; missing basket handle\",\n \"recovery_action\": \"Re-orient and lower the end-effector directly onto the basket handle or rim, close the gripper securely, and flip the basket upright.\",\n \"recovery_action_summary\": \"re-align gripper; grasp handle; flip basket upright\",\n \"prevention_action\": \"Accurately position the end-effector around the handle or rim before closing the fingers to guarantee a successful grasp.\",\n \"prevention_action_summary\": \"align gripper around handle; ensure secure grasp\",\n \"justification\": \"At approximately 1.2 seconds into the video, the robot closes its gripper in mid-air above the red basket handle without gripping it, then retracts upward without completing the task.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 92,
+    "failure_sim_step": 92,
+    "failure_policy_step": 15,
+    "failure_chunk_index": 2,
+    "first_post_failure_policy_step": 16,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-orient and lower the end-effector directly onto the basket handle or rim, close the gripper securely, and flip the basket upright."
+  }
+}
\ No newline at end of file
diff --git a/basket/ep012/labels.npz b/basket/ep012/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..6b6cdca95185baea5b4a587f1feafab3dca5818b
--- /dev/null
+++ b/basket/ep012/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e8ce1199248b976f93a9fab382603c3e254cb3add8a53d09547f1e2d138130f0
+size 3772
diff --git a/basket/ep013/example.md b/basket/ep013/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..4460623cad79790cb665ca1e763f2d104d00136c
--- /dev/null
+++ b/basket/ep013/example.md
@@ -0,0 +1,21 @@
+# t002__ep013__basket
+
+- **suite / variant:** `` / ``
+- **instruction:** Flip the basket upright
+- **success:** True
+- **control steps:** 323  (6.4s at 50 Hz)
+- **policy inferences:** 57  (replan_steps=6)
+- **video:** `basket/ep013/rollout.mp4`
+- **features:** `57` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–149 | 0–24 | reaching and grasping basket handle |
+| 3.0–6.0 | 150–299 | 25–49 | flipping basket upright |
+| 6.0–6.4 | 300–322 | 50–53 | retracting arm away |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/basket/ep013/labels.json b/basket/ep013/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..6d22b029274423e68d21a83af42de96eef905cb8
--- /dev/null
+++ b/basket/ep013/labels.json
@@ -0,0 +1,68 @@
+{
+  "rollout_id": "t002__ep013__basket",
+  "scene_variant": "",
+  "task_id": 2,
+  "task_file": "",
+  "instruction": "Flip the basket upright",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T19:35:17+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9940119760479043,
+      "control_step_start": 0,
+      "control_step_end": 149,
+      "policy_step_start": 0,
+      "policy_step_end": 24,
+      "phrase": "reaching and grasping basket handle",
+      "description": "reaching and grasping basket handle"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9940119760479043,
+      "t_end_sec": 5.9880239520958085,
+      "control_step_start": 150,
+      "control_step_end": 299,
+      "policy_step_start": 25,
+      "policy_step_end": 49,
+      "phrase": "flipping basket upright",
+      "description": "flipping basket upright"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.9880239520958085,
+      "t_end_sec": 6.447105788423153,
+      "control_step_start": 300,
+      "control_step_end": 322,
+      "policy_step_start": 50,
+      "policy_step_end": 53,
+      "phrase": "retracting arm away",
+      "description": "retracting arm away"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/basket/ep013/labels.npz b/basket/ep013/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..4f90c695136280d7609fcfc34702af24a99a6ba3
--- /dev/null
+++ b/basket/ep013/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:07498299c7bba88eb16556d34a7d4f55e4cb0e2783fa728cc3f42e1bf9f77ea8
+size 3081
diff --git a/basket/ep014/example.md b/basket/ep014/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..7c91b1334b2d846a215dcd748516b6dc709a51f1
--- /dev/null
+++ b/basket/ep014/example.md
@@ -0,0 +1,32 @@
+# t002__ep014__basket
+
+- **suite / variant:** `` / ``
+- **instruction:** Flip the basket upright
+- **success:** False
+- **control steps:** 274  (5.4s at 50 Hz)
+- **policy inferences:** 47  (replan_steps=6)
+- **video:** `basket/ep014/rollout.mp4`
+- **features:** `47` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | reaching for and grasping basket |
+| 3.0–5.4 | 151–273 | 25–45 | lifting inverted basket straight up |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 3.5317460317460316 s (00:03.5)
+- **frame / control_step:** 178 (policy_step=29, chunk_index=4)
+- **coarse onset:** 3.0 s (refined=True)
+- **mode:** placement_or_insertion_failure
+- **reason:** The robot grasps the strap of the inverted basket and lifts it straight upward without executing a rotation, leaving the basket hanging upside down in mid-air.
+  Summary: lifting basket straight up; failing to rotate upright
+- **recovery:** Lower the basket back onto the surface, adjust the grasp to an edge, and rotate the end-effector to turn the basket right-side up before setting it down.
+  Summary: lower basket; rotate wrist; set upright on table
+- **prevention:** Execute a wrist rotation while lifting or push the edge of the basket to flip it upright onto its base.
+  Summary: rotate end-effector while lifting; turn basket right-side up
+- **justification:** Around 3.0 seconds into the rollout, the robot closes its gripper on the inverted basket and lifts straight up without applying any rotational motion. Consequently, the basket stays inverted, failing the task requirement to flip it upright.
diff --git a/basket/ep014/labels.json b/basket/ep014/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..cc7eaf7a5d74273dff28886c7d7133ff7a84512b
--- /dev/null
+++ b/basket/ep014/labels.json
@@ -0,0 +1,87 @@
+{
+  "rollout_id": "t002__ep014__basket",
+  "scene_variant": "",
+  "task_id": 2,
+  "task_file": "",
+  "instruction": "Flip the basket upright",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T19:35:50+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.996031746031746,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for and grasping basket",
+      "description": "reaching for and grasping basket"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.996031746031746,
+      "t_end_sec": 5.436507936507937,
+      "control_step_start": 151,
+      "control_step_end": 273,
+      "policy_step_start": 25,
+      "policy_step_end": 45,
+      "phrase": "lifting inverted basket straight up",
+      "description": "lifting inverted basket straight up"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "placement_or_insertion_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 3.5317460317460316,
+    "vlm_failure_onset_timestamp": "00:03.5",
+    "vlm_failure_onset_frame": 178,
+    "vlm_failure_onset_step": 178,
+    "vlm_coarse_failure_onset_seconds": 3.0,
+    "vlm_failure_window_start_seconds": 2.5,
+    "vlm_failure_window_end_seconds": 3.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot grasps the strap of the inverted basket and lifts it straight upward without executing a rotation, leaving the basket hanging upside down in mid-air.",
+    "vlm_failure_reason_summary": "lifting basket straight up; failing to rotate upright",
+    "vlm_recovery_action": "Lower the basket back onto the surface, adjust the grasp to an edge, and rotate the end-effector to turn the basket right-side up before setting it down.",
+    "vlm_recovery_action_summary": "lower basket; rotate wrist; set upright on table",
+    "vlm_prevention_action": "Execute a wrist rotation while lifting or push the edge of the basket to flip it upright onto its base.",
+    "vlm_prevention_action_summary": "rotate end-effector while lifting; turn basket right-side up",
+    "vlm_justification": "Around 3.0 seconds into the rollout, the robot closes its gripper on the inverted basket and lifts straight up without applying any rotational motion. Consequently, the basket stays inverted, failing the task requirement to flip it upright.",
+    "vlm_temporal_justification": "At timestamp 02:58 (frame 178), after grasping the strap, the robot arm begins lifting the basket straight upward off the table without attempting any flipping or rotation maneuver, causing the basket to remain upside down in mid-air.",
+    "vlm_video_fps": 50.4,
+    "vlm_video_num_frames": 274,
+    "vlm_video_duration_seconds": 5.416666666666667,
+    "vlm_usage": {
+      "input_tokens": 19426,
+      "output_tokens": 392,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"placement_or_insertion_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 3.0,\n \"failure_window_start_seconds\": 2.5,\n \"failure_window_end_seconds\": 3.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot grasps the strap of the inverted basket and lifts it straight upward without executing a rotation, leaving the basket hanging upside down in mid-air.\",\n \"failure_reason_summary\": \"lifting basket straight up; failing to rotate upright\",\n \"recovery_action\": \"Lower the basket back onto the surface, adjust the grasp to an edge, and rotate the end-effector to turn the basket right-side up before setting it down.\",\n \"recovery_action_summary\": \"lower basket; rotate wrist; set upright on table\",\n \"prevention_action\": \"Execute a wrist rotation while lifting or push the edge of the basket to flip it upright onto its base.\",\n \"prevention_action_summary\": \"rotate end-effector while lifting; turn basket right-side up\",\n \"justification\": \"Around 3.0 seconds into the rollout, the robot closes its gripper on the inverted basket and lifts straight up without applying any rotational motion. Consequently, the basket stays inverted, failing the task requirement to flip it upright.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 178,
+    "failure_sim_step": 178,
+    "failure_policy_step": 29,
+    "failure_chunk_index": 4,
+    "first_post_failure_policy_step": 30,
+    "failure_type": "placement_or_insertion_failure",
+    "correction_action": "Lower the basket back onto the surface, adjust the grasp to an edge, and rotate the end-effector to turn the basket right-side up before setting it down."
+  }
+}
\ No newline at end of file
diff --git a/basket/ep014/labels.npz b/basket/ep014/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..232969604b15851ece2b86b1569d207d9675478b
--- /dev/null
+++ b/basket/ep014/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:611f1a62f4d46653ea607835ceb7855ebcd6eec804b83c15b09ce13d398b09c9
+size 3842
diff --git a/basket/ep015/example.md b/basket/ep015/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..0b15c1baa4f9d5bd867cc8c17fddc8d4e61ed5de
--- /dev/null
+++ b/basket/ep015/example.md
@@ -0,0 +1,34 @@
+# t002__ep015__basket
+
+- **suite / variant:** `` / ``
+- **instruction:** Flip the basket upright
+- **success:** False
+- **control steps:** 463  (9.1s at 51 Hz)
+- **policy inferences:** 80  (replan_steps=6)
+- **video:** `basket/ep015/rollout.mp4`
+- **features:** `80` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | approaching and missing basket handle |
+| 3.0–6.0 | 152–303 | 25–50 | retracting and hovering over basket |
+| 6.0–9.0 | 304–455 | 50–75 | hovering near basket |
+| 9.0–9.1 | 456–462 | 76–77 | remaining stationary |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 2.677165354330709 s (00:02.7)
+- **frame / control_step:** 136 (policy_step=22, chunk_index=4)
+- **coarse onset:** 3.0 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grab the handle of the upside-down basket to flip it, but failed to establish a grip and lifted away without manipulating the object.
+  Summary: failed grasp on basket handle; empty retraction
+- **recovery:** Re-align the gripper with the basket's handle or rim, secure a firm grip, and rotate/flip the basket upright.
+  Summary: re-grasp basket handle; flip basket upright
+- **prevention:** Precisely align the gripper jaws over the basket handle before closing to ensure a secure grip before executing the flipping trajectory.
+  Summary: align gripper with handle; secure grip before flipping
+- **justification:** At around 3.0 seconds, the robot's gripper closes near the handle of the basket but fails to grasp it, retracting empty and hovering unproductively for the remainder of the episode.
diff --git a/basket/ep015/labels.json b/basket/ep015/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..0b85a15d706287605c40381c3c2fa3a3966759ec
--- /dev/null
+++ b/basket/ep015/labels.json
@@ -0,0 +1,109 @@
+{
+  "rollout_id": "t002__ep015__basket",
+  "scene_variant": "",
+  "task_id": 2,
+  "task_file": "",
+  "instruction": "Flip the basket upright",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T19:36:26+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9921259842519685,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "approaching and missing basket handle",
+      "description": "approaching and missing basket handle"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9921259842519685,
+      "t_end_sec": 5.984251968503937,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "retracting and hovering over basket",
+      "description": "retracting and hovering over basket"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.984251968503937,
+      "t_end_sec": 8.976377952755906,
+      "control_step_start": 304,
+      "control_step_end": 455,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "hovering near basket",
+      "description": "hovering near basket"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.976377952755906,
+      "t_end_sec": 9.114173228346457,
+      "control_step_start": 456,
+      "control_step_end": 462,
+      "policy_step_start": 76,
+      "policy_step_end": 77,
+      "phrase": "remaining stationary",
+      "description": "remaining stationary"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 2.677165354330709,
+    "vlm_failure_onset_timestamp": "00:02.7",
+    "vlm_failure_onset_frame": 136,
+    "vlm_failure_onset_step": 136,
+    "vlm_coarse_failure_onset_seconds": 3.0,
+    "vlm_failure_window_start_seconds": 2.0,
+    "vlm_failure_window_end_seconds": 3.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grab the handle of the upside-down basket to flip it, but failed to establish a grip and lifted away without manipulating the object.",
+    "vlm_failure_reason_summary": "failed grasp on basket handle; empty retraction",
+    "vlm_recovery_action": "Re-align the gripper with the basket's handle or rim, secure a firm grip, and rotate/flip the basket upright.",
+    "vlm_recovery_action_summary": "re-grasp basket handle; flip basket upright",
+    "vlm_prevention_action": "Precisely align the gripper jaws over the basket handle before closing to ensure a secure grip before executing the flipping trajectory.",
+    "vlm_prevention_action_summary": "align gripper with handle; secure grip before flipping",
+    "vlm_justification": "At around 3.0 seconds, the robot's gripper closes near the handle of the basket but fails to grasp it, retracting empty and hovering unproductively for the remainder of the episode.",
+    "vlm_temporal_justification": "At timestamp 02:16 (frame 136), the robot gripper finishes closing without properly capturing or securing the wire handle of the basket. As the arm lifts shortly after, the handle remains on the table, making the grasp failure visibly apparent.",
+    "vlm_video_fps": 50.8,
+    "vlm_video_num_frames": 463,
+    "vlm_video_duration_seconds": 9.094488188976378,
+    "vlm_usage": {
+      "input_tokens": 21627,
+      "output_tokens": 374,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 3.0,\n \"failure_window_start_seconds\": 2.0,\n \"failure_window_end_seconds\": 3.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grab the handle of the upside-down basket to flip it, but failed to establish a grip and lifted away without manipulating the object.\",\n \"failure_reason_summary\": \"failed grasp on basket handle; empty retraction\",\n \"recovery_action\": \"Re-align the gripper with the basket's handle or rim, secure a firm grip, and rotate/flip the basket upright.\",\n \"recovery_action_summary\": \"re-grasp basket handle; flip basket upright\",\n \"prevention_action\": \"Precisely align the gripper jaws over the basket handle before closing to ensure a secure grip before executing the flipping trajectory.\",\n \"prevention_action_summary\": \"align gripper with handle; secure grip before flipping\",\n \"justification\": \"At around 3.0 seconds, the robot's gripper closes near the handle of the basket but fails to grasp it, retracting empty and hovering unproductively for the remainder of the episode.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 136,
+    "failure_sim_step": 136,
+    "failure_policy_step": 22,
+    "failure_chunk_index": 4,
+    "first_post_failure_policy_step": 23,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-align the gripper with the basket's handle or rim, secure a firm grip, and rotate/flip the basket upright."
+  }
+}
\ No newline at end of file
diff --git a/basket/ep015/labels.npz b/basket/ep015/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..a74c7f8ebbe4ad8d99470a5d63a3fabba91ff121
--- /dev/null
+++ b/basket/ep015/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e988cb64cfb79fef1ba3c6d9c82f20fd82972730f5ff3a9b52bf0be9374e7c24
+size 3829
diff --git a/basket/ep016/example.md b/basket/ep016/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..655c508cf550f6760966489e23ea32d527f385d5
--- /dev/null
+++ b/basket/ep016/example.md
@@ -0,0 +1,33 @@
+# t002__ep016__basket
+
+- **suite / variant:** `` / ``
+- **instruction:** Flip the basket upright
+- **success:** False
+- **control steps:** 332  (6.6s at 50 Hz)
+- **policy inferences:** 58  (replan_steps=6)
+- **video:** `basket/ep016/rollout.mp4`
+- **features:** `58` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching down to basket edge |
+| 3.0–6.0 | 152–303 | 25–50 | lifting empty gripper and retreating |
+| 6.0–6.6 | 304–331 | 50–55 | hovering above table |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 3.405940594059406 s (00:03.4)
+- **frame / control_step:** 172 (policy_step=28, chunk_index=4)
+- **coarse onset:** 2.8 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the edge of the inverted red basket, but failed to secure a grip or hook beneath it before lifting away, leaving the basket unmanipulated.
+  Summary: failed grasp on basket edge; lifting away without object
+- **recovery:** Re-align the gripper with the rim or edge of the red basket, secure a firm hold underneath, and flip the basket upright.
+  Summary: re-align gripper with basket rim; secure grip and flip upright
+- **prevention:** Position the gripper lower and further beneath the basket rim before closing the fingers to guarantee a secure grasp for flipping.
+  Summary: position gripper beneath basket rim; close fingers securely
+- **justification:** At approximately 2.8 seconds, the robot attempts to grasp the top edge of the basket, but fails to establish contact or a secure hold. It lifts away empty around 4.0 seconds, failing to flip the basket.
diff --git a/basket/ep016/labels.json b/basket/ep016/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..af3a7a8400ed67cbb27e9584a9874a5187a7b32d
--- /dev/null
+++ b/basket/ep016/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t002__ep016__basket",
+  "scene_variant": "",
+  "task_id": 2,
+  "task_file": "",
+  "instruction": "Flip the basket upright",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T19:37:03+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00990099009901,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching down to basket edge",
+      "description": "reaching down to basket edge"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00990099009901,
+      "t_end_sec": 6.01980198019802,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "lifting empty gripper and retreating",
+      "description": "lifting empty gripper and retreating"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01980198019802,
+      "t_end_sec": 6.574257425742574,
+      "control_step_start": 304,
+      "control_step_end": 331,
+      "policy_step_start": 50,
+      "policy_step_end": 55,
+      "phrase": "hovering above table",
+      "description": "hovering above table"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 3.405940594059406,
+    "vlm_failure_onset_timestamp": "00:03.4",
+    "vlm_failure_onset_frame": 172,
+    "vlm_failure_onset_step": 172,
+    "vlm_coarse_failure_onset_seconds": 2.8,
+    "vlm_failure_window_start_seconds": 2.0,
+    "vlm_failure_window_end_seconds": 4.0,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the edge of the inverted red basket, but failed to secure a grip or hook beneath it before lifting away, leaving the basket unmanipulated.",
+    "vlm_failure_reason_summary": "failed grasp on basket edge; lifting away without object",
+    "vlm_recovery_action": "Re-align the gripper with the rim or edge of the red basket, secure a firm hold underneath, and flip the basket upright.",
+    "vlm_recovery_action_summary": "re-align gripper with basket rim; secure grip and flip upright",
+    "vlm_prevention_action": "Position the gripper lower and further beneath the basket rim before closing the fingers to guarantee a secure grasp for flipping.",
+    "vlm_prevention_action_summary": "position gripper beneath basket rim; close fingers securely",
+    "vlm_justification": "At approximately 2.8 seconds, the robot attempts to grasp the top edge of the basket, but fails to establish contact or a secure hold. It lifts away empty around 4.0 seconds, failing to flip the basket.",
+    "vlm_temporal_justification": "At second 172 (02:52 in the clip), the robot gripper reaches down to the red basket to attempt insertion/hooking into the basket's top hole/edge. However, it fails to achieve any secure grip or leverage and lifts away starting around second 180 (03:00), leaving the basket untouched.",
+    "vlm_video_fps": 50.5,
+    "vlm_video_num_frames": 332,
+    "vlm_video_duration_seconds": 6.554455445544554,
+    "vlm_usage": {
+      "input_tokens": 20811,
+      "output_tokens": 409,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 2.8,\n \"failure_window_start_seconds\": 2.0,\n \"failure_window_end_seconds\": 4.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the edge of the inverted red basket, but failed to secure a grip or hook beneath it before lifting away, leaving the basket unmanipulated.\",\n \"failure_reason_summary\": \"failed grasp on basket edge; lifting away without object\",\n \"recovery_action\": \"Re-align the gripper with the rim or edge of the red basket, secure a firm hold underneath, and flip the basket upright.\",\n \"recovery_action_summary\": \"re-align gripper with basket rim; secure grip and flip upright\",\n \"prevention_action\": \"Position the gripper lower and further beneath the basket rim before closing the fingers to guarantee a secure grasp for flipping.\",\n \"prevention_action_summary\": \"position gripper beneath basket rim; close fingers securely\",\n \"justification\": \"At approximately 2.8 seconds, the robot attempts to grasp the top edge of the basket, but fails to establish contact or a secure hold. It lifts away empty around 4.0 seconds, failing to flip the basket.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 172,
+    "failure_sim_step": 172,
+    "failure_policy_step": 28,
+    "failure_chunk_index": 4,
+    "first_post_failure_policy_step": 29,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-align the gripper with the rim or edge of the red basket, secure a firm hold underneath, and flip the basket upright."
+  }
+}
\ No newline at end of file
diff --git a/basket/ep016/labels.npz b/basket/ep016/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..70178ab9f3961d8661a40e3a535e6302139bdb94
--- /dev/null
+++ b/basket/ep016/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0a5b4d551b6b29d0bb2c5ed48dc2ede30d727935bdd3c2b6c872f27c3aad97d4
+size 3867
diff --git a/basket/ep017/example.md b/basket/ep017/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..edaff3d268a89492d1bbd5df4534e22f30719036
--- /dev/null
+++ b/basket/ep017/example.md
@@ -0,0 +1,33 @@
+# t002__ep017__basket
+
+- **suite / variant:** `` / ``
+- **instruction:** Flip the basket upright
+- **success:** False
+- **control steps:** 386  (7.6s at 50 Hz)
+- **policy inferences:** 67  (replan_steps=6)
+- **video:** `basket/ep017/rollout.mp4`
+- **features:** `67` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | approaching basket; opening gripper |
+| 3.0–6.0 | 152–303 | 25–50 | hovering near basket |
+| 6.0–7.6 | 304–385 | 50–64 | remaining idle near basket |
+
+## Failure
+
+- **yes**
+- **onset type:** operator_intervention
+- **onset:** 1.8613861386138615 s (00:01.9)
+- **frame / control_step:** 94 (policy_step=15, chunk_index=4)
+- **coarse onset:** 3.0 s (refined=True)
+- **mode:** stuck_or_no_progress
+- **reason:** The robot moved near the red basket and opened its gripper, but then froze and hovered in place without attempting to grasp or flip the basket.
+  Summary: hovering near basket; freezing without attempting flip
+- **recovery:** Lower the gripper to make contact with the edge or strap of the red basket, secure a grip, and flip it upright.
+  Summary: grasp basket edge; flip basket upright
+- **prevention:** Lower the end-effector directly onto the basket edge or handle and execute the flipping maneuver continuously without pausing in mid-air.
+  Summary: lower onto basket edge; execute continuous flip
+- **justification:** By around 3.0 seconds into the rollout, the robot reaches a hovering position above the red basket with open fingers and ceases making progress, hovering passively until the rollout completes.
diff --git a/basket/ep017/labels.json b/basket/ep017/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..9c87270648888159fcf0e59febbfd69819512c8d
--- /dev/null
+++ b/basket/ep017/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t002__ep017__basket",
+  "scene_variant": "",
+  "task_id": 2,
+  "task_file": "",
+  "instruction": "Flip the basket upright",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T19:37:43+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00990099009901,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "approaching basket; opening gripper",
+      "description": "approaching basket; opening gripper"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00990099009901,
+      "t_end_sec": 6.01980198019802,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "hovering near basket",
+      "description": "hovering near basket"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01980198019802,
+      "t_end_sec": 7.643564356435643,
+      "control_step_start": 304,
+      "control_step_end": 385,
+      "policy_step_start": 50,
+      "policy_step_end": 64,
+      "phrase": "remaining idle near basket",
+      "description": "remaining idle near basket"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "stuck_or_no_progress",
+    "vlm_failure_onset_type": "operator_intervention",
+    "vlm_failure_onset_seconds": 1.8613861386138615,
+    "vlm_failure_onset_timestamp": "00:01.9",
+    "vlm_failure_onset_frame": 94,
+    "vlm_failure_onset_step": 94,
+    "vlm_coarse_failure_onset_seconds": 3.0,
+    "vlm_failure_window_start_seconds": 2.5,
+    "vlm_failure_window_end_seconds": 3.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot moved near the red basket and opened its gripper, but then froze and hovered in place without attempting to grasp or flip the basket.",
+    "vlm_failure_reason_summary": "hovering near basket; freezing without attempting flip",
+    "vlm_recovery_action": "Lower the gripper to make contact with the edge or strap of the red basket, secure a grip, and flip it upright.",
+    "vlm_recovery_action_summary": "grasp basket edge; flip basket upright",
+    "vlm_prevention_action": "Lower the end-effector directly onto the basket edge or handle and execute the flipping maneuver continuously without pausing in mid-air.",
+    "vlm_prevention_action_summary": "lower onto basket edge; execute continuous flip",
+    "vlm_justification": "By around 3.0 seconds into the rollout, the robot reaches a hovering position above the red basket with open fingers and ceases making progress, hovering passively until the rollout completes.",
+    "vlm_temporal_justification": "The robot arm moves towards the red basket and positions its open gripper above the handle by 01:33 (93 seconds). Starting at 01:34 (94 seconds), all arm movement ceases completely and the robot remains frozen hovering above the basket for the remainder of the rollout.",
+    "vlm_video_fps": 50.5,
+    "vlm_video_num_frames": 386,
+    "vlm_video_duration_seconds": 7.623762376237623,
+    "vlm_usage": {
+      "input_tokens": 21564,
+      "output_tokens": 375,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"stuck_or_no_progress\",\n \"onset_type\": \"operator_intervention\",\n \"failure_onset_seconds\": 3.0,\n \"failure_window_start_seconds\": 2.5,\n \"failure_window_end_seconds\": 3.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot moved near the red basket and opened its gripper, but then froze and hovered in place without attempting to grasp or flip the basket.\",\n \"failure_reason_summary\": \"hovering near basket; freezing without attempting flip\",\n \"recovery_action\": \"Lower the gripper to make contact with the edge or strap of the red basket, secure a grip, and flip it upright.\",\n \"recovery_action_summary\": \"grasp basket edge; flip basket upright\",\n \"prevention_action\": \"Lower the end-effector directly onto the basket edge or handle and execute the flipping maneuver continuously without pausing in mid-air.\",\n \"prevention_action_summary\": \"lower onto basket edge; execute continuous flip\",\n \"justification\": \"By around 3.0 seconds into the rollout, the robot reaches a hovering position above the red basket with open fingers and ceases making progress, hovering passively until the rollout completes.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 94,
+    "failure_sim_step": 94,
+    "failure_policy_step": 15,
+    "failure_chunk_index": 4,
+    "first_post_failure_policy_step": 16,
+    "failure_type": "stuck_or_no_progress",
+    "correction_action": "Lower the gripper to make contact with the edge or strap of the red basket, secure a grip, and flip it upright."
+  }
+}
\ No newline at end of file
diff --git a/basket/ep017/labels.npz b/basket/ep017/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..4aa1295f322b7a0e9ffa92ec42bb6c8514ea4364
--- /dev/null
+++ b/basket/ep017/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:10d7c6672bfa4798f14f830b8c13ef374858d7ae2fee106d401c913143b6fd17
+size 3785
diff --git a/basket/ep018/example.md b/basket/ep018/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..55b0b21e63d7b1719de44d254ea82237d0303721
--- /dev/null
+++ b/basket/ep018/example.md
@@ -0,0 +1,33 @@
+# t002__ep018__basket
+
+- **suite / variant:** `` / ``
+- **instruction:** Flip the basket upright
+- **success:** False
+- **control steps:** 431  (8.5s at 50 Hz)
+- **policy inferences:** 75  (replan_steps=6)
+- **video:** `basket/ep018/rollout.mp4`
+- **features:** `75` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching and grasping basket edge |
+| 3.0–6.0 | 152–303 | 25–50 | pressing down on basket top |
+| 6.0–8.5 | 304–430 | 50–71 | hovering idle over basket |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 2.4752475247524752 s (00:02.5)
+- **frame / control_step:** 125 (policy_step=20, chunk_index=5)
+- **coarse onset:** 2.5 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the edge or handle of the red basket to flip it, but the gripper slipped off without acquiring a firm hold.
+  Summary: failed grasp on basket handle; unable to flip
+- **recovery:** Release the basket, reposition the end-effector near the handle or outer rim, acquire a firm grip, and rotate it upright.
+  Summary: reposition gripper; grasp edge; flip upright
+- **prevention:** Align the gripper carefully with the basket handle or rim before closing the fingers to ensure a secure grip for flipping.
+  Summary: clamp handle securely; tilt and flip basket
+- **justification:** Around 2.5 seconds into the rollout, the robot attempts to grab the red basket's edge/handle to flip it over, but loses its grip and slips onto the top of the basket, leaving it upside down for the remainder of the episode.
diff --git a/basket/ep018/labels.json b/basket/ep018/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..90da1f9e732fb3830c8498271215dc332fa00ff2
--- /dev/null
+++ b/basket/ep018/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t002__ep018__basket",
+  "scene_variant": "",
+  "task_id": 2,
+  "task_file": "",
+  "instruction": "Flip the basket upright",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T19:38:17+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00990099009901,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching and grasping basket edge",
+      "description": "reaching and grasping basket edge"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00990099009901,
+      "t_end_sec": 6.01980198019802,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "pressing down on basket top",
+      "description": "pressing down on basket top"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01980198019802,
+      "t_end_sec": 8.534653465346535,
+      "control_step_start": 304,
+      "control_step_end": 430,
+      "policy_step_start": 50,
+      "policy_step_end": 71,
+      "phrase": "hovering idle over basket",
+      "description": "hovering idle over basket"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 2.4752475247524752,
+    "vlm_failure_onset_timestamp": "00:02.5",
+    "vlm_failure_onset_frame": 125,
+    "vlm_failure_onset_step": 125,
+    "vlm_coarse_failure_onset_seconds": 2.5,
+    "vlm_failure_window_start_seconds": 2.0,
+    "vlm_failure_window_end_seconds": 3.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the edge or handle of the red basket to flip it, but the gripper slipped off without acquiring a firm hold.",
+    "vlm_failure_reason_summary": "failed grasp on basket handle; unable to flip",
+    "vlm_recovery_action": "Release the basket, reposition the end-effector near the handle or outer rim, acquire a firm grip, and rotate it upright.",
+    "vlm_recovery_action_summary": "reposition gripper; grasp edge; flip upright",
+    "vlm_prevention_action": "Align the gripper carefully with the basket handle or rim before closing the fingers to ensure a secure grip for flipping.",
+    "vlm_prevention_action_summary": "clamp handle securely; tilt and flip basket",
+    "vlm_justification": "Around 2.5 seconds into the rollout, the robot attempts to grab the red basket's edge/handle to flip it over, but loses its grip and slips onto the top of the basket, leaving it upside down for the remainder of the episode.",
+    "vlm_temporal_justification": "The robot attempts to grasp the basket handle/rim and closes its gripper around second 100 (01:40). At second 125 (02:05), as the arm begins pulling upward to flip the basket, the gripper slips off the handle without securing a hold, leaving the basket on the table.",
+    "vlm_video_fps": 50.5,
+    "vlm_video_num_frames": 431,
+    "vlm_video_duration_seconds": 8.514851485148515,
+    "vlm_usage": {
+      "input_tokens": 19986,
+      "output_tokens": 396,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 2.5,\n \"failure_window_start_seconds\": 2.0,\n \"failure_window_end_seconds\": 3.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the edge or handle of the red basket to flip it, but the gripper slipped off without acquiring a firm hold.\",\n \"failure_reason_summary\": \"failed grasp on basket handle; unable to flip\",\n \"recovery_action\": \"Release the basket, reposition the end-effector near the handle or outer rim, acquire a firm grip, and rotate it upright.\",\n \"recovery_action_summary\": \"reposition gripper; grasp edge; flip upright\",\n \"prevention_action\": \"Align the gripper carefully with the basket handle or rim before closing the fingers to ensure a secure grip for flipping.\",\n \"prevention_action_summary\": \"clamp handle securely; tilt and flip basket\",\n \"justification\": \"Around 2.5 seconds into the rollout, the robot attempts to grab the red basket's edge/handle to flip it over, but loses its grip and slips onto the top of the basket, leaving it upside down for the remainder of the episode.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 125,
+    "failure_sim_step": 125,
+    "failure_policy_step": 20,
+    "failure_chunk_index": 5,
+    "first_post_failure_policy_step": 21,
+    "failure_type": "grasp_failure",
+    "correction_action": "Release the basket, reposition the end-effector near the handle or outer rim, acquire a firm grip, and rotate it upright."
+  }
+}
\ No newline at end of file
diff --git a/basket/ep018/labels.npz b/basket/ep018/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..30a08d176dede5ed01ed47a9ffdaaf8131d8723d
--- /dev/null
+++ b/basket/ep018/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:736fc7b348da8c293f9b5e21d8c17bde0dc620afaeaccc7113667a6f71fdcd9c
+size 3778
diff --git a/basket/ep019/example.md b/basket/ep019/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..9e7b93de0c5d29dbbfcc625e76429ec2010896ce
--- /dev/null
+++ b/basket/ep019/example.md
@@ -0,0 +1,34 @@
+# t002__ep019__basket
+
+- **suite / variant:** `` / ``
+- **instruction:** Flip the basket upright
+- **success:** False
+- **control steps:** 473  (9.4s at 50 Hz)
+- **policy inferences:** 81  (replan_steps=6)
+- **video:** `basket/ep019/rollout.mp4`
+- **features:** `81` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–149 | 0–24 | approaching basket handle |
+| 3.0–6.0 | 150–299 | 25–49 | grasping handle; dropping basket |
+| 6.0–9.0 | 300–449 | 50–74 | releasing grip; lifting arm |
+| 9.0–9.4 | 450–472 | 75–78 | hovering above table |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 4.790419161676646 s (00:04.8)
+- **frame / control_step:** 240 (policy_step=40, chunk_index=0)
+- **coarse onset:** 5.0 s (refined=True)
+- **mode:** object_displacement
+- **reason:** The robot grasped the basket handle to flip it, but lost its grip mid-lift and dropped the basket back onto the table in an inverted position.
+  Summary: losing grip on handle; dropping basket
+- **recovery:** Re-grasp the handle securely and lift while rotating the basket until it rests right-side up on the table.
+  Summary: re-grasp handle; flip basket upright
+- **prevention:** Maintain a firm grip on the handle throughout the upward and rotational motion to ensure the basket is fully flipped upright.
+  Summary: secure handle grasp; complete rotational flip
+- **justification:** At approximately 5.0 seconds, while trying to lift and rotate the basket upright, the end effector slips off the handle, dropping the basket onto its side/top rather than completing the flip.
diff --git a/basket/ep019/labels.json b/basket/ep019/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..9f237890a47e4e0261cbfb9a5e1711ddf1a0fc81
--- /dev/null
+++ b/basket/ep019/labels.json
@@ -0,0 +1,109 @@
+{
+  "rollout_id": "t002__ep019__basket",
+  "scene_variant": "",
+  "task_id": 2,
+  "task_file": "",
+  "instruction": "Flip the basket upright",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T19:38:55+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9940119760479043,
+      "control_step_start": 0,
+      "control_step_end": 149,
+      "policy_step_start": 0,
+      "policy_step_end": 24,
+      "phrase": "approaching basket handle",
+      "description": "approaching basket handle"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9940119760479043,
+      "t_end_sec": 5.9880239520958085,
+      "control_step_start": 150,
+      "control_step_end": 299,
+      "policy_step_start": 25,
+      "policy_step_end": 49,
+      "phrase": "grasping handle; dropping basket",
+      "description": "grasping handle; dropping basket"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.9880239520958085,
+      "t_end_sec": 8.982035928143713,
+      "control_step_start": 300,
+      "control_step_end": 449,
+      "policy_step_start": 50,
+      "policy_step_end": 74,
+      "phrase": "releasing grip; lifting arm",
+      "description": "releasing grip; lifting arm"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.982035928143713,
+      "t_end_sec": 9.441117764471057,
+      "control_step_start": 450,
+      "control_step_end": 472,
+      "policy_step_start": 75,
+      "policy_step_end": 78,
+      "phrase": "hovering above table",
+      "description": "hovering above table"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "object_displacement",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 4.790419161676646,
+    "vlm_failure_onset_timestamp": "00:04.8",
+    "vlm_failure_onset_frame": 240,
+    "vlm_failure_onset_step": 240,
+    "vlm_coarse_failure_onset_seconds": 5.0,
+    "vlm_failure_window_start_seconds": 4.8,
+    "vlm_failure_window_end_seconds": 5.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot grasped the basket handle to flip it, but lost its grip mid-lift and dropped the basket back onto the table in an inverted position.",
+    "vlm_failure_reason_summary": "losing grip on handle; dropping basket",
+    "vlm_recovery_action": "Re-grasp the handle securely and lift while rotating the basket until it rests right-side up on the table.",
+    "vlm_recovery_action_summary": "re-grasp handle; flip basket upright",
+    "vlm_prevention_action": "Maintain a firm grip on the handle throughout the upward and rotational motion to ensure the basket is fully flipped upright.",
+    "vlm_prevention_action_summary": "secure handle grasp; complete rotational flip",
+    "vlm_justification": "At approximately 5.0 seconds, while trying to lift and rotate the basket upright, the end effector slips off the handle, dropping the basket onto its side/top rather than completing the flip.",
+    "vlm_temporal_justification": "At 02:20 (140s into the clip), the robot lifts the basket by its handle to flip it, but loses its grip immediately at 02:21, dropping the basket back down onto the table in an upside-down position.",
+    "vlm_video_fps": 50.1,
+    "vlm_video_num_frames": 473,
+    "vlm_video_duration_seconds": 9.421157684630739,
+    "vlm_usage": {
+      "input_tokens": 21380,
+      "output_tokens": 367,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"object_displacement\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 5.0,\n \"failure_window_start_seconds\": 4.8,\n \"failure_window_end_seconds\": 5.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot grasped the basket handle to flip it, but lost its grip mid-lift and dropped the basket back onto the table in an inverted position.\",\n \"failure_reason_summary\": \"losing grip on handle; dropping basket\",\n \"recovery_action\": \"Re-grasp the handle securely and lift while rotating the basket until it rests right-side up on the table.\",\n \"recovery_action_summary\": \"re-grasp handle; flip basket upright\",\n \"prevention_action\": \"Maintain a firm grip on the handle throughout the upward and rotational motion to ensure the basket is fully flipped upright.\",\n \"prevention_action_summary\": \"secure handle grasp; complete rotational flip\",\n \"justification\": \"At approximately 5.0 seconds, while trying to lift and rotate the basket upright, the end effector slips off the handle, dropping the basket onto its side/top rather than completing the flip.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 240,
+    "failure_sim_step": 240,
+    "failure_policy_step": 40,
+    "failure_chunk_index": 0,
+    "first_post_failure_policy_step": 41,
+    "failure_type": "object_displacement",
+    "correction_action": "Re-grasp the handle securely and lift while rotating the basket until it rests right-side up on the table."
+  }
+}
\ No newline at end of file
diff --git a/basket/ep019/labels.npz b/basket/ep019/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..91be9f1a2f37e845b740485872d7a3f506cf145c
--- /dev/null
+++ b/basket/ep019/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:09ca5d3367a6106d4d4e5474c56784de4ff613365b8898ed5145127d4c9f1c2b
+size 3798
diff --git a/basket/ep020/example.md b/basket/ep020/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..d2098fee063c2d0a2c0681ec271ad4f2ba0d51e1
--- /dev/null
+++ b/basket/ep020/example.md
@@ -0,0 +1,32 @@
+# t002__ep020__basket
+
+- **suite / variant:** `` / ``
+- **instruction:** Flip the basket upright
+- **success:** False
+- **control steps:** 242  (4.8s at 50 Hz)
+- **policy inferences:** 42  (replan_steps=6)
+- **video:** `basket/ep020/rollout.mp4`
+- **features:** `42` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | approaching and attempting handle grasp |
+| 3.0–4.8 | 151–241 | 25–40 | retracting gripper without object |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 2.485089463220676 s (00:02.5)
+- **frame / control_step:** 125 (policy_step=20, chunk_index=5)
+- **coarse onset:** 2.5 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the basket's handle to flip it upright, but failed to secure a firm grip and slipped off without moving the basket.
+  Summary: unsuccessful grasp on handle; gripper slipping
+- **recovery:** Re-align the gripper directly over the handle or rim of the red basket, establish a firm grip, and rotate the basket upright.
+  Summary: re-align gripper; secure grip on handle; flip basket upright
+- **prevention:** Position the end-effector accurately over the handle or lip of the basket and close the fingers tightly before attempting to lift and flip it.
+  Summary: align gripper properly; grasp handle securely; rotate basket upright
+- **justification:** At 2.5 seconds, the end-effector closes near the handle of the upside-down basket but fails to grasp it securely. As the arm moves up and rotates, the gripper slips off without flipping the basket, leaving it in its original state.
diff --git a/basket/ep020/labels.json b/basket/ep020/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..31fc17b0ed08762585ad569926dc735ddb0588a7
--- /dev/null
+++ b/basket/ep020/labels.json
@@ -0,0 +1,87 @@
+{
+  "rollout_id": "t002__ep020__basket",
+  "scene_variant": "",
+  "task_id": 2,
+  "task_file": "",
+  "instruction": "Flip the basket upright",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T19:39:32+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.001988071570577,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "approaching and attempting handle grasp",
+      "description": "approaching and attempting handle grasp"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.001988071570577,
+      "t_end_sec": 4.811133200795229,
+      "control_step_start": 151,
+      "control_step_end": 241,
+      "policy_step_start": 25,
+      "policy_step_end": 40,
+      "phrase": "retracting gripper without object",
+      "description": "retracting gripper without object"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 2.485089463220676,
+    "vlm_failure_onset_timestamp": "00:02.5",
+    "vlm_failure_onset_frame": 125,
+    "vlm_failure_onset_step": 125,
+    "vlm_coarse_failure_onset_seconds": 2.5,
+    "vlm_failure_window_start_seconds": 2.0,
+    "vlm_failure_window_end_seconds": 3.0,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the basket's handle to flip it upright, but failed to secure a firm grip and slipped off without moving the basket.",
+    "vlm_failure_reason_summary": "unsuccessful grasp on handle; gripper slipping",
+    "vlm_recovery_action": "Re-align the gripper directly over the handle or rim of the red basket, establish a firm grip, and rotate the basket upright.",
+    "vlm_recovery_action_summary": "re-align gripper; secure grip on handle; flip basket upright",
+    "vlm_prevention_action": "Position the end-effector accurately over the handle or lip of the basket and close the fingers tightly before attempting to lift and flip it.",
+    "vlm_prevention_action_summary": "align gripper properly; grasp handle securely; rotate basket upright",
+    "vlm_justification": "At 2.5 seconds, the end-effector closes near the handle of the upside-down basket but fails to grasp it securely. As the arm moves up and rotates, the gripper slips off without flipping the basket, leaving it in its original state.",
+    "vlm_temporal_justification": "At 02:05 (125 seconds into the clip), the gripper fingers finish closing completely without securing the basket's black handle, confirming the grasp failure as the handle slips out before the robot attempts to lift.",
+    "vlm_video_fps": 50.3,
+    "vlm_video_num_frames": 242,
+    "vlm_video_duration_seconds": 4.791252485089464,
+    "vlm_usage": {
+      "input_tokens": 17405,
+      "output_tokens": 388,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 2.5,\n \"failure_window_start_seconds\": 2.0,\n \"failure_window_end_seconds\": 3.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the basket's handle to flip it upright, but failed to secure a firm grip and slipped off without moving the basket.\",\n \"failure_reason_summary\": \"unsuccessful grasp on handle; gripper slipping\",\n \"recovery_action\": \"Re-align the gripper directly over the handle or rim of the red basket, establish a firm grip, and rotate the basket upright.\",\n \"recovery_action_summary\": \"re-align gripper; secure grip on handle; flip basket upright\",\n \"prevention_action\": \"Position the end-effector accurately over the handle or lip of the basket and close the fingers tightly before attempting to lift and flip it.\",\n \"prevention_action_summary\": \"align gripper properly; grasp handle securely; rotate basket upright\",\n \"justification\": \"At 2.5 seconds, the end-effector closes near the handle of the upside-down basket but fails to grasp it securely. As the arm moves up and rotates, the gripper slips off without flipping the basket, leaving it in its original state.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 125,
+    "failure_sim_step": 125,
+    "failure_policy_step": 20,
+    "failure_chunk_index": 5,
+    "first_post_failure_policy_step": 21,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-align the gripper directly over the handle or rim of the red basket, establish a firm grip, and rotate the basket upright."
+  }
+}
\ No newline at end of file
diff --git a/basket/ep020/labels.npz b/basket/ep020/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..1bca020af2196d99a7dd03c41f1f56cf76188c3d
--- /dev/null
+++ b/basket/ep020/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:84c19895723be31098a9711a26c99bf0eab6729d544bd0b57c1f53818596b902
+size 3799
diff --git a/basket/ep021/example.md b/basket/ep021/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..60514351b2c9555e813f89c3d65bc1260dec7f39
--- /dev/null
+++ b/basket/ep021/example.md
@@ -0,0 +1,33 @@
+# t002__ep021__basket
+
+- **suite / variant:** `` / ``
+- **instruction:** Flip the basket upright
+- **success:** False
+- **control steps:** 311  (6.1s at 51 Hz)
+- **policy inferences:** 53  (replan_steps=6)
+- **video:** `basket/ep021/rollout.mp4`
+- **features:** `53` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–152 | 0–25 | approaching and grasping red basket |
+| 3.0–6.0 | 153–305 | 25–50 | slipping off basket and lifting |
+| 6.0–6.1 | 306–310 | 51–51 | idle above red basket |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 2.986247544204322 s (00:03.0)
+- **frame / control_step:** 152 (policy_step=25, chunk_index=2)
+- **coarse onset:** 3.2 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot end-effector lowered to grip the red basket but failed to establish a secure hold on its edge or handle, slipping off without picking up or flipping the basket.
+  Summary: failed grasp on basket; slipped off edge
+- **recovery:** Re-align the gripper over the rim of the red basket, close the fingers firmly to establish a secure hold, and rotate the basket to flip it upright.
+  Summary: reposition gripper; grasp basket rim; flip basket upright
+- **prevention:** Precisely align the gripper jaws with the rim or handle of the red basket before closing them to ensure a secure grip before applying rotation.
+  Summary: align gripper with rim; secure firm grip; flip basket
+- **justification:** None
diff --git a/basket/ep021/labels.json b/basket/ep021/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..cc7cfb1b6bc425a001c08a323d776132df335dc4
--- /dev/null
+++ b/basket/ep021/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t002__ep021__basket",
+  "scene_variant": "",
+  "task_id": 2,
+  "task_file": "",
+  "instruction": "Flip the basket upright",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T19:40:03+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.005893909626719,
+      "control_step_start": 0,
+      "control_step_end": 152,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "approaching and grasping red basket",
+      "description": "approaching and grasping red basket"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.005893909626719,
+      "t_end_sec": 6.011787819253438,
+      "control_step_start": 153,
+      "control_step_end": 305,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "slipping off basket and lifting",
+      "description": "slipping off basket and lifting"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.011787819253438,
+      "t_end_sec": 6.110019646365423,
+      "control_step_start": 306,
+      "control_step_end": 310,
+      "policy_step_start": 51,
+      "policy_step_end": 51,
+      "phrase": "idle above red basket",
+      "description": "idle above red basket"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 2.986247544204322,
+    "vlm_failure_onset_timestamp": "00:03.0",
+    "vlm_failure_onset_frame": 152,
+    "vlm_failure_onset_step": 152,
+    "vlm_coarse_failure_onset_seconds": 3.2,
+    "vlm_failure_window_start_seconds": 2.7,
+    "vlm_failure_window_end_seconds": 3.7,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot end-effector lowered to grip the red basket but failed to establish a secure hold on its edge or handle, slipping off without picking up or flipping the basket.",
+    "vlm_failure_reason_summary": "failed grasp on basket; slipped off edge",
+    "vlm_recovery_action": "Re-align the gripper over the rim of the red basket, close the fingers firmly to establish a secure hold, and rotate the basket to flip it upright.",
+    "vlm_recovery_action_summary": "reposition gripper; grasp basket rim; flip basket upright",
+    "vlm_prevention_action": "Precisely align the gripper jaws with the rim or handle of the red basket before closing them to ensure a secure grip before applying rotation.",
+    "vlm_prevention_action_summary": "align gripper with rim; secure firm grip; flip basket",
+    "vlm_justification": null,
+    "vlm_temporal_justification": "The robot positions its end-effector over the black loop handle and closes its gripper around 02:20. By 02:22 (second 142), the gripper fingers close completely without clamping onto the handle or basket edge, leaving the object untouched and resulting in a grasp failure.",
+    "vlm_video_fps": 50.9,
+    "vlm_video_num_frames": 311,
+    "vlm_video_duration_seconds": 6.090373280943026,
+    "vlm_usage": {
+      "input_tokens": 21192,
+      "output_tokens": 353,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 3.2,\n \"failure_window_start_seconds\": 2.7,\n \"failure_window_end_seconds\": 3.7,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot end-effector lowered to grip the red basket but failed to establish a secure hold on its edge or handle, slipping off without picking up or flipping the basket.\",\n \"failure_reason_summary\": \"failed grasp on basket; slipped off edge\",\n \"recovery_action\": \"Re-align the gripper over the rim of the red basket, close the fingers firmly to establish a secure hold, and rotate the basket to flip it upright.\",\n \"recovery_action_summary\": \"reposition gripper; grasp basket rim; flip basket upright\",\n \"prevention_action\": \"Precisely align the gripper jaws with the rim or handle of the red basket before closing them to ensure a secure grip before applying rotation.\",\n \"prevention_action_summary\": \"align gripper with rim; secure firm grip; flip basket\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 152,
+    "failure_sim_step": 152,
+    "failure_policy_step": 25,
+    "failure_chunk_index": 2,
+    "first_post_failure_policy_step": 26,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-align the gripper over the rim of the red basket, close the fingers firmly to establish a secure hold, and rotate the basket to flip it upright."
+  }
+}
\ No newline at end of file
diff --git a/basket/ep021/labels.npz b/basket/ep021/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..6ddbc686796735ef9188ad0df9f92b4e32ea5b71
--- /dev/null
+++ b/basket/ep021/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c35a0bfa3d889bb54d892781af934153212c6d85d007e578c632b79f2aa33b2a
+size 3842
diff --git a/basket/ep022/example.md b/basket/ep022/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..0fff532e4f4b8209e05da1f36dd8415fe352cef8
--- /dev/null
+++ b/basket/ep022/example.md
@@ -0,0 +1,33 @@
+# t002__ep022__basket
+
+- **suite / variant:** `` / ``
+- **instruction:** Flip the basket upright
+- **success:** False
+- **control steps:** 413  (8.2s at 50 Hz)
+- **policy inferences:** 71  (replan_steps=6)
+- **video:** `basket/ep022/rollout.mp4`
+- **features:** `71` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | approaching red basket |
+| 3.0–6.0 | 152–303 | 25–50 | attempting grasp on basket |
+| 6.0–8.2 | 304–412 | 50–68 | hovering above table |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 3.405940594059406 s (00:03.4)
+- **frame / control_step:** 172 (policy_step=28, chunk_index=4)
+- **coarse onset:** 3.0 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the rim or handle of the red basket to flip it, but failed to secure a firm grip, causing the gripper to slip off empty.
+  Summary: failed grasp on basket; fingers slipped off rim
+- **recovery:** Reposition the end-effector over the edge of the basket, open the gripper, establish a firm grip on the structure, and lift to flip it upright.
+  Summary: reposition gripper; grasp basket edge; flip basket upright
+- **prevention:** Align the gripper precisely with the rim or handle of the basket before closing the fingers to ensure a reliable hold prior to lifting.
+  Summary: align gripper properly; secure firm grip; flip basket
+- **justification:** Around 3.0 seconds, the robot's gripper attempts to grasp the red basket but fails to pinch the edge securely. The fingers slip off and close without holding the object, preventing the robot from flipping the basket.
diff --git a/basket/ep022/labels.json b/basket/ep022/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..065cdafb02b54127ec7cab297fbcdf31c450576e
--- /dev/null
+++ b/basket/ep022/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t002__ep022__basket",
+  "scene_variant": "",
+  "task_id": 2,
+  "task_file": "",
+  "instruction": "Flip the basket upright",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T19:40:37+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00990099009901,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "approaching red basket",
+      "description": "approaching red basket"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00990099009901,
+      "t_end_sec": 6.01980198019802,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "attempting grasp on basket",
+      "description": "attempting grasp on basket"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01980198019802,
+      "t_end_sec": 8.178217821782178,
+      "control_step_start": 304,
+      "control_step_end": 412,
+      "policy_step_start": 50,
+      "policy_step_end": 68,
+      "phrase": "hovering above table",
+      "description": "hovering above table"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 3.405940594059406,
+    "vlm_failure_onset_timestamp": "00:03.4",
+    "vlm_failure_onset_frame": 172,
+    "vlm_failure_onset_step": 172,
+    "vlm_coarse_failure_onset_seconds": 3.0,
+    "vlm_failure_window_start_seconds": 2.5,
+    "vlm_failure_window_end_seconds": 3.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the rim or handle of the red basket to flip it, but failed to secure a firm grip, causing the gripper to slip off empty.",
+    "vlm_failure_reason_summary": "failed grasp on basket; fingers slipped off rim",
+    "vlm_recovery_action": "Reposition the end-effector over the edge of the basket, open the gripper, establish a firm grip on the structure, and lift to flip it upright.",
+    "vlm_recovery_action_summary": "reposition gripper; grasp basket edge; flip basket upright",
+    "vlm_prevention_action": "Align the gripper precisely with the rim or handle of the basket before closing the fingers to ensure a reliable hold prior to lifting.",
+    "vlm_prevention_action_summary": "align gripper properly; secure firm grip; flip basket",
+    "vlm_justification": "Around 3.0 seconds, the robot's gripper attempts to grasp the red basket but fails to pinch the edge securely. The fingers slip off and close without holding the object, preventing the robot from flipping the basket.",
+    "vlm_temporal_justification": "At 02:52 (second 172 in the video clip), the robot gripper closes completely around the handle area but fails to secure a grip, slipping off empty as it begins to lift.",
+    "vlm_video_fps": 50.5,
+    "vlm_video_num_frames": 413,
+    "vlm_video_duration_seconds": 8.158415841584159,
+    "vlm_usage": {
+      "input_tokens": 21566,
+      "output_tokens": 381,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 3.0,\n \"failure_window_start_seconds\": 2.5,\n \"failure_window_end_seconds\": 3.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the rim or handle of the red basket to flip it, but failed to secure a firm grip, causing the gripper to slip off empty.\",\n \"failure_reason_summary\": \"failed grasp on basket; fingers slipped off rim\",\n \"recovery_action\": \"Reposition the end-effector over the edge of the basket, open the gripper, establish a firm grip on the structure, and lift to flip it upright.\",\n \"recovery_action_summary\": \"reposition gripper; grasp basket edge; flip basket upright\",\n \"prevention_action\": \"Align the gripper precisely with the rim or handle of the basket before closing the fingers to ensure a reliable hold prior to lifting.\",\n \"prevention_action_summary\": \"align gripper properly; secure firm grip; flip basket\",\n \"justification\": \"Around 3.0 seconds, the robot's gripper attempts to grasp the red basket but fails to pinch the edge securely. The fingers slip off and close without holding the object, preventing the robot from flipping the basket.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 172,
+    "failure_sim_step": 172,
+    "failure_policy_step": 28,
+    "failure_chunk_index": 4,
+    "first_post_failure_policy_step": 29,
+    "failure_type": "grasp_failure",
+    "correction_action": "Reposition the end-effector over the edge of the basket, open the gripper, establish a firm grip on the structure, and lift to flip it upright."
+  }
+}
\ No newline at end of file
diff --git a/basket/ep022/labels.npz b/basket/ep022/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..1ce4abab44006f71ec4a2472454f16319bbce4b3
--- /dev/null
+++ b/basket/ep022/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d4f53586c52c62ebc1fb56a2e5c41f2b3562864af60d45283777243ca603872f
+size 3819
diff --git a/basket/ep023/example.md b/basket/ep023/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..064682ad8b5427bb42a9c53d14b186f476d8feb1
--- /dev/null
+++ b/basket/ep023/example.md
@@ -0,0 +1,33 @@
+# t002__ep023__basket
+
+- **suite / variant:** `` / ``
+- **instruction:** Flip the basket upright
+- **success:** False
+- **control steps:** 405  (8.0s at 50 Hz)
+- **policy inferences:** 69  (replan_steps=6)
+- **video:** `basket/ep023/rollout.mp4`
+- **features:** `69` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | lowering arm toward basket handle |
+| 3.0–6.0 | 151–301 | 25–50 | attempting grasp and slipping off |
+| 6.0–8.0 | 302–404 | 50–67 | hovering above basket |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 3.373015873015873 s (00:03.4)
+- **frame / control_step:** 170 (policy_step=28, chunk_index=2)
+- **coarse onset:** 3.5 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the basket handle to flip it upright, but the gripper closed improperly and slipped off the handle without securing a firm grip.
+  Summary: failed grasp on basket handle; gripper slippage
+- **recovery:** Re-align the gripper directly over the basket handle, close the fingers firmly to secure the grip, and then execute the rotation to flip the basket upright.
+  Summary: re-align gripper; secure basket handle; flip upright
+- **prevention:** Center the end-effector accurately over the basket handle and ensure a firm grasp is established before applying lift or rotation forces.
+  Summary: center gripper over handle; establish firm grasp before rotating
+- **justification:** At approximately 3.5 seconds, the end-effector attempts to close on the red basket's handle to flip it, but slips off due to an insecure grasp, leaving the basket unchanged for the remainder of the episode.
diff --git a/basket/ep023/labels.json b/basket/ep023/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..6522e50c2f7f8ec472ee2340c228b809a3443ba6
--- /dev/null
+++ b/basket/ep023/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t002__ep023__basket",
+  "scene_variant": "",
+  "task_id": 2,
+  "task_file": "",
+  "instruction": "Flip the basket upright",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T19:41:10+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.996031746031746,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "lowering arm toward basket handle",
+      "description": "lowering arm toward basket handle"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.996031746031746,
+      "t_end_sec": 5.992063492063492,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "attempting grasp and slipping off",
+      "description": "attempting grasp and slipping off"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.992063492063492,
+      "t_end_sec": 8.035714285714286,
+      "control_step_start": 302,
+      "control_step_end": 404,
+      "policy_step_start": 50,
+      "policy_step_end": 67,
+      "phrase": "hovering above basket",
+      "description": "hovering above basket"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 3.373015873015873,
+    "vlm_failure_onset_timestamp": "00:03.4",
+    "vlm_failure_onset_frame": 170,
+    "vlm_failure_onset_step": 170,
+    "vlm_coarse_failure_onset_seconds": 3.5,
+    "vlm_failure_window_start_seconds": 3.0,
+    "vlm_failure_window_end_seconds": 4.0,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the basket handle to flip it upright, but the gripper closed improperly and slipped off the handle without securing a firm grip.",
+    "vlm_failure_reason_summary": "failed grasp on basket handle; gripper slippage",
+    "vlm_recovery_action": "Re-align the gripper directly over the basket handle, close the fingers firmly to secure the grip, and then execute the rotation to flip the basket upright.",
+    "vlm_recovery_action_summary": "re-align gripper; secure basket handle; flip upright",
+    "vlm_prevention_action": "Center the end-effector accurately over the basket handle and ensure a firm grasp is established before applying lift or rotation forces.",
+    "vlm_prevention_action_summary": "center gripper over handle; establish firm grasp before rotating",
+    "vlm_justification": "At approximately 3.5 seconds, the end-effector attempts to close on the red basket's handle to flip it, but slips off due to an insecure grasp, leaving the basket unchanged for the remainder of the episode.",
+    "vlm_temporal_justification": "At timestamp 02:25 (second 145 in the clip), the robot gripper begins closing to grasp the basket handle. However, due to poor positioning, the fingers close improperly around the wire loop, leading directly to the gripper slipping off without lifting or flipping the basket.",
+    "vlm_video_fps": 50.4,
+    "vlm_video_num_frames": 405,
+    "vlm_video_duration_seconds": 8.015873015873016,
+    "vlm_usage": {
+      "input_tokens": 21439,
+      "output_tokens": 394,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 3.5,\n \"failure_window_start_seconds\": 3.0,\n \"failure_window_end_seconds\": 4.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the basket handle to flip it upright, but the gripper closed improperly and slipped off the handle without securing a firm grip.\",\n \"failure_reason_summary\": \"failed grasp on basket handle; gripper slippage\",\n \"recovery_action\": \"Re-align the gripper directly over the basket handle, close the fingers firmly to secure the grip, and then execute the rotation to flip the basket upright.\",\n \"recovery_action_summary\": \"re-align gripper; secure basket handle; flip upright\",\n \"prevention_action\": \"Center the end-effector accurately over the basket handle and ensure a firm grasp is established before applying lift or rotation forces.\",\n \"prevention_action_summary\": \"center gripper over handle; establish firm grasp before rotating\",\n \"justification\": \"At approximately 3.5 seconds, the end-effector attempts to close on the red basket's handle to flip it, but slips off due to an insecure grasp, leaving the basket unchanged for the remainder of the episode.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 170,
+    "failure_sim_step": 170,
+    "failure_policy_step": 28,
+    "failure_chunk_index": 2,
+    "first_post_failure_policy_step": 29,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-align the gripper directly over the basket handle, close the fingers firmly to secure the grip, and then execute the rotation to flip the basket upright."
+  }
+}
\ No newline at end of file
diff --git a/basket/ep023/labels.npz b/basket/ep023/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..209e2adc1220739f17a613424b7b0dc291aac781
--- /dev/null
+++ b/basket/ep023/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:337c44dcf9faa1abf90cc10e7313dd0fcfd1d6ddc9b1cb131e086b1c9735bd19
+size 3858
diff --git a/basket/ep024/example.md b/basket/ep024/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..f5693c8f232263e10c73e0f32b497f2438f887f1
--- /dev/null
+++ b/basket/ep024/example.md
@@ -0,0 +1,33 @@
+# t002__ep024__basket
+
+- **suite / variant:** `` / ``
+- **instruction:** Flip the basket upright
+- **success:** False
+- **control steps:** 355  (7.0s at 51 Hz)
+- **policy inferences:** 62  (replan_steps=6)
+- **video:** `basket/ep024/rollout.mp4`
+- **features:** `62` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching for basket and gripping rim |
+| 3.0–6.0 | 152–303 | 25–50 | pressing on top of basket |
+| 6.0–7.0 | 304–354 | 50–59 | raising arm away from basket |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 2.3320158102766797 s (00:02.3)
+- **frame / control_step:** 118 (policy_step=19, chunk_index=4)
+- **coarse onset:** 2.3 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot fails to securely grasp the rim of the basket, slipping off the edge and pressing down on the side instead of flipping it.
+  Summary: failed grasp on basket rim; basket slipping
+- **recovery:** Re-align the gripper with the basket's edge, establish a firm grip on the rim, and rotate the basket into an upright position.
+  Summary: re-grasp basket rim; rotate basket upright
+- **prevention:** Position the end-effector accurately over the basket rim and clamp down firmly before executing the upward flipping trajectory.
+  Summary: align gripper with rim; secure grasp before flipping
+- **justification:** At around 2.3 seconds, the robot attempts to grasp the edge of the red basket to flip it, but fails to gain a secure grip. The gripper slips off the rim and simply presses down on the object, leaving it unflipped on the table.
diff --git a/basket/ep024/labels.json b/basket/ep024/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..4468a0e79de8537564eb9315bf7c7eeb9339e44a
--- /dev/null
+++ b/basket/ep024/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t002__ep024__basket",
+  "scene_variant": "",
+  "task_id": 2,
+  "task_file": "",
+  "instruction": "Flip the basket upright",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T19:41:45+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0039525691699605,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for basket and gripping rim",
+      "description": "reaching for basket and gripping rim"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0039525691699605,
+      "t_end_sec": 6.007905138339921,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "pressing on top of basket",
+      "description": "pressing on top of basket"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.007905138339921,
+      "t_end_sec": 7.015810276679842,
+      "control_step_start": 304,
+      "control_step_end": 354,
+      "policy_step_start": 50,
+      "policy_step_end": 59,
+      "phrase": "raising arm away from basket",
+      "description": "raising arm away from basket"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 2.3320158102766797,
+    "vlm_failure_onset_timestamp": "00:02.3",
+    "vlm_failure_onset_frame": 118,
+    "vlm_failure_onset_step": 118,
+    "vlm_coarse_failure_onset_seconds": 2.3,
+    "vlm_failure_window_start_seconds": 1.8,
+    "vlm_failure_window_end_seconds": 2.8,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot fails to securely grasp the rim of the basket, slipping off the edge and pressing down on the side instead of flipping it.",
+    "vlm_failure_reason_summary": "failed grasp on basket rim; basket slipping",
+    "vlm_recovery_action": "Re-align the gripper with the basket's edge, establish a firm grip on the rim, and rotate the basket into an upright position.",
+    "vlm_recovery_action_summary": "re-grasp basket rim; rotate basket upright",
+    "vlm_prevention_action": "Position the end-effector accurately over the basket rim and clamp down firmly before executing the upward flipping trajectory.",
+    "vlm_prevention_action_summary": "align gripper with rim; secure grasp before flipping",
+    "vlm_justification": "At around 2.3 seconds, the robot attempts to grasp the edge of the red basket to flip it, but fails to gain a secure grip. The gripper slips off the rim and simply presses down on the object, leaving it unflipped on the table.",
+    "vlm_temporal_justification": "At clip second 118 (01:58), the robot's gripper slips off the edge of the basket rim, failing to establish a firm grasp and instead pressing down on the side of the basket.",
+    "vlm_video_fps": 50.6,
+    "vlm_video_num_frames": 355,
+    "vlm_video_duration_seconds": 6.996047430830039,
+    "vlm_usage": {
+      "input_tokens": 19229,
+      "output_tokens": 374,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 2.3,\n \"failure_window_start_seconds\": 1.8,\n \"failure_window_end_seconds\": 2.8,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot fails to securely grasp the rim of the basket, slipping off the edge and pressing down on the side instead of flipping it.\",\n \"failure_reason_summary\": \"failed grasp on basket rim; basket slipping\",\n \"recovery_action\": \"Re-align the gripper with the basket's edge, establish a firm grip on the rim, and rotate the basket into an upright position.\",\n \"recovery_action_summary\": \"re-grasp basket rim; rotate basket upright\",\n \"prevention_action\": \"Position the end-effector accurately over the basket rim and clamp down firmly before executing the upward flipping trajectory.\",\n \"prevention_action_summary\": \"align gripper with rim; secure grasp before flipping\",\n \"justification\": \"At around 2.3 seconds, the robot attempts to grasp the edge of the red basket to flip it, but fails to gain a secure grip. The gripper slips off the rim and simply presses down on the object, leaving it unflipped on the table.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 118,
+    "failure_sim_step": 118,
+    "failure_policy_step": 19,
+    "failure_chunk_index": 4,
+    "first_post_failure_policy_step": 20,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-align the gripper with the basket's edge, establish a firm grip on the rim, and rotate the basket into an upright position."
+  }
+}
\ No newline at end of file
diff --git a/basket/ep024/labels.npz b/basket/ep024/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..6b25687b16abaf72fb150b01e9c6203ac6c9b6be
--- /dev/null
+++ b/basket/ep024/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b9b6e9050942b25a6783c726625bead741baa51f785edd3aa5f1e1698ea948e1
+size 3756
diff --git a/basket/ep025/example.md b/basket/ep025/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..fd184fb6b1c64704e09086bb46eb0a892ebeae68
--- /dev/null
+++ b/basket/ep025/example.md
@@ -0,0 +1,33 @@
+# t002__ep025__basket
+
+- **suite / variant:** `` / ``
+- **instruction:** Flip the basket upright
+- **success:** False
+- **control steps:** 325  (6.4s at 51 Hz)
+- **policy inferences:** 57  (replan_steps=6)
+- **video:** `basket/ep025/rollout.mp4`
+- **features:** `57` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching and attempting grasp on basket |
+| 3.0–6.0 | 152–303 | 25–50 | retracting and re-attempting grasp on basket |
+| 6.0–6.4 | 304–324 | 50–54 | retracting end-effector above table |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 1.5779092702169624 s (00:01.6)
+- **frame / control_step:** 80 (policy_step=13, chunk_index=2)
+- **coarse onset:** 2.1 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempts to grasp the handle of the red basket to flip it upright, but fails to achieve a secure grip and retracts repeatedly without moving the basket.
+  Summary: failed grasp on basket handle; repeated missing of object
+- **recovery:** Realign the end-effector over the red basket handle, close the gripper firmly around the handle, and lift to flip the basket upright.
+  Summary: reposition gripper; grasp basket handle; flip basket upright
+- **prevention:** Approach the basket handle with accurate gripper orientation and depth to secure a solid grasp before initiating the flip motion.
+  Summary: align gripper precisely; secure handle firmly; execute flip motion
+- **justification:** At 2.1 seconds, the robot's gripper reaches for the red basket handle but fails to establish a grip before retracting. It repeats another unsuccessful grasp attempt, failing to perform the required flipping task.
diff --git a/basket/ep025/labels.json b/basket/ep025/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..05d20ac33f8b5930f5787f96936b35eb6797da66
--- /dev/null
+++ b/basket/ep025/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t002__ep025__basket",
+  "scene_variant": "",
+  "task_id": 2,
+  "task_file": "",
+  "instruction": "Flip the basket upright",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T19:42:17+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9980276134122286,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching and attempting grasp on basket",
+      "description": "reaching and attempting grasp on basket"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9980276134122286,
+      "t_end_sec": 5.996055226824457,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "retracting and re-attempting grasp on basket",
+      "description": "retracting and re-attempting grasp on basket"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.996055226824457,
+      "t_end_sec": 6.41025641025641,
+      "control_step_start": 304,
+      "control_step_end": 324,
+      "policy_step_start": 50,
+      "policy_step_end": 54,
+      "phrase": "retracting end-effector above table",
+      "description": "retracting end-effector above table"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 1.5779092702169624,
+    "vlm_failure_onset_timestamp": "00:01.6",
+    "vlm_failure_onset_frame": 80,
+    "vlm_failure_onset_step": 80,
+    "vlm_coarse_failure_onset_seconds": 2.1,
+    "vlm_failure_window_start_seconds": 2.1,
+    "vlm_failure_window_end_seconds": 3.0,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempts to grasp the handle of the red basket to flip it upright, but fails to achieve a secure grip and retracts repeatedly without moving the basket.",
+    "vlm_failure_reason_summary": "failed grasp on basket handle; repeated missing of object",
+    "vlm_recovery_action": "Realign the end-effector over the red basket handle, close the gripper firmly around the handle, and lift to flip the basket upright.",
+    "vlm_recovery_action_summary": "reposition gripper; grasp basket handle; flip basket upright",
+    "vlm_prevention_action": "Approach the basket handle with accurate gripper orientation and depth to secure a solid grasp before initiating the flip motion.",
+    "vlm_prevention_action_summary": "align gripper precisely; secure handle firmly; execute flip motion",
+    "vlm_justification": "At 2.1 seconds, the robot's gripper reaches for the red basket handle but fails to establish a grip before retracting. It repeats another unsuccessful grasp attempt, failing to perform the required flipping task.",
+    "vlm_temporal_justification": "At second 80 (01:20), the robot completes its first grasp attempt on the basket's handle and begins lifting away, revealing that it failed to secure a grip on the handle.",
+    "vlm_video_fps": 50.7,
+    "vlm_video_num_frames": 325,
+    "vlm_video_duration_seconds": 6.390532544378698,
+    "vlm_usage": {
+      "input_tokens": 18541,
+      "output_tokens": 370,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 2.1,\n \"failure_window_start_seconds\": 2.1,\n \"failure_window_end_seconds\": 3.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempts to grasp the handle of the red basket to flip it upright, but fails to achieve a secure grip and retracts repeatedly without moving the basket.\",\n \"failure_reason_summary\": \"failed grasp on basket handle; repeated missing of object\",\n \"recovery_action\": \"Realign the end-effector over the red basket handle, close the gripper firmly around the handle, and lift to flip the basket upright.\",\n \"recovery_action_summary\": \"reposition gripper; grasp basket handle; flip basket upright\",\n \"prevention_action\": \"Approach the basket handle with accurate gripper orientation and depth to secure a solid grasp before initiating the flip motion.\",\n \"prevention_action_summary\": \"align gripper precisely; secure handle firmly; execute flip motion\",\n \"justification\": \"At 2.1 seconds, the robot's gripper reaches for the red basket handle but fails to establish a grip before retracting. It repeats another unsuccessful grasp attempt, failing to perform the required flipping task.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 80,
+    "failure_sim_step": 80,
+    "failure_policy_step": 13,
+    "failure_chunk_index": 2,
+    "first_post_failure_policy_step": 14,
+    "failure_type": "grasp_failure",
+    "correction_action": "Realign the end-effector over the red basket handle, close the gripper firmly around the handle, and lift to flip the basket upright."
+  }
+}
\ No newline at end of file
diff --git a/basket/ep025/labels.npz b/basket/ep025/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..2ab98e4de164c4f146bd4e19ee93ac331e972266
--- /dev/null
+++ b/basket/ep025/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1e4bf7baf5e619a5f7e0372ca4691b29d8d7685eec54344df66536d13c6ab90f
+size 3864
diff --git a/basket/ep026/example.md b/basket/ep026/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..8e47dd3c85879d19dd9a0b23a2d3a3ef1ed8e85f
--- /dev/null
+++ b/basket/ep026/example.md
@@ -0,0 +1,33 @@
+# t002__ep026__basket
+
+- **suite / variant:** `` / ``
+- **instruction:** Flip the basket upright
+- **success:** False
+- **control steps:** 362  (7.1s at 51 Hz)
+- **policy inferences:** 62  (replan_steps=6)
+- **video:** `basket/ep026/rollout.mp4`
+- **features:** `62` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–153 | 0–25 | reaching and grasping basket handle |
+| 3.0–6.0 | 154–307 | 25–51 | pressing down against basket |
+| 6.0–7.1 | 308–361 | 51–60 | hovering over basket |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 2.5536062378167643 s (00:02.6)
+- **frame / control_step:** 131 (policy_step=21, chunk_index=5)
+- **coarse onset:** 2.0 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the basket handle to flip it upright, but the gripper misaligned and failed to secure a grip, slipping off the basket.
+  Summary: improper grasp alignment; failure to secure handle grip
+- **recovery:** Reposition the end effector directly over the basket handle, close the gripper securely, and rotate the wrist to flip the basket upright.
+  Summary: reposition end effector; secure handle grasp; rotate basket upright
+- **prevention:** Align the gripper accurately with the basket handle and ensure a firm grip is established before initiating the flipping rotation.
+  Summary: align gripper with handle; establish firm grip; lift and flip
+- **justification:** At 2.0 seconds, the robot's gripper closes on the basket handle but fails to gain purchase, slipping off and pushing the basket down instead of lifting and flipping it upright.
diff --git a/basket/ep026/labels.json b/basket/ep026/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..742446d4d4283e250c8a9c8a5bce4532b99dacf0
--- /dev/null
+++ b/basket/ep026/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t002__ep026__basket",
+  "scene_variant": "",
+  "task_id": 2,
+  "task_file": "",
+  "instruction": "Flip the basket upright",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T19:42:46+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0019493177387915,
+      "control_step_start": 0,
+      "control_step_end": 153,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching and grasping basket handle",
+      "description": "reaching and grasping basket handle"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0019493177387915,
+      "t_end_sec": 6.003898635477583,
+      "control_step_start": 154,
+      "control_step_end": 307,
+      "policy_step_start": 25,
+      "policy_step_end": 51,
+      "phrase": "pressing down against basket",
+      "description": "pressing down against basket"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.003898635477583,
+      "t_end_sec": 7.056530214424952,
+      "control_step_start": 308,
+      "control_step_end": 361,
+      "policy_step_start": 51,
+      "policy_step_end": 60,
+      "phrase": "hovering over basket",
+      "description": "hovering over basket"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 2.5536062378167643,
+    "vlm_failure_onset_timestamp": "00:02.6",
+    "vlm_failure_onset_frame": 131,
+    "vlm_failure_onset_step": 131,
+    "vlm_coarse_failure_onset_seconds": 2.0,
+    "vlm_failure_window_start_seconds": 2.0,
+    "vlm_failure_window_end_seconds": 3.0,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the basket handle to flip it upright, but the gripper misaligned and failed to secure a grip, slipping off the basket.",
+    "vlm_failure_reason_summary": "improper grasp alignment; failure to secure handle grip",
+    "vlm_recovery_action": "Reposition the end effector directly over the basket handle, close the gripper securely, and rotate the wrist to flip the basket upright.",
+    "vlm_recovery_action_summary": "reposition end effector; secure handle grasp; rotate basket upright",
+    "vlm_prevention_action": "Align the gripper accurately with the basket handle and ensure a firm grip is established before initiating the flipping rotation.",
+    "vlm_prevention_action_summary": "align gripper with handle; establish firm grip; lift and flip",
+    "vlm_justification": "At 2.0 seconds, the robot's gripper closes on the basket handle but fails to gain purchase, slipping off and pushing the basket down instead of lifting and flipping it upright.",
+    "vlm_temporal_justification": "At 02:11 (frame 131), the robot closes its gripper, but because it is misaligned with the handle, it fails to secure a grip on the basket handle and instead pinches onto the outer surface of the basket.",
+    "vlm_video_fps": 51.3,
+    "vlm_video_num_frames": 362,
+    "vlm_video_duration_seconds": 7.037037037037037,
+    "vlm_usage": {
+      "input_tokens": 18539,
+      "output_tokens": 374,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 2.0,\n \"failure_window_start_seconds\": 2.0,\n \"failure_window_end_seconds\": 3.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the basket handle to flip it upright, but the gripper misaligned and failed to secure a grip, slipping off the basket.\",\n \"failure_reason_summary\": \"improper grasp alignment; failure to secure handle grip\",\n \"recovery_action\": \"Reposition the end effector directly over the basket handle, close the gripper securely, and rotate the wrist to flip the basket upright.\",\n \"recovery_action_summary\": \"reposition end effector; secure handle grasp; rotate basket upright\",\n \"prevention_action\": \"Align the gripper accurately with the basket handle and ensure a firm grip is established before initiating the flipping rotation.\",\n \"prevention_action_summary\": \"align gripper with handle; establish firm grip; lift and flip\",\n \"justification\": \"At 2.0 seconds, the robot's gripper closes on the basket handle but fails to gain purchase, slipping off and pushing the basket down instead of lifting and flipping it upright.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 131,
+    "failure_sim_step": 131,
+    "failure_policy_step": 21,
+    "failure_chunk_index": 5,
+    "first_post_failure_policy_step": 22,
+    "failure_type": "grasp_failure",
+    "correction_action": "Reposition the end effector directly over the basket handle, close the gripper securely, and rotate the wrist to flip the basket upright."
+  }
+}
\ No newline at end of file
diff --git a/basket/ep026/labels.npz b/basket/ep026/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..fd6d25484d18846b48b6d508174f4e3c3f27cfd7
--- /dev/null
+++ b/basket/ep026/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f3d69a238043d57f9d5df616a42d6d4b6f2fc255caa4dd14480b08689c9eafc6
+size 3828
diff --git a/basket/ep027/example.md b/basket/ep027/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..e96c83e79e389c64f34cf8073477797977c6abc0
--- /dev/null
+++ b/basket/ep027/example.md
@@ -0,0 +1,33 @@
+# t002__ep027__basket
+
+- **suite / variant:** `` / ``
+- **instruction:** Flip the basket upright
+- **success:** False
+- **control steps:** 342  (6.8s at 50 Hz)
+- **policy inferences:** 60  (replan_steps=6)
+- **video:** `basket/ep027/rollout.mp4`
+- **features:** `60` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | approaching and grasping basket handle |
+| 3.0–6.0 | 151–301 | 25–50 | attempting flip and releasing basket |
+| 6.0–6.8 | 302–341 | 50–56 | retracting arm above table |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 2.862823061630219 s (00:02.9)
+- **frame / control_step:** 144 (policy_step=24, chunk_index=0)
+- **coarse onset:** 3.2 s (refined=True)
+- **mode:** placement_or_insertion_failure
+- **reason:** The robot gripped the handle of the red basket but failed to execute a full rotation to flip it upright, releasing the basket while it was still resting upside down on the table.
+  Summary: failed to rotate basket; incorrect orientation
+- **recovery:** Re-grasp the handle or edge of the red basket, lift and rotate it 180 degrees so the opening faces upward, and place it stably on the table.
+  Summary: re-grasp basket handle; rotate basket 180 degrees upright
+- **prevention:** Securely grasp the basket handle and follow a trajectory that rotates the object 180 degrees into an upright orientation before releasing.
+  Summary: execute 180-degree flip arc; maintain firm handle grip
+- **justification:** The robot attempts to flip the basket starting around 2.7 seconds and grasps the handle around 3.2 seconds. However, instead of executing a rotational motion to invert the basket upright, it sets it back down upside down by 4.2 seconds, resulting in a placement/orientation failure.
diff --git a/basket/ep027/labels.json b/basket/ep027/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..7c4cdbc4d6b788fb56251cbf92bee088367398ec
--- /dev/null
+++ b/basket/ep027/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t002__ep027__basket",
+  "scene_variant": "",
+  "task_id": 2,
+  "task_file": "",
+  "instruction": "Flip the basket upright",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T19:43:29+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.001988071570577,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "approaching and grasping basket handle",
+      "description": "approaching and grasping basket handle"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.001988071570577,
+      "t_end_sec": 6.003976143141154,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "attempting flip and releasing basket",
+      "description": "attempting flip and releasing basket"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.003976143141154,
+      "t_end_sec": 6.79920477137177,
+      "control_step_start": 302,
+      "control_step_end": 341,
+      "policy_step_start": 50,
+      "policy_step_end": 56,
+      "phrase": "retracting arm above table",
+      "description": "retracting arm above table"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "placement_or_insertion_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 2.862823061630219,
+    "vlm_failure_onset_timestamp": "00:02.9",
+    "vlm_failure_onset_frame": 144,
+    "vlm_failure_onset_step": 144,
+    "vlm_coarse_failure_onset_seconds": 3.2,
+    "vlm_failure_window_start_seconds": 2.7,
+    "vlm_failure_window_end_seconds": 4.2,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot gripped the handle of the red basket but failed to execute a full rotation to flip it upright, releasing the basket while it was still resting upside down on the table.",
+    "vlm_failure_reason_summary": "failed to rotate basket; incorrect orientation",
+    "vlm_recovery_action": "Re-grasp the handle or edge of the red basket, lift and rotate it 180 degrees so the opening faces upward, and place it stably on the table.",
+    "vlm_recovery_action_summary": "re-grasp basket handle; rotate basket 180 degrees upright",
+    "vlm_prevention_action": "Securely grasp the basket handle and follow a trajectory that rotates the object 180 degrees into an upright orientation before releasing.",
+    "vlm_prevention_action_summary": "execute 180-degree flip arc; maintain firm handle grip",
+    "vlm_justification": "The robot attempts to flip the basket starting around 2.7 seconds and grasps the handle around 3.2 seconds. However, instead of executing a rotational motion to invert the basket upright, it sets it back down upside down by 4.2 seconds, resulting in a placement/orientation failure.",
+    "vlm_temporal_justification": "At 02:14 (134 seconds), the robot's gripper slips off the basket handle mid-flip rather than completing the rotation, causing the basket to fall back flat upside down on the table.",
+    "vlm_video_fps": 50.3,
+    "vlm_video_num_frames": 342,
+    "vlm_video_duration_seconds": 6.779324055666004,
+    "vlm_usage": {
+      "input_tokens": 21386,
+      "output_tokens": 415,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"placement_or_insertion_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 3.2,\n \"failure_window_start_seconds\": 2.7,\n \"failure_window_end_seconds\": 4.2,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot gripped the handle of the red basket but failed to execute a full rotation to flip it upright, releasing the basket while it was still resting upside down on the table.\",\n \"failure_reason_summary\": \"failed to rotate basket; incorrect orientation\",\n \"recovery_action\": \"Re-grasp the handle or edge of the red basket, lift and rotate it 180 degrees so the opening faces upward, and place it stably on the table.\",\n \"recovery_action_summary\": \"re-grasp basket handle; rotate basket 180 degrees upright\",\n \"prevention_action\": \"Securely grasp the basket handle and follow a trajectory that rotates the object 180 degrees into an upright orientation before releasing.\",\n \"prevention_action_summary\": \"execute 180-degree flip arc; maintain firm handle grip\",\n \"justification\": \"The robot attempts to flip the basket starting around 2.7 seconds and grasps the handle around 3.2 seconds. However, instead of executing a rotational motion to invert the basket upright, it sets it back down upside down by 4.2 seconds, resulting in a placement/orientation failure.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 144,
+    "failure_sim_step": 144,
+    "failure_policy_step": 24,
+    "failure_chunk_index": 0,
+    "first_post_failure_policy_step": 25,
+    "failure_type": "placement_or_insertion_failure",
+    "correction_action": "Re-grasp the handle or edge of the red basket, lift and rotate it 180 degrees so the opening faces upward, and place it stably on the table."
+  }
+}
\ No newline at end of file
diff --git a/basket/ep027/labels.npz b/basket/ep027/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..826a3685a3a14cc72066468bc9441ed63581be6c
--- /dev/null
+++ b/basket/ep027/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8ed551b2f229b65e3e2e2bd7505d302e7761911875900543e531d706af24b06f
+size 3917
diff --git a/basket/ep028/example.md b/basket/ep028/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..64c333e41a824aa5b224a5be006fab8f911991fd
--- /dev/null
+++ b/basket/ep028/example.md
@@ -0,0 +1,34 @@
+# t002__ep028__basket
+
+- **suite / variant:** `` / ``
+- **instruction:** Flip the basket upright
+- **success:** False
+- **control steps:** 560  (11.1s at 50 Hz)
+- **policy inferences:** 97  (replan_steps=6)
+- **video:** `basket/ep028/rollout.mp4`
+- **features:** `97` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | reaching and attempting handle grasp |
+| 3.0–6.0 | 151–301 | 25–50 | slipping, retracting, and re-approaching |
+| 6.0–9.0 | 302–452 | 50–75 | grasping handle and lifting basket |
+| 9.0–11.1 | 453–559 | 75–93 | releasing handle and retracting arm |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 3.001988071570577 s (00:03.0)
+- **frame / control_step:** 151 (policy_step=25, chunk_index=1)
+- **coarse onset:** 3.0 s (refined=False)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the basket's handle to flip it upright, but failed to secure a firm grip, causing the handle to slip out during both lift attempts.
+  Summary: failed grasp on handle; handle slipping from gripper
+- **recovery:** Re-align the gripper directly over the black handle, close the fingers firmly around it, and lift smoothly to rotate the basket upright.
+  Summary: re-align gripper; firmly grasp handle; rotate basket upright
+- **prevention:** Position the gripper lower and more centrally around the black handle before closing to ensure a secure grip during the flipping motion.
+  Summary: position gripper lower; ensure secure handle grip
+- **justification:** At around 3.0 seconds, the robot's gripper attempts to grab the black handle on top of the red basket but slips off without achieving a stable hold. A second attempt at 7.0–8.0 seconds similarly slips, leaving the basket inverted at the end of the episode.
diff --git a/basket/ep028/labels.json b/basket/ep028/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..a0cfe0d8258f728514cb69244fee1888ec85ecdc
--- /dev/null
+++ b/basket/ep028/labels.json
@@ -0,0 +1,109 @@
+{
+  "rollout_id": "t002__ep028__basket",
+  "scene_variant": "",
+  "task_id": 2,
+  "task_file": "",
+  "instruction": "Flip the basket upright",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T19:44:06+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.001988071570577,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching and attempting handle grasp",
+      "description": "reaching and attempting handle grasp"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.001988071570577,
+      "t_end_sec": 6.003976143141154,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "slipping, retracting, and re-approaching",
+      "description": "slipping, retracting, and re-approaching"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.003976143141154,
+      "t_end_sec": 9.00596421471173,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "grasping handle and lifting basket",
+      "description": "grasping handle and lifting basket"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.00596421471173,
+      "t_end_sec": 11.133200795228628,
+      "control_step_start": 453,
+      "control_step_end": 559,
+      "policy_step_start": 75,
+      "policy_step_end": 93,
+      "phrase": "releasing handle and retracting arm",
+      "description": "releasing handle and retracting arm"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 3.001988071570577,
+    "vlm_failure_onset_timestamp": "00:03.0",
+    "vlm_failure_onset_frame": 151,
+    "vlm_failure_onset_step": 151,
+    "vlm_coarse_failure_onset_seconds": 3.0,
+    "vlm_failure_window_start_seconds": 3.0,
+    "vlm_failure_window_end_seconds": 9.0,
+    "vlm_temporal_refined": false,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the basket's handle to flip it upright, but failed to secure a firm grip, causing the handle to slip out during both lift attempts.",
+    "vlm_failure_reason_summary": "failed grasp on handle; handle slipping from gripper",
+    "vlm_recovery_action": "Re-align the gripper directly over the black handle, close the fingers firmly around it, and lift smoothly to rotate the basket upright.",
+    "vlm_recovery_action_summary": "re-align gripper; firmly grasp handle; rotate basket upright",
+    "vlm_prevention_action": "Position the gripper lower and more centrally around the black handle before closing to ensure a secure grip during the flipping motion.",
+    "vlm_prevention_action_summary": "position gripper lower; ensure secure handle grip",
+    "vlm_justification": "At around 3.0 seconds, the robot's gripper attempts to grab the black handle on top of the red basket but slips off without achieving a stable hold. A second attempt at 7.0\u20138.0 seconds similarly slips, leaving the basket inverted at the end of the episode.",
+    "vlm_temporal_justification": "The robot closes its gripper on the basket's handle and begins lifting upward, but at 02:08 the handle visibly begins slipping out of the gripper fingers due to an insecure grip.",
+    "vlm_video_fps": 50.3,
+    "vlm_video_num_frames": 560,
+    "vlm_video_duration_seconds": 11.113320079522863,
+    "vlm_usage": {
+      "input_tokens": 21632,
+      "output_tokens": 389,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 3.0,\n \"failure_window_start_seconds\": 3.0,\n \"failure_window_end_seconds\": 9.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the basket's handle to flip it upright, but failed to secure a firm grip, causing the handle to slip out during both lift attempts.\",\n \"failure_reason_summary\": \"failed grasp on handle; handle slipping from gripper\",\n \"recovery_action\": \"Re-align the gripper directly over the black handle, close the fingers firmly around it, and lift smoothly to rotate the basket upright.\",\n \"recovery_action_summary\": \"re-align gripper; firmly grasp handle; rotate basket upright\",\n \"prevention_action\": \"Position the gripper lower and more centrally around the black handle before closing to ensure a secure grip during the flipping motion.\",\n \"prevention_action_summary\": \"position gripper lower; ensure secure handle grip\",\n \"justification\": \"At around 3.0 seconds, the robot's gripper attempts to grab the black handle on top of the red basket but slips off without achieving a stable hold. A second attempt at 7.0\u20138.0 seconds similarly slips, leaving the basket inverted at the end of the episode.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 151,
+    "failure_sim_step": 151,
+    "failure_policy_step": 25,
+    "failure_chunk_index": 1,
+    "first_post_failure_policy_step": 26,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-align the gripper directly over the black handle, close the fingers firmly around it, and lift smoothly to rotate the basket upright."
+  }
+}
\ No newline at end of file
diff --git a/basket/ep028/labels.npz b/basket/ep028/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..c22afcf69114137bdd9d1bd0b5ca20444d1078ad
--- /dev/null
+++ b/basket/ep028/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0fbddfdf036cef5b111b6726c82a7fb86e191e6f94ac41396b9a4e2f6d641ba3
+size 3875
diff --git a/basket/ep029/example.md b/basket/ep029/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..2782d503f8d9eb0498ae3be1b7442cf5dbf6ef6f
--- /dev/null
+++ b/basket/ep029/example.md
@@ -0,0 +1,33 @@
+# t002__ep029__basket
+
+- **suite / variant:** `` / ``
+- **instruction:** Flip the basket upright
+- **success:** False
+- **control steps:** 371  (7.4s at 50 Hz)
+- **policy inferences:** 64  (replan_steps=6)
+- **video:** `basket/ep029/rollout.mp4`
+- **features:** `64` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | lifting handle and dropping basket |
+| 3.0–6.0 | 151–301 | 25–50 | pressing down on basket surface |
+| 6.0–7.4 | 302–370 | 50–61 | hovering above basket |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 2.4652087475149105 s (00:02.5)
+- **frame / control_step:** 124 (policy_step=20, chunk_index=4)
+- **coarse onset:** 2.2 s (refined=True)
+- **mode:** object_displacement
+- **reason:** The robot attempted to flip the basket by grabbing its handle, but lost its grip during the lift and dropped the basket back onto the table.
+  Summary: losing grip on basket handle; dropping object
+- **recovery:** Re-align the gripper with the basket handle or rim, secure a firm grasp, and carefully flip the basket upright onto its base.
+  Summary: re-grasp handle; rotate basket upright
+- **prevention:** Maintain a firm and stable grip on the basket handle while applying smooth rotational motion to prevent the basket from slipping out of the gripper.
+  Summary: secure firm grip on handle; rotate smoothly upright
+- **justification:** At around 2.2 seconds, the robot lifts the basket by its handle, but the handle slips from the gripper's fingers, causing the basket to drop back down un-flipped. Subsequent attempts to push/grasp the top surface fail to flip the basket.
diff --git a/basket/ep029/labels.json b/basket/ep029/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..eedac5578c650ce0f43e8a66e02a64c96ce2dd57
--- /dev/null
+++ b/basket/ep029/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t002__ep029__basket",
+  "scene_variant": "",
+  "task_id": 2,
+  "task_file": "",
+  "instruction": "Flip the basket upright",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T19:44:42+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.001988071570577,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "lifting handle and dropping basket",
+      "description": "lifting handle and dropping basket"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.001988071570577,
+      "t_end_sec": 6.003976143141154,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "pressing down on basket surface",
+      "description": "pressing down on basket surface"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.003976143141154,
+      "t_end_sec": 7.3757455268389664,
+      "control_step_start": 302,
+      "control_step_end": 370,
+      "policy_step_start": 50,
+      "policy_step_end": 61,
+      "phrase": "hovering above basket",
+      "description": "hovering above basket"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "object_displacement",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 2.4652087475149105,
+    "vlm_failure_onset_timestamp": "00:02.5",
+    "vlm_failure_onset_frame": 124,
+    "vlm_failure_onset_step": 124,
+    "vlm_coarse_failure_onset_seconds": 2.2,
+    "vlm_failure_window_start_seconds": 1.7,
+    "vlm_failure_window_end_seconds": 2.7,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to flip the basket by grabbing its handle, but lost its grip during the lift and dropped the basket back onto the table.",
+    "vlm_failure_reason_summary": "losing grip on basket handle; dropping object",
+    "vlm_recovery_action": "Re-align the gripper with the basket handle or rim, secure a firm grasp, and carefully flip the basket upright onto its base.",
+    "vlm_recovery_action_summary": "re-grasp handle; rotate basket upright",
+    "vlm_prevention_action": "Maintain a firm and stable grip on the basket handle while applying smooth rotational motion to prevent the basket from slipping out of the gripper.",
+    "vlm_prevention_action_summary": "secure firm grip on handle; rotate smoothly upright",
+    "vlm_justification": "At around 2.2 seconds, the robot lifts the basket by its handle, but the handle slips from the gripper's fingers, causing the basket to drop back down un-flipped. Subsequent attempts to push/grasp the top surface fail to flip the basket.",
+    "vlm_temporal_justification": "At second 124 (02:04), the robot loses its grip on the basket's handle as it attempts to lift and flip it, causing the basket to slip free from the gripper and fall back onto the table.",
+    "vlm_video_fps": 50.3,
+    "vlm_video_num_frames": 371,
+    "vlm_video_duration_seconds": 7.355864811133201,
+    "vlm_usage": {
+      "input_tokens": 18852,
+      "output_tokens": 383,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"object_displacement\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 2.2,\n \"failure_window_start_seconds\": 1.7,\n \"failure_window_end_seconds\": 2.7,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to flip the basket by grabbing its handle, but lost its grip during the lift and dropped the basket back onto the table.\",\n \"failure_reason_summary\": \"losing grip on basket handle; dropping object\",\n \"recovery_action\": \"Re-align the gripper with the basket handle or rim, secure a firm grasp, and carefully flip the basket upright onto its base.\",\n \"recovery_action_summary\": \"re-grasp handle; rotate basket upright\",\n \"prevention_action\": \"Maintain a firm and stable grip on the basket handle while applying smooth rotational motion to prevent the basket from slipping out of the gripper.\",\n \"prevention_action_summary\": \"secure firm grip on handle; rotate smoothly upright\",\n \"justification\": \"At around 2.2 seconds, the robot lifts the basket by its handle, but the handle slips from the gripper's fingers, causing the basket to drop back down un-flipped. Subsequent attempts to push/grasp the top surface fail to flip the basket.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 124,
+    "failure_sim_step": 124,
+    "failure_policy_step": 20,
+    "failure_chunk_index": 4,
+    "first_post_failure_policy_step": 21,
+    "failure_type": "object_displacement",
+    "correction_action": "Re-align the gripper with the basket handle or rim, secure a firm grasp, and carefully flip the basket upright onto its base."
+  }
+}
\ No newline at end of file
diff --git a/basket/ep029/labels.npz b/basket/ep029/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..a337d5cb915539d5a2bdc37cc3a3257bd20f05d3
--- /dev/null
+++ b/basket/ep029/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:05e1504329322aabf7ece1749979ec2d4b402a5f4e9745a22f12120a9e8b858a
+size 3813
diff --git a/basket/ep030/example.md b/basket/ep030/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..a9b62e1af18ca838b8407078bf219a644e6aa9ce
--- /dev/null
+++ b/basket/ep030/example.md
@@ -0,0 +1,33 @@
+# t002__ep030__basket
+
+- **suite / variant:** `` / ``
+- **instruction:** Flip the basket upright
+- **success:** False
+- **control steps:** 351  (6.9s at 51 Hz)
+- **policy inferences:** 60  (replan_steps=6)
+- **video:** `basket/ep030/rollout.mp4`
+- **features:** `60` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–152 | 0–25 | approaching and attempting handle grasp |
+| 3.0–6.0 | 153–305 | 25–50 | retracting arm away from basket |
+| 6.0–6.9 | 306–350 | 51–58 | holding position above table |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 2.5294117647058822 s (00:02.5)
+- **frame / control_step:** 129 (policy_step=21, chunk_index=3)
+- **coarse onset:** 2.2 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the handle of the basket to flip it, but the gripper closed prematurely and failed to secure a hold on the handle before retracting.
+  Summary: failed grasp on basket handle; gripper missed handle
+- **recovery:** Re-align the gripper over the basket handle, descend to firmly clamp the handle, and rotate the end-effector to flip the basket upright.
+  Summary: re-align gripper; grasp handle firmly; flip basket upright
+- **prevention:** Lower the end-effector closer to the handle and ensure proper alignment before closing the fingers to achieve a secure grip before pulling or flipping.
+  Summary: lower gripper further; align precisely with handle before grasping
+- **justification:** At around 2.2 seconds, the gripper closes above/around the basket handle without securing it. The arm then pulls away empty at 3.0 seconds, leaving the basket upside down for the remainder of the episode.
diff --git a/basket/ep030/labels.json b/basket/ep030/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..8b9af856439d6e42d076434f2a26495d80d7205f
--- /dev/null
+++ b/basket/ep030/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t002__ep030__basket",
+  "scene_variant": "",
+  "task_id": 2,
+  "task_file": "",
+  "instruction": "Flip the basket upright",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T19:45:15+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0,
+      "control_step_start": 0,
+      "control_step_end": 152,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "approaching and attempting handle grasp",
+      "description": "approaching and attempting handle grasp"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0,
+      "t_end_sec": 6.0,
+      "control_step_start": 153,
+      "control_step_end": 305,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "retracting arm away from basket",
+      "description": "retracting arm away from basket"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.0,
+      "t_end_sec": 6.882352941176471,
+      "control_step_start": 306,
+      "control_step_end": 350,
+      "policy_step_start": 51,
+      "policy_step_end": 58,
+      "phrase": "holding position above table",
+      "description": "holding position above table"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 2.5294117647058822,
+    "vlm_failure_onset_timestamp": "00:02.5",
+    "vlm_failure_onset_frame": 129,
+    "vlm_failure_onset_step": 129,
+    "vlm_coarse_failure_onset_seconds": 2.2,
+    "vlm_failure_window_start_seconds": 1.7,
+    "vlm_failure_window_end_seconds": 2.7,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the handle of the basket to flip it, but the gripper closed prematurely and failed to secure a hold on the handle before retracting.",
+    "vlm_failure_reason_summary": "failed grasp on basket handle; gripper missed handle",
+    "vlm_recovery_action": "Re-align the gripper over the basket handle, descend to firmly clamp the handle, and rotate the end-effector to flip the basket upright.",
+    "vlm_recovery_action_summary": "re-align gripper; grasp handle firmly; flip basket upright",
+    "vlm_prevention_action": "Lower the end-effector closer to the handle and ensure proper alignment before closing the fingers to achieve a secure grip before pulling or flipping.",
+    "vlm_prevention_action_summary": "lower gripper further; align precisely with handle before grasping",
+    "vlm_justification": "At around 2.2 seconds, the gripper closes above/around the basket handle without securing it. The arm then pulls away empty at 3.0 seconds, leaving the basket upside down for the remainder of the episode.",
+    "vlm_temporal_justification": "At clip second 02:09 (corresponding to frame 129 of the original rollout), the robot's gripper begins closing prematurely before properly encircling the handle. As a result, the fingers close completely without grabbing the handle, leading directly to the grasp failure.",
+    "vlm_video_fps": 51.0,
+    "vlm_video_num_frames": 351,
+    "vlm_video_duration_seconds": 6.862745098039215,
+    "vlm_usage": {
+      "input_tokens": 19045,
+      "output_tokens": 395,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 2.2,\n \"failure_window_start_seconds\": 1.7,\n \"failure_window_end_seconds\": 2.7,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the handle of the basket to flip it, but the gripper closed prematurely and failed to secure a hold on the handle before retracting.\",\n \"failure_reason_summary\": \"failed grasp on basket handle; gripper missed handle\",\n \"recovery_action\": \"Re-align the gripper over the basket handle, descend to firmly clamp the handle, and rotate the end-effector to flip the basket upright.\",\n \"recovery_action_summary\": \"re-align gripper; grasp handle firmly; flip basket upright\",\n \"prevention_action\": \"Lower the end-effector closer to the handle and ensure proper alignment before closing the fingers to achieve a secure grip before pulling or flipping.\",\n \"prevention_action_summary\": \"lower gripper further; align precisely with handle before grasping\",\n \"justification\": \"At around 2.2 seconds, the gripper closes above/around the basket handle without securing it. The arm then pulls away empty at 3.0 seconds, leaving the basket upside down for the remainder of the episode.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 129,
+    "failure_sim_step": 129,
+    "failure_policy_step": 21,
+    "failure_chunk_index": 3,
+    "first_post_failure_policy_step": 22,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-align the gripper over the basket handle, descend to firmly clamp the handle, and rotate the end-effector to flip the basket upright."
+  }
+}
\ No newline at end of file
diff --git a/basket/ep030/labels.npz b/basket/ep030/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..4e7c73049645bd40856f4027daf8b4c72b1d18bd
--- /dev/null
+++ b/basket/ep030/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4755add10d451b9aa29ca89488cf8d4d8aa8e06ba99c5c44810623e0d047fea0
+size 3872
diff --git a/basket/ep031/example.md b/basket/ep031/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..c9819b02febf429534e3ec6a09ca990ecdc2023a
--- /dev/null
+++ b/basket/ep031/example.md
@@ -0,0 +1,35 @@
+# t002__ep031__basket
+
+- **suite / variant:** `` / ``
+- **instruction:** Flip the basket upright
+- **success:** False
+- **control steps:** 639  (12.7s at 50 Hz)
+- **policy inferences:** 109  (replan_steps=6)
+- **video:** `basket/ep031/rollout.mp4`
+- **features:** `109` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | approaching and grasping basket edge |
+| 3.0–6.0 | 151–301 | 25–50 | attempting to flip basket |
+| 6.0–9.0 | 302–452 | 50–75 | pressing against top of basket |
+| 9.0–12.0 | 453–603 | 75–100 | holding position over basket |
+| 12.0–12.7 | 604–638 | 100–106 | remaining stationary near basket |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 1.7658730158730158 s (00:01.8)
+- **frame / control_step:** 89 (policy_step=14, chunk_index=5)
+- **coarse onset:** 3.0 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the rim of the red basket to flip it upright, but failed to achieve a firm grip, leaving the basket upside down on the table.
+  Summary: failed grasp on basket rim; failure to flip object
+- **recovery:** Re-align the end-effector with the rim or handle of the basket, establish a firm grasp, and lift and rotate it until it is upright.
+  Summary: re-align gripper; secure basket rim; flip upright
+- **prevention:** Position the gripper accurately over the edge or handle of the basket before closing the fingers to guarantee a firm grasp for flipping.
+  Summary: align with basket rim; secure grip before flipping
+- **justification:** At around 3.0 seconds, the robot descends and attempts to grasp the edge of the red basket, but the gripper fails to establish a secure grip. As a result, the robot cannot rotate or flip the basket, leaving it stuck in place for the remainder of the episode.
diff --git a/basket/ep031/labels.json b/basket/ep031/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..b534f591b4e5de9b19ecd6fad78b457fe58d4e0f
--- /dev/null
+++ b/basket/ep031/labels.json
@@ -0,0 +1,120 @@
+{
+  "rollout_id": "t002__ep031__basket",
+  "scene_variant": "",
+  "task_id": 2,
+  "task_file": "",
+  "instruction": "Flip the basket upright",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T19:45:49+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.996031746031746,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "approaching and grasping basket edge",
+      "description": "approaching and grasping basket edge"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.996031746031746,
+      "t_end_sec": 5.992063492063492,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "attempting to flip basket",
+      "description": "attempting to flip basket"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.992063492063492,
+      "t_end_sec": 8.988095238095239,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "pressing against top of basket",
+      "description": "pressing against top of basket"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.988095238095239,
+      "t_end_sec": 11.984126984126984,
+      "control_step_start": 453,
+      "control_step_end": 603,
+      "policy_step_start": 75,
+      "policy_step_end": 100,
+      "phrase": "holding position over basket",
+      "description": "holding position over basket"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 11.984126984126984,
+      "t_end_sec": 12.678571428571429,
+      "control_step_start": 604,
+      "control_step_end": 638,
+      "policy_step_start": 100,
+      "policy_step_end": 106,
+      "phrase": "remaining stationary near basket",
+      "description": "remaining stationary near basket"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 1.7658730158730158,
+    "vlm_failure_onset_timestamp": "00:01.8",
+    "vlm_failure_onset_frame": 89,
+    "vlm_failure_onset_step": 89,
+    "vlm_coarse_failure_onset_seconds": 3.0,
+    "vlm_failure_window_start_seconds": 3.0,
+    "vlm_failure_window_end_seconds": 5.0,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the rim of the red basket to flip it upright, but failed to achieve a firm grip, leaving the basket upside down on the table.",
+    "vlm_failure_reason_summary": "failed grasp on basket rim; failure to flip object",
+    "vlm_recovery_action": "Re-align the end-effector with the rim or handle of the basket, establish a firm grasp, and lift and rotate it until it is upright.",
+    "vlm_recovery_action_summary": "re-align gripper; secure basket rim; flip upright",
+    "vlm_prevention_action": "Position the gripper accurately over the edge or handle of the basket before closing the fingers to guarantee a firm grasp for flipping.",
+    "vlm_prevention_action_summary": "align with basket rim; secure grip before flipping",
+    "vlm_justification": "At around 3.0 seconds, the robot descends and attempts to grasp the edge of the red basket, but the gripper fails to establish a secure grip. As a result, the robot cannot rotate or flip the basket, leaving it stuck in place for the remainder of the episode.",
+    "vlm_temporal_justification": "At 01:29 (second 89), the robot's gripper closes its fingers completely without making solid contact or securing a grip on the handle/rim of the red basket, failing the grasp attempt.",
+    "vlm_video_fps": 50.4,
+    "vlm_video_num_frames": 639,
+    "vlm_video_duration_seconds": 12.65873015873016,
+    "vlm_usage": {
+      "input_tokens": 21757,
+      "output_tokens": 392,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 3.0,\n \"failure_window_start_seconds\": 3.0,\n \"failure_window_end_seconds\": 5.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the rim of the red basket to flip it upright, but failed to achieve a firm grip, leaving the basket upside down on the table.\",\n \"failure_reason_summary\": \"failed grasp on basket rim; failure to flip object\",\n \"recovery_action\": \"Re-align the end-effector with the rim or handle of the basket, establish a firm grasp, and lift and rotate it until it is upright.\",\n \"recovery_action_summary\": \"re-align gripper; secure basket rim; flip upright\",\n \"prevention_action\": \"Position the gripper accurately over the edge or handle of the basket before closing the fingers to guarantee a firm grasp for flipping.\",\n \"prevention_action_summary\": \"align with basket rim; secure grip before flipping\",\n \"justification\": \"At around 3.0 seconds, the robot descends and attempts to grasp the edge of the red basket, but the gripper fails to establish a secure grip. As a result, the robot cannot rotate or flip the basket, leaving it stuck in place for the remainder of the episode.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 89,
+    "failure_sim_step": 89,
+    "failure_policy_step": 14,
+    "failure_chunk_index": 5,
+    "first_post_failure_policy_step": 15,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-align the end-effector with the rim or handle of the basket, establish a firm grasp, and lift and rotate it until it is upright."
+  }
+}
\ No newline at end of file
diff --git a/basket/ep031/labels.npz b/basket/ep031/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..d13379ba9079bd1cda62d72380a1134470d1adad
--- /dev/null
+++ b/basket/ep031/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:74d3619d7d9f87b7b7d85e5ff75a01f3d2f22528560196f2040e53325c7c4895
+size 3929
diff --git a/basket/ep032/example.md b/basket/ep032/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..3352890fefccc4e3c6f9723b3ee085ca97323e3e
--- /dev/null
+++ b/basket/ep032/example.md
@@ -0,0 +1,33 @@
+# t002__ep032__basket
+
+- **suite / variant:** `` / ``
+- **instruction:** Flip the basket upright
+- **success:** False
+- **control steps:** 427  (8.4s at 51 Hz)
+- **policy inferences:** 74  (replan_steps=6)
+- **video:** `basket/ep032/rollout.mp4`
+- **features:** `74` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching for basket handle |
+| 3.0–6.0 | 152–303 | 25–50 | slipping off handle and pulling up |
+| 6.0–8.4 | 304–426 | 50–71 | hovering above basket |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 3.0039525691699605 s (00:03.0)
+- **frame / control_step:** 152 (policy_step=25, chunk_index=2)
+- **coarse onset:** 2.5 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the handle of the red basket, but failed to close its gripper securely around the handle, causing the gripper to slip off without flipping the basket upright.
+  Summary: failed grasp on basket handle; gripper slipped
+- **recovery:** Reposition the end-effector over the basket handle, align the fingers, and securely close the gripper before attempting to lift and flip the basket upright.
+  Summary: re-align gripper; grasp handle securely; flip basket
+- **prevention:** Approach the basket handle with proper gripper alignment and depth to ensure a firm hold before initiating the flipping motion.
+  Summary: align gripper with handle; secure grasp before flip
+- **justification:** At approximately 2.5 seconds, the end-effector closes near the handle of the red basket but fails to establish a secure hold, slipping off and leaving the basket unflipped for the remainder of the episode.
diff --git a/basket/ep032/labels.json b/basket/ep032/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..6b11312edcc2d40ec0545391daffa2360434d7f7
--- /dev/null
+++ b/basket/ep032/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t002__ep032__basket",
+  "scene_variant": "",
+  "task_id": 2,
+  "task_file": "",
+  "instruction": "Flip the basket upright",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T19:46:23+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0039525691699605,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for basket handle",
+      "description": "reaching for basket handle"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0039525691699605,
+      "t_end_sec": 6.007905138339921,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "slipping off handle and pulling up",
+      "description": "slipping off handle and pulling up"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.007905138339921,
+      "t_end_sec": 8.438735177865613,
+      "control_step_start": 304,
+      "control_step_end": 426,
+      "policy_step_start": 50,
+      "policy_step_end": 71,
+      "phrase": "hovering above basket",
+      "description": "hovering above basket"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 3.0039525691699605,
+    "vlm_failure_onset_timestamp": "00:03.0",
+    "vlm_failure_onset_frame": 152,
+    "vlm_failure_onset_step": 152,
+    "vlm_coarse_failure_onset_seconds": 2.5,
+    "vlm_failure_window_start_seconds": 2.0,
+    "vlm_failure_window_end_seconds": 3.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the handle of the red basket, but failed to close its gripper securely around the handle, causing the gripper to slip off without flipping the basket upright.",
+    "vlm_failure_reason_summary": "failed grasp on basket handle; gripper slipped",
+    "vlm_recovery_action": "Reposition the end-effector over the basket handle, align the fingers, and securely close the gripper before attempting to lift and flip the basket upright.",
+    "vlm_recovery_action_summary": "re-align gripper; grasp handle securely; flip basket",
+    "vlm_prevention_action": "Approach the basket handle with proper gripper alignment and depth to ensure a firm hold before initiating the flipping motion.",
+    "vlm_prevention_action_summary": "align gripper with handle; secure grasp before flip",
+    "vlm_justification": "At approximately 2.5 seconds, the end-effector closes near the handle of the red basket but fails to establish a secure hold, slipping off and leaving the basket unflipped for the remainder of the episode.",
+    "vlm_temporal_justification": "At second 152 (02:32 in the video), the robot gripper closes its fingers to grasp the blue wire handle, but fails to position properly around it, causing the gripper to slip off without securing the handle as it lifts.",
+    "vlm_video_fps": 50.6,
+    "vlm_video_num_frames": 427,
+    "vlm_video_duration_seconds": 8.41897233201581,
+    "vlm_usage": {
+      "input_tokens": 19930,
+      "output_tokens": 381,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 2.5,\n \"failure_window_start_seconds\": 2.0,\n \"failure_window_end_seconds\": 3.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the handle of the red basket, but failed to close its gripper securely around the handle, causing the gripper to slip off without flipping the basket upright.\",\n \"failure_reason_summary\": \"failed grasp on basket handle; gripper slipped\",\n \"recovery_action\": \"Reposition the end-effector over the basket handle, align the fingers, and securely close the gripper before attempting to lift and flip the basket upright.\",\n \"recovery_action_summary\": \"re-align gripper; grasp handle securely; flip basket\",\n \"prevention_action\": \"Approach the basket handle with proper gripper alignment and depth to ensure a firm hold before initiating the flipping motion.\",\n \"prevention_action_summary\": \"align gripper with handle; secure grasp before flip\",\n \"justification\": \"At approximately 2.5 seconds, the end-effector closes near the handle of the red basket but fails to establish a secure hold, slipping off and leaving the basket unflipped for the remainder of the episode.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 152,
+    "failure_sim_step": 152,
+    "failure_policy_step": 25,
+    "failure_chunk_index": 2,
+    "first_post_failure_policy_step": 26,
+    "failure_type": "grasp_failure",
+    "correction_action": "Reposition the end-effector over the basket handle, align the fingers, and securely close the gripper before attempting to lift and flip the basket upright."
+  }
+}
\ No newline at end of file
diff --git a/basket/ep032/labels.npz b/basket/ep032/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..126c4c039cbcb58b240fec5d60329682aaa59383
--- /dev/null
+++ b/basket/ep032/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8f2e6ec64ef768ecc3a1cd62716de1ede6ca6356f8235b0ecbc259c62fdba5f7
+size 3845
diff --git a/basket/ep033/example.md b/basket/ep033/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..5ecd80975a7e8b8a094547e1a21ed2d56090fe60
--- /dev/null
+++ b/basket/ep033/example.md
@@ -0,0 +1,22 @@
+# t002__ep033__basket
+
+- **suite / variant:** `` / ``
+- **instruction:** Flip the basket upright
+- **success:** True
+- **control steps:** 580  (11.5s at 50 Hz)
+- **policy inferences:** 100  (replan_steps=6)
+- **video:** `basket/ep033/rollout.mp4`
+- **features:** `100` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | reaching for basket handle |
+| 3.0–6.0 | 151–301 | 25–50 | grasping handle and lifting basket |
+| 6.0–9.0 | 302–452 | 50–75 | flipping basket upright |
+| 9.0–11.5 | 453–579 | 75–96 | releasing basket and retracting |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/basket/ep033/labels.json b/basket/ep033/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..0d07aa6d32df1cb773777b6e6842aeea1372f514
--- /dev/null
+++ b/basket/ep033/labels.json
@@ -0,0 +1,79 @@
+{
+  "rollout_id": "t002__ep033__basket",
+  "scene_variant": "",
+  "task_id": 2,
+  "task_file": "",
+  "instruction": "Flip the basket upright",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T19:46:32+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.996031746031746,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for basket handle",
+      "description": "reaching for basket handle"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.996031746031746,
+      "t_end_sec": 5.992063492063492,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "grasping handle and lifting basket",
+      "description": "grasping handle and lifting basket"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.992063492063492,
+      "t_end_sec": 8.988095238095239,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "flipping basket upright",
+      "description": "flipping basket upright"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.988095238095239,
+      "t_end_sec": 11.507936507936508,
+      "control_step_start": 453,
+      "control_step_end": 579,
+      "policy_step_start": 75,
+      "policy_step_end": 96,
+      "phrase": "releasing basket and retracting",
+      "description": "releasing basket and retracting"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/basket/ep033/labels.npz b/basket/ep033/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..0e48302ce7d85a3ad90d6fe4440ac26af11b4cd8
--- /dev/null
+++ b/basket/ep033/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bb95f99606029a3224ba8069e33ba6f4d84009496d8ec1bde02e7eace6601058
+size 3125
diff --git a/basket/ep034/example.md b/basket/ep034/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..bd0eb09f9bda4ec1e454697e5f2c55dc4370c690
--- /dev/null
+++ b/basket/ep034/example.md
@@ -0,0 +1,33 @@
+# t002__ep034__basket
+
+- **suite / variant:** `` / ``
+- **instruction:** Flip the basket upright
+- **success:** False
+- **control steps:** 369  (7.3s at 51 Hz)
+- **policy inferences:** 64  (replan_steps=6)
+- **video:** `basket/ep034/rollout.mp4`
+- **features:** `64` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | approaching and grasping basket rim |
+| 3.0–6.0 | 152–303 | 25–50 | attempting flip and losing grip |
+| 6.0–7.3 | 304–368 | 50–61 | resting gripper on basket |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 4.0828402366863905 s (00:04.1)
+- **frame / control_step:** 207 (policy_step=34, chunk_index=3)
+- **coarse onset:** 3.3 s (refined=True)
+- **mode:** object_displacement
+- **reason:** The robot attempted to flip the red basket upright, but its gripper lost hold of the basket rim, causing the basket to drop back flat onto the table.
+  Summary: lost grasp on basket rim; dropped basket back down
+- **recovery:** Re-grip the edge of the red basket securely, lift and rotate it until it stands upright on its base.
+  Summary: re-grip basket edge; rotate basket upright
+- **prevention:** Maintain a tight grip on the basket rim while executing the flipping motion to prevent it from slipping.
+  Summary: maintain secure grip; rotate basket upright
+- **justification:** Around 3.3 seconds into the trajectory, the robot attempts to flip the red basket, but its gripper slips off the rim. The basket falls flat onto the table and the robot makes no further effective attempt to flip it upright.
diff --git a/basket/ep034/labels.json b/basket/ep034/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..5775c78f06011865b38d85378f8916804a40773a
--- /dev/null
+++ b/basket/ep034/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t002__ep034__basket",
+  "scene_variant": "",
+  "task_id": 2,
+  "task_file": "",
+  "instruction": "Flip the basket upright",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T19:47:08+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9980276134122286,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "approaching and grasping basket rim",
+      "description": "approaching and grasping basket rim"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9980276134122286,
+      "t_end_sec": 5.996055226824457,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "attempting flip and losing grip",
+      "description": "attempting flip and losing grip"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.996055226824457,
+      "t_end_sec": 7.278106508875739,
+      "control_step_start": 304,
+      "control_step_end": 368,
+      "policy_step_start": 50,
+      "policy_step_end": 61,
+      "phrase": "resting gripper on basket",
+      "description": "resting gripper on basket"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "object_displacement",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 4.0828402366863905,
+    "vlm_failure_onset_timestamp": "00:04.1",
+    "vlm_failure_onset_frame": 207,
+    "vlm_failure_onset_step": 207,
+    "vlm_coarse_failure_onset_seconds": 3.3,
+    "vlm_failure_window_start_seconds": 3.0,
+    "vlm_failure_window_end_seconds": 4.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to flip the red basket upright, but its gripper lost hold of the basket rim, causing the basket to drop back flat onto the table.",
+    "vlm_failure_reason_summary": "lost grasp on basket rim; dropped basket back down",
+    "vlm_recovery_action": "Re-grip the edge of the red basket securely, lift and rotate it until it stands upright on its base.",
+    "vlm_recovery_action_summary": "re-grip basket edge; rotate basket upright",
+    "vlm_prevention_action": "Maintain a tight grip on the basket rim while executing the flipping motion to prevent it from slipping.",
+    "vlm_prevention_action_summary": "maintain secure grip; rotate basket upright",
+    "vlm_justification": "Around 3.3 seconds into the trajectory, the robot attempts to flip the red basket, but its gripper slips off the rim. The basket falls flat onto the table and the robot makes no further effective attempt to flip it upright.",
+    "vlm_temporal_justification": "At 03:12 into the clip (second 192), the robot's gripper loses its secure hold on the basket handle as it lifts upward, causing the handle to slip out of the fingers and the basket to drop back flat onto the table.",
+    "vlm_video_fps": 50.7,
+    "vlm_video_num_frames": 369,
+    "vlm_video_duration_seconds": 7.258382642998027,
+    "vlm_usage": {
+      "input_tokens": 21504,
+      "output_tokens": 374,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"object_displacement\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 3.3,\n \"failure_window_start_seconds\": 3.0,\n \"failure_window_end_seconds\": 4.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to flip the red basket upright, but its gripper lost hold of the basket rim, causing the basket to drop back flat onto the table.\",\n \"failure_reason_summary\": \"lost grasp on basket rim; dropped basket back down\",\n \"recovery_action\": \"Re-grip the edge of the red basket securely, lift and rotate it until it stands upright on its base.\",\n \"recovery_action_summary\": \"re-grip basket edge; rotate basket upright\",\n \"prevention_action\": \"Maintain a tight grip on the basket rim while executing the flipping motion to prevent it from slipping.\",\n \"prevention_action_summary\": \"maintain secure grip; rotate basket upright\",\n \"justification\": \"Around 3.3 seconds into the trajectory, the robot attempts to flip the red basket, but its gripper slips off the rim. The basket falls flat onto the table and the robot makes no further effective attempt to flip it upright.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 207,
+    "failure_sim_step": 207,
+    "failure_policy_step": 34,
+    "failure_chunk_index": 3,
+    "first_post_failure_policy_step": 35,
+    "failure_type": "object_displacement",
+    "correction_action": "Re-grip the edge of the red basket securely, lift and rotate it until it stands upright on its base."
+  }
+}
\ No newline at end of file
diff --git a/basket/ep034/labels.npz b/basket/ep034/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..2ee72f7c9aa75db37b08dce16a492395da1e908b
--- /dev/null
+++ b/basket/ep034/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:263722e81f4702db2719f5b432bb2029b8a8d6c54d1e25f09cf74c2242912faa
+size 3747
diff --git a/basket/ep035/example.md b/basket/ep035/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..e93b5c5d83c008a215f8d2bcce459b335d538e9b
--- /dev/null
+++ b/basket/ep035/example.md
@@ -0,0 +1,21 @@
+# t002__ep035__basket
+
+- **suite / variant:** `` / ``
+- **instruction:** Flip the basket upright
+- **success:** True
+- **control steps:** 321  (6.4s at 50 Hz)
+- **policy inferences:** 55  (replan_steps=6)
+- **video:** `basket/ep035/rollout.mp4`
+- **features:** `55` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | grasping and flipping basket |
+| 3.0–6.0 | 152–303 | 25–50 | releasing basket and retracting |
+| 6.0–6.4 | 304–320 | 50–53 | idle above the table |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/basket/ep035/labels.json b/basket/ep035/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..f1296e4f1e3cb09a439c7b0fdeba4c3afd087ae6
--- /dev/null
+++ b/basket/ep035/labels.json
@@ -0,0 +1,68 @@
+{
+  "rollout_id": "t002__ep035__basket",
+  "scene_variant": "",
+  "task_id": 2,
+  "task_file": "",
+  "instruction": "Flip the basket upright",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T19:47:15+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00990099009901,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "grasping and flipping basket",
+      "description": "grasping and flipping basket"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00990099009901,
+      "t_end_sec": 6.01980198019802,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "releasing basket and retracting",
+      "description": "releasing basket and retracting"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01980198019802,
+      "t_end_sec": 6.356435643564357,
+      "control_step_start": 304,
+      "control_step_end": 320,
+      "policy_step_start": 50,
+      "policy_step_end": 53,
+      "phrase": "idle above the table",
+      "description": "idle above the table"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/basket/ep035/labels.npz b/basket/ep035/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..3654e8de02a1700e2faeccd88d31d1b9ef526974
--- /dev/null
+++ b/basket/ep035/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b6980163e759d456f047a625d01a9f6571a36293d9dd4765fcf28ef76a67fe5c
+size 3080
diff --git a/basket/ep036/example.md b/basket/ep036/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..628f5fc8bfc614a2263322c9a3ed202f1fb1b3bf
--- /dev/null
+++ b/basket/ep036/example.md
@@ -0,0 +1,32 @@
+# t002__ep036__basket
+
+- **suite / variant:** `` / ``
+- **instruction:** Flip the basket upright
+- **success:** False
+- **control steps:** 304  (6.0s at 50 Hz)
+- **policy inferences:** 53  (replan_steps=6)
+- **video:** `basket/ep036/rollout.mp4`
+- **features:** `53` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | approaching and grasping basket rim |
+| 3.0–6.0 | 152–303 | 25–50 | pressing down on basket top |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 2.99009900990099 s (00:03.0)
+- **frame / control_step:** 151 (policy_step=25, chunk_index=1)
+- **coarse onset:** 2.7 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the rim of the red basket to flip it upright, but failed to establish a secure hold, causing the gripper to slip off.
+  Summary: failed grasp on basket rim; gripper slipped off
+- **recovery:** Re-align the gripper with the rim of the basket, establish a secure grasp on the edge, and lift while rotating to flip it upright.
+  Summary: re-align gripper; grasp rim securely; flip upright
+- **prevention:** Position the gripper deeper over the basket rim before closing the fingers to ensure a stable grip before applying lifting force.
+  Summary: deepen grasp on rim; securely flip basket
+- **justification:** At around 2.7 seconds (frame 00:03), the robot's gripper attempts to grab the edge of the basket to flip it. However, the grasp is insecure and slips off, leaving the basket lying on its side.
diff --git a/basket/ep036/labels.json b/basket/ep036/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..2221018785e199b511cf2fc20a10748a0ed1a8b1
--- /dev/null
+++ b/basket/ep036/labels.json
@@ -0,0 +1,87 @@
+{
+  "rollout_id": "t002__ep036__basket",
+  "scene_variant": "",
+  "task_id": 2,
+  "task_file": "",
+  "instruction": "Flip the basket upright",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T19:47:52+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00990099009901,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "approaching and grasping basket rim",
+      "description": "approaching and grasping basket rim"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00990099009901,
+      "t_end_sec": 6.01980198019802,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "pressing down on basket top",
+      "description": "pressing down on basket top"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 2.99009900990099,
+    "vlm_failure_onset_timestamp": "00:03.0",
+    "vlm_failure_onset_frame": 151,
+    "vlm_failure_onset_step": 151,
+    "vlm_coarse_failure_onset_seconds": 2.7,
+    "vlm_failure_window_start_seconds": 2.2,
+    "vlm_failure_window_end_seconds": 3.2,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the rim of the red basket to flip it upright, but failed to establish a secure hold, causing the gripper to slip off.",
+    "vlm_failure_reason_summary": "failed grasp on basket rim; gripper slipped off",
+    "vlm_recovery_action": "Re-align the gripper with the rim of the basket, establish a secure grasp on the edge, and lift while rotating to flip it upright.",
+    "vlm_recovery_action_summary": "re-align gripper; grasp rim securely; flip upright",
+    "vlm_prevention_action": "Position the gripper deeper over the basket rim before closing the fingers to ensure a stable grip before applying lifting force.",
+    "vlm_prevention_action_summary": "deepen grasp on rim; securely flip basket",
+    "vlm_justification": "At around 2.7 seconds (frame 00:03), the robot's gripper attempts to grab the edge of the basket to flip it. However, the grasp is insecure and slips off, leaving the basket lying on its side.",
+    "vlm_temporal_justification": "At 02:31 (second 151 of the clip), as the robot arm rotates to flip the basket upright, the gripper loses a firm hold on the rim and visibly begins to slip off, causing the attempt to fail.",
+    "vlm_video_fps": 50.5,
+    "vlm_video_num_frames": 304,
+    "vlm_video_duration_seconds": 6.0,
+    "vlm_usage": {
+      "input_tokens": 20430,
+      "output_tokens": 385,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 2.7,\n \"failure_window_start_seconds\": 2.2,\n \"failure_window_end_seconds\": 3.2,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the rim of the red basket to flip it upright, but failed to establish a secure hold, causing the gripper to slip off.\",\n \"failure_reason_summary\": \"failed grasp on basket rim; gripper slipped off\",\n \"recovery_action\": \"Re-align the gripper with the rim of the basket, establish a secure grasp on the edge, and lift while rotating to flip it upright.\",\n \"recovery_action_summary\": \"re-align gripper; grasp rim securely; flip upright\",\n \"prevention_action\": \"Position the gripper deeper over the basket rim before closing the fingers to ensure a stable grip before applying lifting force.\",\n \"prevention_action_summary\": \"deepen grasp on rim; securely flip basket\",\n \"justification\": \"At around 2.7 seconds (frame 00:03), the robot's gripper attempts to grab the edge of the basket to flip it. However, the grasp is insecure and slips off, leaving the basket lying on its side.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 151,
+    "failure_sim_step": 151,
+    "failure_policy_step": 25,
+    "failure_chunk_index": 1,
+    "first_post_failure_policy_step": 26,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-align the gripper with the rim of the basket, establish a secure grasp on the edge, and lift while rotating to flip it upright."
+  }
+}
\ No newline at end of file
diff --git a/basket/ep036/labels.npz b/basket/ep036/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..f367970566dba7446e564bfe757c4aa49bdb2925
--- /dev/null
+++ b/basket/ep036/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c81b76981c45c243be70ef2d4737fae62c8a8452cffa57b303393500dfa16014
+size 3744
diff --git a/basket/ep037/example.md b/basket/ep037/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..791411294c9b38a7a527cbe6f61d661e7301169e
--- /dev/null
+++ b/basket/ep037/example.md
@@ -0,0 +1,21 @@
+# t002__ep037__basket
+
+- **suite / variant:** `` / ``
+- **instruction:** Flip the basket upright
+- **success:** True
+- **control steps:** 325  (6.4s at 50 Hz)
+- **policy inferences:** 56  (replan_steps=6)
+- **video:** `basket/ep037/rollout.mp4`
+- **features:** `56` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | grasping the basket handle |
+| 3.0–6.0 | 151–301 | 25–50 | flipping the basket upright |
+| 6.0–6.4 | 302–324 | 50–54 | releasing basket and retracting |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/basket/ep037/labels.json b/basket/ep037/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..6dfc258ee4ee5387065ea20456d3ada930663ce2
--- /dev/null
+++ b/basket/ep037/labels.json
@@ -0,0 +1,68 @@
+{
+  "rollout_id": "t002__ep037__basket",
+  "scene_variant": "",
+  "task_id": 2,
+  "task_file": "",
+  "instruction": "Flip the basket upright",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T19:48:00+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.996031746031746,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "grasping the basket handle",
+      "description": "grasping the basket handle"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.996031746031746,
+      "t_end_sec": 5.992063492063492,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "flipping the basket upright",
+      "description": "flipping the basket upright"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.992063492063492,
+      "t_end_sec": 6.448412698412699,
+      "control_step_start": 302,
+      "control_step_end": 324,
+      "policy_step_start": 50,
+      "policy_step_end": 54,
+      "phrase": "releasing basket and retracting",
+      "description": "releasing basket and retracting"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/basket/ep037/labels.npz b/basket/ep037/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..c26285ae9c5fd9e6aeb0f09cf0cd8689b361ea75
--- /dev/null
+++ b/basket/ep037/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2cb39837e438a31940921d5076447f56651628d5c1bd6a2883f027109fafa1e2
+size 3085
diff --git a/basket/ep038/example.md b/basket/ep038/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..84a527497fd32258bfbc2b57140b8905f90b3c3e
--- /dev/null
+++ b/basket/ep038/example.md
@@ -0,0 +1,32 @@
+# t002__ep038__basket
+
+- **suite / variant:** `` / ``
+- **instruction:** Flip the basket upright
+- **success:** False
+- **control steps:** 302  (6.0s at 50 Hz)
+- **policy inferences:** 52  (replan_steps=6)
+- **video:** `basket/ep038/rollout.mp4`
+- **features:** `52` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | reaching for and grasping basket |
+| 3.0–6.0 | 151–301 | 25–50 | dropping basket and retracting arm |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 3.8568588469184895 s (00:03.9)
+- **frame / control_step:** 194 (policy_step=32, chunk_index=2)
+- **coarse onset:** 3.5 s (refined=True)
+- **mode:** object_displacement
+- **reason:** The robot attempted to lift and flip the red basket by its cord, but lost its grip mid-motion, dropping the basket back onto the table upside down.
+  Summary: losing grip on basket; dropping basket unflipped
+- **recovery:** Re-grasp the basket handle or rim securely and execute a full rotation to rest the basket upright on the table surface.
+  Summary: re-grasp basket; rotate and place upright
+- **prevention:** Secure a firm grasp on the basket and maintain control through the full flipping rotation before releasing.
+  Summary: secure firm grasp; complete rotation before releasing
+- **justification:** The robot successfully grasps and lifts the basket at 00:03, but slips and drops it around 00:03.5. The basket lands back in its original upside-down orientation, and the arm retracts without completing the task.
diff --git a/basket/ep038/labels.json b/basket/ep038/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..70b10ac80175eb788aa36e8469b9cbfe6b86affd
--- /dev/null
+++ b/basket/ep038/labels.json
@@ -0,0 +1,87 @@
+{
+  "rollout_id": "t002__ep038__basket",
+  "scene_variant": "",
+  "task_id": 2,
+  "task_file": "",
+  "instruction": "Flip the basket upright",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T19:48:35+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.001988071570577,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for and grasping basket",
+      "description": "reaching for and grasping basket"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.001988071570577,
+      "t_end_sec": 6.003976143141154,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "dropping basket and retracting arm",
+      "description": "dropping basket and retracting arm"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "object_displacement",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 3.8568588469184895,
+    "vlm_failure_onset_timestamp": "00:03.9",
+    "vlm_failure_onset_frame": 194,
+    "vlm_failure_onset_step": 194,
+    "vlm_coarse_failure_onset_seconds": 3.5,
+    "vlm_failure_window_start_seconds": 3.0,
+    "vlm_failure_window_end_seconds": 4.0,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to lift and flip the red basket by its cord, but lost its grip mid-motion, dropping the basket back onto the table upside down.",
+    "vlm_failure_reason_summary": "losing grip on basket; dropping basket unflipped",
+    "vlm_recovery_action": "Re-grasp the basket handle or rim securely and execute a full rotation to rest the basket upright on the table surface.",
+    "vlm_recovery_action_summary": "re-grasp basket; rotate and place upright",
+    "vlm_prevention_action": "Secure a firm grasp on the basket and maintain control through the full flipping rotation before releasing.",
+    "vlm_prevention_action_summary": "secure firm grasp; complete rotation before releasing",
+    "vlm_justification": "The robot successfully grasps and lifts the basket at 00:03, but slips and drops it around 00:03.5. The basket lands back in its original upside-down orientation, and the arm retracts without completing the task.",
+    "vlm_temporal_justification": "At 02:49 (169 seconds into the clip), the robot gripper opens and loses its grip on the basket's cord, causing the basket to drop back onto the table without successfully flipping upright.",
+    "vlm_video_fps": 50.3,
+    "vlm_video_num_frames": 302,
+    "vlm_video_duration_seconds": 5.984095427435388,
+    "vlm_usage": {
+      "input_tokens": 19678,
+      "output_tokens": 373,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"object_displacement\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 3.5,\n \"failure_window_start_seconds\": 3.0,\n \"failure_window_end_seconds\": 4.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to lift and flip the red basket by its cord, but lost its grip mid-motion, dropping the basket back onto the table upside down.\",\n \"failure_reason_summary\": \"losing grip on basket; dropping basket unflipped\",\n \"recovery_action\": \"Re-grasp the basket handle or rim securely and execute a full rotation to rest the basket upright on the table surface.\",\n \"recovery_action_summary\": \"re-grasp basket; rotate and place upright\",\n \"prevention_action\": \"Secure a firm grasp on the basket and maintain control through the full flipping rotation before releasing.\",\n \"prevention_action_summary\": \"secure firm grasp; complete rotation before releasing\",\n \"justification\": \"The robot successfully grasps and lifts the basket at 00:03, but slips and drops it around 00:03.5. The basket lands back in its original upside-down orientation, and the arm retracts without completing the task.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 194,
+    "failure_sim_step": 194,
+    "failure_policy_step": 32,
+    "failure_chunk_index": 2,
+    "first_post_failure_policy_step": 33,
+    "failure_type": "object_displacement",
+    "correction_action": "Re-grasp the basket handle or rim securely and execute a full rotation to rest the basket upright on the table surface."
+  }
+}
\ No newline at end of file
diff --git a/basket/ep038/labels.npz b/basket/ep038/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..d36f399f84fab17d12be1a36a906029e45d21d7d
--- /dev/null
+++ b/basket/ep038/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7ef9adfb2d72c19a9dba7b9e80a472d2806a225040809a3ecf100149ccc865e5
+size 3738
diff --git a/basket/ep039/example.md b/basket/ep039/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..ceee5300cf44a37c7848c083b5fe87db9bdae9a9
--- /dev/null
+++ b/basket/ep039/example.md
@@ -0,0 +1,21 @@
+# t002__ep039__basket
+
+- **suite / variant:** `` / ``
+- **instruction:** Flip the basket upright
+- **success:** True
+- **control steps:** 317  (6.3s at 50 Hz)
+- **policy inferences:** 55  (replan_steps=6)
+- **video:** `basket/ep039/rollout.mp4`
+- **features:** `55` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | reaching and grasping basket |
+| 3.0–6.0 | 151–301 | 25–50 | flipping basket upright |
+| 6.0–6.3 | 302–316 | 50–52 | releasing basket and retracting |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/basket/ep039/labels.json b/basket/ep039/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..c81ddbb45e941363ee922a9206e7864d1691935b
--- /dev/null
+++ b/basket/ep039/labels.json
@@ -0,0 +1,68 @@
+{
+  "rollout_id": "t002__ep039__basket",
+  "scene_variant": "",
+  "task_id": 2,
+  "task_file": "",
+  "instruction": "Flip the basket upright",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T19:48:42+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.001988071570577,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching and grasping basket",
+      "description": "reaching and grasping basket"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.001988071570577,
+      "t_end_sec": 6.003976143141154,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "flipping basket upright",
+      "description": "flipping basket upright"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.003976143141154,
+      "t_end_sec": 6.3021868787276345,
+      "control_step_start": 302,
+      "control_step_end": 316,
+      "policy_step_start": 50,
+      "policy_step_end": 52,
+      "phrase": "releasing basket and retracting",
+      "description": "releasing basket and retracting"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/basket/ep039/labels.npz b/basket/ep039/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..abacfa7427bc533ae75af6569d509065becec805
--- /dev/null
+++ b/basket/ep039/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8e7b7e4d0f983397bcc0cc8d9b9003917c1431b494882112b25dd2cd4315dc78
+size 3077
diff --git a/basket/ep040/example.md b/basket/ep040/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..7994bf3a0b77cbeb593e95a29b3a9b9f47b4d105
--- /dev/null
+++ b/basket/ep040/example.md
@@ -0,0 +1,34 @@
+# t002__ep040__basket
+
+- **suite / variant:** `` / ``
+- **instruction:** Flip the basket upright
+- **success:** False
+- **control steps:** 469  (9.3s at 50 Hz)
+- **policy inferences:** 80  (replan_steps=6)
+- **video:** `basket/ep040/rollout.mp4`
+- **features:** `80` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching for basket handle |
+| 3.0–6.0 | 152–303 | 25–50 | slipping and re-grasping basket |
+| 6.0–9.0 | 304–455 | 50–75 | tilting basket partially |
+| 9.0–9.3 | 456–468 | 76–78 | holding basket stationary |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 3.128712871287129 s (00:03.1)
+- **frame / control_step:** 158 (policy_step=26, chunk_index=2)
+- **coarse onset:** 3.0 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempts to grasp the basket handle to flip it, but fails to close securely on the handle and slips off.
+  Summary: grasp failure on basket handle; slipping off handle
+- **recovery:** Realign the gripper over the basket handle, establish a firm clamp, and rotate the basket until it is resting upright.
+  Summary: realign gripper; grasp handle firmly; flip basket upright
+- **prevention:** Align the end-effector accurately over the handle center before closing the gripper to ensure a secure grasp before initiating the flipping motion.
+  Summary: align with handle precisely; secure grip before lifting
+- **justification:** None
diff --git a/basket/ep040/labels.json b/basket/ep040/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..eeff172abbd55fab0f5866513405a84edcbcac19
--- /dev/null
+++ b/basket/ep040/labels.json
@@ -0,0 +1,109 @@
+{
+  "rollout_id": "t002__ep040__basket",
+  "scene_variant": "",
+  "task_id": 2,
+  "task_file": "",
+  "instruction": "Flip the basket upright",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T19:49:18+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00990099009901,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for basket handle",
+      "description": "reaching for basket handle"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00990099009901,
+      "t_end_sec": 6.01980198019802,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "slipping and re-grasping basket",
+      "description": "slipping and re-grasping basket"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01980198019802,
+      "t_end_sec": 9.029702970297029,
+      "control_step_start": 304,
+      "control_step_end": 455,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "tilting basket partially",
+      "description": "tilting basket partially"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.029702970297029,
+      "t_end_sec": 9.287128712871286,
+      "control_step_start": 456,
+      "control_step_end": 468,
+      "policy_step_start": 76,
+      "policy_step_end": 78,
+      "phrase": "holding basket stationary",
+      "description": "holding basket stationary"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 3.128712871287129,
+    "vlm_failure_onset_timestamp": "00:03.1",
+    "vlm_failure_onset_frame": 158,
+    "vlm_failure_onset_step": 158,
+    "vlm_coarse_failure_onset_seconds": 3.0,
+    "vlm_failure_window_start_seconds": 2.5,
+    "vlm_failure_window_end_seconds": 3.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempts to grasp the basket handle to flip it, but fails to close securely on the handle and slips off.",
+    "vlm_failure_reason_summary": "grasp failure on basket handle; slipping off handle",
+    "vlm_recovery_action": "Realign the gripper over the basket handle, establish a firm clamp, and rotate the basket until it is resting upright.",
+    "vlm_recovery_action_summary": "realign gripper; grasp handle firmly; flip basket upright",
+    "vlm_prevention_action": "Align the end-effector accurately over the handle center before closing the gripper to ensure a secure grasp before initiating the flipping motion.",
+    "vlm_prevention_action_summary": "align with handle precisely; secure grip before lifting",
+    "vlm_justification": null,
+    "vlm_temporal_justification": "At clip second 158 (02:38), the robot arm begins moving upward to flip the basket, but the gripper fails to maintain a secure grip on the handle, causing it to slip off without moving the basket.",
+    "vlm_video_fps": 50.5,
+    "vlm_video_num_frames": 469,
+    "vlm_video_duration_seconds": 9.267326732673267,
+    "vlm_usage": {
+      "input_tokens": 21620,
+      "output_tokens": 317,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 3.0,\n \"failure_window_start_seconds\": 2.5,\n \"failure_window_end_seconds\": 3.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempts to grasp the basket handle to flip it, but fails to close securely on the handle and slips off.\",\n \"failure_reason_summary\": \"grasp failure on basket handle; slipping off handle\",\n \"recovery_action\": \"Realign the gripper over the basket handle, establish a firm clamp, and rotate the basket until it is resting upright.\",\n \"recovery_action_summary\": \"realign gripper; grasp handle firmly; flip basket upright\",\n \"prevention_action\": \"Align the end-effector accurately over the handle center before closing the gripper to ensure a secure grasp before initiating the flipping motion.\",\n \"prevention_action_summary\": \"align with handle precisely; secure grip before lifting\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 158,
+    "failure_sim_step": 158,
+    "failure_policy_step": 26,
+    "failure_chunk_index": 2,
+    "first_post_failure_policy_step": 27,
+    "failure_type": "grasp_failure",
+    "correction_action": "Realign the gripper over the basket handle, establish a firm clamp, and rotate the basket until it is resting upright."
+  }
+}
\ No newline at end of file
diff --git a/basket/ep040/labels.npz b/basket/ep040/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..c520076cf72c3b0312cab3caa13cd997b72c77c8
--- /dev/null
+++ b/basket/ep040/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d78613bc1f25ed0faf3689be010e063420c21ccaf86152857638712ff0d82527
+size 3830
diff --git a/basket/ep041/example.md b/basket/ep041/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..4a451cec3bc9583676372448c3092d5a0393116b
--- /dev/null
+++ b/basket/ep041/example.md
@@ -0,0 +1,32 @@
+# t002__ep041__basket
+
+- **suite / variant:** `` / ``
+- **instruction:** Flip the basket upright
+- **success:** False
+- **control steps:** 303  (5.9s at 51 Hz)
+- **policy inferences:** 52  (replan_steps=6)
+- **video:** `basket/ep041/rollout.mp4`
+- **features:** `52` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–152 | 0–25 | approaching and grasping basket rim |
+| 3.0–5.9 | 153–302 | 25–50 | slipping off rim and retracting |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 3.176470588235294 s (00:03.2)
+- **frame / control_step:** 162 (policy_step=27, chunk_index=0)
+- **coarse onset:** 3.0 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the edge of the red basket to flip it upright, but failed to secure a firm grip, causing the gripper to slip off without flipping the object.
+  Summary: failed grasp on basket lip; gripper slipping off
+- **recovery:** Re-align the end effector with the rim of the red basket, secure a firm grasp on the edge, and lift and rotate it until it rests upright.
+  Summary: re-grasp basket rim; flip basket upright
+- **prevention:** Position the gripper deeper over the rim of the red basket before closing the fingers to establish a secure hold before attempting the flipping motion.
+  Summary: deepen grip on basket rim; rotate securely upright
+- **justification:** None
diff --git a/basket/ep041/labels.json b/basket/ep041/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..318c803b2679a0232ee7ac62474081997205b0bf
--- /dev/null
+++ b/basket/ep041/labels.json
@@ -0,0 +1,87 @@
+{
+  "rollout_id": "t002__ep041__basket",
+  "scene_variant": "",
+  "task_id": 2,
+  "task_file": "",
+  "instruction": "Flip the basket upright",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T19:49:47+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0,
+      "control_step_start": 0,
+      "control_step_end": 152,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "approaching and grasping basket rim",
+      "description": "approaching and grasping basket rim"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0,
+      "t_end_sec": 5.9411764705882355,
+      "control_step_start": 153,
+      "control_step_end": 302,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "slipping off rim and retracting",
+      "description": "slipping off rim and retracting"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 3.176470588235294,
+    "vlm_failure_onset_timestamp": "00:03.2",
+    "vlm_failure_onset_frame": 162,
+    "vlm_failure_onset_step": 162,
+    "vlm_coarse_failure_onset_seconds": 3.0,
+    "vlm_failure_window_start_seconds": 2.5,
+    "vlm_failure_window_end_seconds": 3.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the edge of the red basket to flip it upright, but failed to secure a firm grip, causing the gripper to slip off without flipping the object.",
+    "vlm_failure_reason_summary": "failed grasp on basket lip; gripper slipping off",
+    "vlm_recovery_action": "Re-align the end effector with the rim of the red basket, secure a firm grasp on the edge, and lift and rotate it until it rests upright.",
+    "vlm_recovery_action_summary": "re-grasp basket rim; flip basket upright",
+    "vlm_prevention_action": "Position the gripper deeper over the rim of the red basket before closing the fingers to establish a secure hold before attempting the flipping motion.",
+    "vlm_prevention_action_summary": "deepen grip on basket rim; rotate securely upright",
+    "vlm_justification": null,
+    "vlm_temporal_justification": "At 02:42 (frame 162), the robot gripper slips off the black handle/cord of the red basket while trying to lift and flip it, causing the basket to fall back flat onto the table without flipping upright.",
+    "vlm_video_fps": 51.0,
+    "vlm_video_num_frames": 303,
+    "vlm_video_duration_seconds": 5.921568627450981,
+    "vlm_usage": {
+      "input_tokens": 21316,
+      "output_tokens": 338,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 3.0,\n \"failure_window_start_seconds\": 2.5,\n \"failure_window_end_seconds\": 3.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the edge of the red basket to flip it upright, but failed to secure a firm grip, causing the gripper to slip off without flipping the object.\",\n \"failure_reason_summary\": \"failed grasp on basket lip; gripper slipping off\",\n \"recovery_action\": \"Re-align the end effector with the rim of the red basket, secure a firm grasp on the edge, and lift and rotate it until it rests upright.\",\n \"recovery_action_summary\": \"re-grasp basket rim; flip basket upright\",\n \"prevention_action\": \"Position the gripper deeper over the rim of the red basket before closing the fingers to establish a secure hold before attempting the flipping motion.\",\n \"prevention_action_summary\": \"deepen grip on basket rim; rotate securely upright\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 162,
+    "failure_sim_step": 162,
+    "failure_policy_step": 27,
+    "failure_chunk_index": 0,
+    "first_post_failure_policy_step": 28,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-align the end effector with the rim of the red basket, secure a firm grasp on the edge, and lift and rotate it until it rests upright."
+  }
+}
\ No newline at end of file
diff --git a/basket/ep041/labels.npz b/basket/ep041/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..2ce9554f4f2b09dbd2e0a308d2bff914a880f604
--- /dev/null
+++ b/basket/ep041/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:baae7846a97e4d9c2677ddcbab8996a3c378d04cff1ad9f0d5cb7b313cb4d25b
+size 3783
diff --git a/basket/ep042/example.md b/basket/ep042/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..42192e0636ab613ebb5942bb8ef3021b44ff911b
--- /dev/null
+++ b/basket/ep042/example.md
@@ -0,0 +1,32 @@
+# t002__ep042__basket
+
+- **suite / variant:** `` / ``
+- **instruction:** Flip the basket upright
+- **success:** False
+- **control steps:** 263  (5.1s at 52 Hz)
+- **policy inferences:** 46  (replan_steps=6)
+- **video:** `basket/ep042/rollout.mp4`
+- **features:** `46` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–154 | 0–25 | reaching for and grasping basket edge |
+| 3.0–5.1 | 155–262 | 25–43 | slipping off basket edge |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 2.5968992248062013 s (00:02.6)
+- **frame / control_step:** 134 (policy_step=22, chunk_index=2)
+- **coarse onset:** 3.0 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the edge of the red basket to flip it, but failed to establish a secure hold, causing the gripper to slip off without flipping the basket upright.
+  Summary: lost grip on basket edge; failed to flip basket
+- **recovery:** Re-align the gripper with the edge of the red basket, establish a firm grip, and rotate the basket to flip it upright.
+  Summary: re-grasp basket edge; rotate and flip upright
+- **prevention:** Position the gripper deeper around the rim of the basket before lifting so that the fingers do not slip off during the rotation.
+  Summary: ensure deep pinch grip; rotate basket cleanly
+- **justification:** Around 3.0 seconds, the robot attempts to pull and rotate the red basket upward, but the gripper slips off the edge due to an insecure grasp, leaving the basket upside down on the table.
diff --git a/basket/ep042/labels.json b/basket/ep042/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..1769f77374fed00629ed1501030be1254d58ca3d
--- /dev/null
+++ b/basket/ep042/labels.json
@@ -0,0 +1,87 @@
+{
+  "rollout_id": "t002__ep042__basket",
+  "scene_variant": "",
+  "task_id": 2,
+  "task_file": "",
+  "instruction": "Flip the basket upright",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T19:50:20+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.003875968992248,
+      "control_step_start": 0,
+      "control_step_end": 154,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for and grasping basket edge",
+      "description": "reaching for and grasping basket edge"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.003875968992248,
+      "t_end_sec": 5.096899224806202,
+      "control_step_start": 155,
+      "control_step_end": 262,
+      "policy_step_start": 25,
+      "policy_step_end": 43,
+      "phrase": "slipping off basket edge",
+      "description": "slipping off basket edge"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 2.5968992248062013,
+    "vlm_failure_onset_timestamp": "00:02.6",
+    "vlm_failure_onset_frame": 134,
+    "vlm_failure_onset_step": 134,
+    "vlm_coarse_failure_onset_seconds": 3.0,
+    "vlm_failure_window_start_seconds": 2.7,
+    "vlm_failure_window_end_seconds": 3.7,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the edge of the red basket to flip it, but failed to establish a secure hold, causing the gripper to slip off without flipping the basket upright.",
+    "vlm_failure_reason_summary": "lost grip on basket edge; failed to flip basket",
+    "vlm_recovery_action": "Re-align the gripper with the edge of the red basket, establish a firm grip, and rotate the basket to flip it upright.",
+    "vlm_recovery_action_summary": "re-grasp basket edge; rotate and flip upright",
+    "vlm_prevention_action": "Position the gripper deeper around the rim of the basket before lifting so that the fingers do not slip off during the rotation.",
+    "vlm_prevention_action_summary": "ensure deep pinch grip; rotate basket cleanly",
+    "vlm_justification": "Around 3.0 seconds, the robot attempts to pull and rotate the red basket upward, but the gripper slips off the edge due to an insecure grasp, leaving the basket upside down on the table.",
+    "vlm_temporal_justification": "At timestamp 02:14 (second 134 of the clip), the robot gripper slips off the edge of the red basket while trying to lift and flip it, leaving the basket on the table and failing to establish a secure hold.",
+    "vlm_video_fps": 51.6,
+    "vlm_video_num_frames": 263,
+    "vlm_video_duration_seconds": 5.077519379844961,
+    "vlm_usage": {
+      "input_tokens": 18733,
+      "output_tokens": 379,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 3.0,\n \"failure_window_start_seconds\": 2.7,\n \"failure_window_end_seconds\": 3.7,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the edge of the red basket to flip it, but failed to establish a secure hold, causing the gripper to slip off without flipping the basket upright.\",\n \"failure_reason_summary\": \"lost grip on basket edge; failed to flip basket\",\n \"recovery_action\": \"Re-align the gripper with the edge of the red basket, establish a firm grip, and rotate the basket to flip it upright.\",\n \"recovery_action_summary\": \"re-grasp basket edge; rotate and flip upright\",\n \"prevention_action\": \"Position the gripper deeper around the rim of the basket before lifting so that the fingers do not slip off during the rotation.\",\n \"prevention_action_summary\": \"ensure deep pinch grip; rotate basket cleanly\",\n \"justification\": \"Around 3.0 seconds, the robot attempts to pull and rotate the red basket upward, but the gripper slips off the edge due to an insecure grasp, leaving the basket upside down on the table.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 134,
+    "failure_sim_step": 134,
+    "failure_policy_step": 22,
+    "failure_chunk_index": 2,
+    "first_post_failure_policy_step": 23,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-align the gripper with the edge of the red basket, establish a firm grip, and rotate the basket to flip it upright."
+  }
+}
\ No newline at end of file
diff --git a/basket/ep042/labels.npz b/basket/ep042/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..af9887700bb79c074cc92b976ef525258c4ab38d
--- /dev/null
+++ b/basket/ep042/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:790aebc7cdfb68ce7c9ecebeee037f6b47f9e0fb863009f3e22aad7d7311d2d6
+size 3742
diff --git a/basket/ep043/example.md b/basket/ep043/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..2447285072892704b1ec01419ccd0eea98a99882
--- /dev/null
+++ b/basket/ep043/example.md
@@ -0,0 +1,34 @@
+# t002__ep043__basket
+
+- **suite / variant:** `` / ``
+- **instruction:** Flip the basket upright
+- **success:** False
+- **control steps:** 510  (10.2s at 50 Hz)
+- **policy inferences:** 88  (replan_steps=6)
+- **video:** `basket/ep043/rollout.mp4`
+- **features:** `88` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–149 | 0–24 | reaching down toward basket handle |
+| 3.0–6.0 | 150–299 | 25–49 | lifting arm and losing grip |
+| 6.0–9.0 | 300–449 | 50–74 | re-approaching basket handle to grasp |
+| 9.0–10.2 | 450–509 | 75–84 | lifting gripper without turning basket |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 4.491017964071856 s (00:04.5)
+- **frame / control_step:** 225 (policy_step=37, chunk_index=3)
+- **coarse onset:** 3.5 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the basket's strap handle but failed to achieve a firm grip, allowing the handle to slip out when lifting.
+  Summary: failed grasp on basket handle; gripper slippage
+- **recovery:** Reposition the gripper accurately over the strap handle or basket edge, close firmly, and lift to flip the basket upright.
+  Summary: reposition gripper; securely grasp handle; flip basket upright
+- **prevention:** Align the end-effector accurately with the center of the strap handle and close the fingers tightly before executing the lifting and flipping motion.
+  Summary: align gripper with handle; close securely before lifting
+- **justification:** None
diff --git a/basket/ep043/labels.json b/basket/ep043/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..5ad61165cbc55606ad42a8ad84786c3e578b810c
--- /dev/null
+++ b/basket/ep043/labels.json
@@ -0,0 +1,109 @@
+{
+  "rollout_id": "t002__ep043__basket",
+  "scene_variant": "",
+  "task_id": 2,
+  "task_file": "",
+  "instruction": "Flip the basket upright",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T19:50:57+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9940119760479043,
+      "control_step_start": 0,
+      "control_step_end": 149,
+      "policy_step_start": 0,
+      "policy_step_end": 24,
+      "phrase": "reaching down toward basket handle",
+      "description": "reaching down toward basket handle"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9940119760479043,
+      "t_end_sec": 5.9880239520958085,
+      "control_step_start": 150,
+      "control_step_end": 299,
+      "policy_step_start": 25,
+      "policy_step_end": 49,
+      "phrase": "lifting arm and losing grip",
+      "description": "lifting arm and losing grip"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.9880239520958085,
+      "t_end_sec": 8.982035928143713,
+      "control_step_start": 300,
+      "control_step_end": 449,
+      "policy_step_start": 50,
+      "policy_step_end": 74,
+      "phrase": "re-approaching basket handle to grasp",
+      "description": "re-approaching basket handle to grasp"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.982035928143713,
+      "t_end_sec": 10.179640718562874,
+      "control_step_start": 450,
+      "control_step_end": 509,
+      "policy_step_start": 75,
+      "policy_step_end": 84,
+      "phrase": "lifting gripper without turning basket",
+      "description": "lifting gripper without turning basket"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 4.491017964071856,
+    "vlm_failure_onset_timestamp": "00:04.5",
+    "vlm_failure_onset_frame": 225,
+    "vlm_failure_onset_step": 225,
+    "vlm_coarse_failure_onset_seconds": 3.5,
+    "vlm_failure_window_start_seconds": 3.0,
+    "vlm_failure_window_end_seconds": 4.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the basket's strap handle but failed to achieve a firm grip, allowing the handle to slip out when lifting.",
+    "vlm_failure_reason_summary": "failed grasp on basket handle; gripper slippage",
+    "vlm_recovery_action": "Reposition the gripper accurately over the strap handle or basket edge, close firmly, and lift to flip the basket upright.",
+    "vlm_recovery_action_summary": "reposition gripper; securely grasp handle; flip basket upright",
+    "vlm_prevention_action": "Align the end-effector accurately with the center of the strap handle and close the fingers tightly before executing the lifting and flipping motion.",
+    "vlm_prevention_action_summary": "align gripper with handle; close securely before lifting",
+    "vlm_justification": null,
+    "vlm_temporal_justification": "The robot closes its gripper around the basket's strap handle and begins lifting upward. At timestamp 03:20 (200 seconds into the clip), the gripper moves upward while the strap handle slips out and stays flat on the basket, revealing the failed grasp.",
+    "vlm_video_fps": 50.1,
+    "vlm_video_num_frames": 510,
+    "vlm_video_duration_seconds": 10.159680638722554,
+    "vlm_usage": {
+      "input_tokens": 21440,
+      "output_tokens": 330,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 3.5,\n \"failure_window_start_seconds\": 3.0,\n \"failure_window_end_seconds\": 4.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the basket's strap handle but failed to achieve a firm grip, allowing the handle to slip out when lifting.\",\n \"failure_reason_summary\": \"failed grasp on basket handle; gripper slippage\",\n \"recovery_action\": \"Reposition the gripper accurately over the strap handle or basket edge, close firmly, and lift to flip the basket upright.\",\n \"recovery_action_summary\": \"reposition gripper; securely grasp handle; flip basket upright\",\n \"prevention_action\": \"Align the end-effector accurately with the center of the strap handle and close the fingers tightly before executing the lifting and flipping motion.\",\n \"prevention_action_summary\": \"align gripper with handle; close securely before lifting\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 225,
+    "failure_sim_step": 225,
+    "failure_policy_step": 37,
+    "failure_chunk_index": 3,
+    "first_post_failure_policy_step": 38,
+    "failure_type": "grasp_failure",
+    "correction_action": "Reposition the gripper accurately over the strap handle or basket edge, close firmly, and lift to flip the basket upright."
+  }
+}
\ No newline at end of file
diff --git a/basket/ep043/labels.npz b/basket/ep043/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..bcb8134473d67e7b9ee7ff87f74a5339143bbeb8
--- /dev/null
+++ b/basket/ep043/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b139e26977e8b3d1b3b0d41547eb68e80aca23899f91e8307016d2f95d2c2280
+size 3897
diff --git a/basket/ep044/example.md b/basket/ep044/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..10e0d58b75826485b8797e0b4d73c0290fde83cd
--- /dev/null
+++ b/basket/ep044/example.md
@@ -0,0 +1,33 @@
+# t002__ep044__basket
+
+- **suite / variant:** `` / ``
+- **instruction:** Flip the basket upright
+- **success:** False
+- **control steps:** 432  (8.6s at 50 Hz)
+- **policy inferences:** 74  (replan_steps=6)
+- **video:** `basket/ep044/rollout.mp4`
+- **features:** `74` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | lowering gripper toward basket handle |
+| 3.0–6.0 | 151–301 | 25–50 | closing gripper and missing handle |
+| 6.0–8.6 | 302–431 | 50–71 | pressing and hovering over basket |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 3.2738095238095237 s (00:03.3)
+- **frame / control_step:** 165 (policy_step=27, chunk_index=3)
+- **coarse onset:** 3.0 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the handle of the upside-down basket but failed to secure a grip, causing the fingers to close on air and slip off.
+  Summary: failed grasp on basket handle; slipping off object
+- **recovery:** Re-align the gripper with the basket's handle or edge, establish a firm grasp, and rotate the basket to flip it upright.
+  Summary: re-align gripper with handle; grasp securely; flip basket upright
+- **prevention:** Accurately position the end effector over the basket handle and ensure proper finger alignment before closing the gripper to flip the basket.
+  Summary: align gripper with handle; secure firm grip; rotate basket
+- **justification:** Around 3.0 seconds into the rollout, the robot closes its gripper near the red basket's black handle but fails to secure a grip. It continues moving around the object without holding it, resulting in a failed task execution.
diff --git a/basket/ep044/labels.json b/basket/ep044/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..90edb3e80fc7e7d601e9ae3c79169db83231d29b
--- /dev/null
+++ b/basket/ep044/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t002__ep044__basket",
+  "scene_variant": "",
+  "task_id": 2,
+  "task_file": "",
+  "instruction": "Flip the basket upright",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T19:51:29+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.996031746031746,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "lowering gripper toward basket handle",
+      "description": "lowering gripper toward basket handle"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.996031746031746,
+      "t_end_sec": 5.992063492063492,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "closing gripper and missing handle",
+      "description": "closing gripper and missing handle"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.992063492063492,
+      "t_end_sec": 8.571428571428571,
+      "control_step_start": 302,
+      "control_step_end": 431,
+      "policy_step_start": 50,
+      "policy_step_end": 71,
+      "phrase": "pressing and hovering over basket",
+      "description": "pressing and hovering over basket"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 3.2738095238095237,
+    "vlm_failure_onset_timestamp": "00:03.3",
+    "vlm_failure_onset_frame": 165,
+    "vlm_failure_onset_step": 165,
+    "vlm_coarse_failure_onset_seconds": 3.0,
+    "vlm_failure_window_start_seconds": 2.5,
+    "vlm_failure_window_end_seconds": 4.0,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the handle of the upside-down basket but failed to secure a grip, causing the fingers to close on air and slip off.",
+    "vlm_failure_reason_summary": "failed grasp on basket handle; slipping off object",
+    "vlm_recovery_action": "Re-align the gripper with the basket's handle or edge, establish a firm grasp, and rotate the basket to flip it upright.",
+    "vlm_recovery_action_summary": "re-align gripper with handle; grasp securely; flip basket upright",
+    "vlm_prevention_action": "Accurately position the end effector over the basket handle and ensure proper finger alignment before closing the gripper to flip the basket.",
+    "vlm_prevention_action_summary": "align gripper with handle; secure firm grip; rotate basket",
+    "vlm_justification": "Around 3.0 seconds into the rollout, the robot closes its gripper near the red basket's black handle but fails to secure a grip. It continues moving around the object without holding it, resulting in a failed task execution.",
+    "vlm_temporal_justification": "At clip time 02:45 (165 seconds), the robot's gripper fingers finish closing without capturing the handle loop, closing completely on air and demonstrating a clear grasp failure.",
+    "vlm_video_fps": 50.4,
+    "vlm_video_num_frames": 432,
+    "vlm_video_duration_seconds": 8.551587301587302,
+    "vlm_usage": {
+      "input_tokens": 21501,
+      "output_tokens": 377,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 3.0,\n \"failure_window_start_seconds\": 2.5,\n \"failure_window_end_seconds\": 4.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the handle of the upside-down basket but failed to secure a grip, causing the fingers to close on air and slip off.\",\n \"failure_reason_summary\": \"failed grasp on basket handle; slipping off object\",\n \"recovery_action\": \"Re-align the gripper with the basket's handle or edge, establish a firm grasp, and rotate the basket to flip it upright.\",\n \"recovery_action_summary\": \"re-align gripper with handle; grasp securely; flip basket upright\",\n \"prevention_action\": \"Accurately position the end effector over the basket handle and ensure proper finger alignment before closing the gripper to flip the basket.\",\n \"prevention_action_summary\": \"align gripper with handle; secure firm grip; rotate basket\",\n \"justification\": \"Around 3.0 seconds into the rollout, the robot closes its gripper near the red basket's black handle but fails to secure a grip. It continues moving around the object without holding it, resulting in a failed task execution.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 165,
+    "failure_sim_step": 165,
+    "failure_policy_step": 27,
+    "failure_chunk_index": 3,
+    "first_post_failure_policy_step": 28,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-align the gripper with the basket's handle or edge, establish a firm grasp, and rotate the basket to flip it upright."
+  }
+}
\ No newline at end of file
diff --git a/basket/ep044/labels.npz b/basket/ep044/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..128c093ab417afabd09d8d8bad9097e056b99318
--- /dev/null
+++ b/basket/ep044/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:114d851eb7d8aa0615557415b5dbbbe9b059b8ca8925c4d347c4e665306514d0
+size 3837
diff --git a/basket/ep045/example.md b/basket/ep045/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..b3b68d7638fca2c066a1bae7b0385ff071ef6860
--- /dev/null
+++ b/basket/ep045/example.md
@@ -0,0 +1,33 @@
+# t002__ep045__basket
+
+- **suite / variant:** `` / ``
+- **instruction:** Flip the basket upright
+- **success:** False
+- **control steps:** 361  (7.1s at 51 Hz)
+- **policy inferences:** 63  (replan_steps=6)
+- **video:** `basket/ep045/rollout.mp4`
+- **features:** `63` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–153 | 0–25 | reaching and attempting handle grasp |
+| 3.0–6.0 | 154–307 | 25–51 | slipping off handle and retracting |
+| 6.0–7.1 | 308–360 | 51–60 | hovering above basket; idle |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 2.63671875 s (00:02.6)
+- **frame / control_step:** 135 (policy_step=22, chunk_index=3)
+- **coarse onset:** 2.8 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the handle on the red basket to flip it, but failed to secure a grip and slipped off the handle.
+  Summary: failed grasp on basket handle; gripper slipped off
+- **recovery:** Re-align the gripper over the basket handle, form a secure grip, and lift while rotating to flip the basket upright.
+  Summary: re-grasp handle securely; rotate basket upright
+- **prevention:** Center the end-effector accurately over the handle and ensure full closure before initiating the upward flipping motion.
+  Summary: center gripper on handle; secure grip before flipping
+- **justification:** At around 2.8 seconds, the robot's gripper attempts to grasp the black handle on top of the red basket, but fails to grip it securely and slips off. The arm then moves back up without having flipped the basket.
diff --git a/basket/ep045/labels.json b/basket/ep045/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..64994cc0eeb2831fc4c465dc30ebb52023744831
--- /dev/null
+++ b/basket/ep045/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t002__ep045__basket",
+  "scene_variant": "",
+  "task_id": 2,
+  "task_file": "",
+  "instruction": "Flip the basket upright",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T19:51:59+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0078125,
+      "control_step_start": 0,
+      "control_step_end": 153,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching and attempting handle grasp",
+      "description": "reaching and attempting handle grasp"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0078125,
+      "t_end_sec": 6.015625,
+      "control_step_start": 154,
+      "control_step_end": 307,
+      "policy_step_start": 25,
+      "policy_step_end": 51,
+      "phrase": "slipping off handle and retracting",
+      "description": "slipping off handle and retracting"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.015625,
+      "t_end_sec": 7.05078125,
+      "control_step_start": 308,
+      "control_step_end": 360,
+      "policy_step_start": 51,
+      "policy_step_end": 60,
+      "phrase": "hovering above basket; idle",
+      "description": "hovering above basket; idle"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 2.63671875,
+    "vlm_failure_onset_timestamp": "00:02.6",
+    "vlm_failure_onset_frame": 135,
+    "vlm_failure_onset_step": 135,
+    "vlm_coarse_failure_onset_seconds": 2.8,
+    "vlm_failure_window_start_seconds": 2.3,
+    "vlm_failure_window_end_seconds": 3.3,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the handle on the red basket to flip it, but failed to secure a grip and slipped off the handle.",
+    "vlm_failure_reason_summary": "failed grasp on basket handle; gripper slipped off",
+    "vlm_recovery_action": "Re-align the gripper over the basket handle, form a secure grip, and lift while rotating to flip the basket upright.",
+    "vlm_recovery_action_summary": "re-grasp handle securely; rotate basket upright",
+    "vlm_prevention_action": "Center the end-effector accurately over the handle and ensure full closure before initiating the upward flipping motion.",
+    "vlm_prevention_action_summary": "center gripper on handle; secure grip before flipping",
+    "vlm_justification": "At around 2.8 seconds, the robot's gripper attempts to grasp the black handle on top of the red basket, but fails to grip it securely and slips off. The arm then moves back up without having flipped the basket.",
+    "vlm_temporal_justification": "At second 135 (02:15 in the clip), the robot gripper closes its fingers near the basket's black handle but fails to secure a grip, slipping off the handle completely as it attempts to move.",
+    "vlm_video_fps": 51.2,
+    "vlm_video_num_frames": 361,
+    "vlm_video_duration_seconds": 7.03125,
+    "vlm_usage": {
+      "input_tokens": 21056,
+      "output_tokens": 367,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 2.8,\n \"failure_window_start_seconds\": 2.3,\n \"failure_window_end_seconds\": 3.3,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the handle on the red basket to flip it, but failed to secure a grip and slipped off the handle.\",\n \"failure_reason_summary\": \"failed grasp on basket handle; gripper slipped off\",\n \"recovery_action\": \"Re-align the gripper over the basket handle, form a secure grip, and lift while rotating to flip the basket upright.\",\n \"recovery_action_summary\": \"re-grasp handle securely; rotate basket upright\",\n \"prevention_action\": \"Center the end-effector accurately over the handle and ensure full closure before initiating the upward flipping motion.\",\n \"prevention_action_summary\": \"center gripper on handle; secure grip before flipping\",\n \"justification\": \"At around 2.8 seconds, the robot's gripper attempts to grasp the black handle on top of the red basket, but fails to grip it securely and slips off. The arm then moves back up without having flipped the basket.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 135,
+    "failure_sim_step": 135,
+    "failure_policy_step": 22,
+    "failure_chunk_index": 3,
+    "first_post_failure_policy_step": 23,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-align the gripper over the basket handle, form a secure grip, and lift while rotating to flip the basket upright."
+  }
+}
\ No newline at end of file
diff --git a/basket/ep045/labels.npz b/basket/ep045/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..abf2453ea178a5e925a0c18baf282f8c576d7f35
--- /dev/null
+++ b/basket/ep045/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:758e37b7ca35d874f83afe56161f7561ddd2e8b0580284d30bda09534ca4a9dc
+size 3759
diff --git a/basket/ep046/example.md b/basket/ep046/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..09199cff5b49885fc8699af745b30f780c359b45
--- /dev/null
+++ b/basket/ep046/example.md
@@ -0,0 +1,34 @@
+# t002__ep046__basket
+
+- **suite / variant:** `` / ``
+- **instruction:** Flip the basket upright
+- **success:** False
+- **control steps:** 499  (9.9s at 50 Hz)
+- **policy inferences:** 86  (replan_steps=6)
+- **video:** `basket/ep046/rollout.mp4`
+- **features:** `86` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | reaching and grasping basket handle |
+| 3.0–6.0 | 151–301 | 25–50 | releasing handle and repositioning gripper |
+| 6.0–9.0 | 302–452 | 50–75 | attempting second grasp on handle |
+| 9.0–9.9 | 453–498 | 75–83 | holding position near basket |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 3.1013916500994037 s (00:03.1)
+- **frame / control_step:** 156 (policy_step=26, chunk_index=0)
+- **coarse onset:** 2.5 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the handle of the basket to flip it upright, but failed to secure a firm grip, causing the handle to slip out during the attempt.
+  Summary: failed grasp on basket handle; handle slipping
+- **recovery:** Realign the gripper directly over the handle, close the fingers tightly to establish a firm grasp, and lift and rotate the basket until it rests upright.
+  Summary: realign gripper; firmly grasp handle; flip basket upright
+- **prevention:** Align the end-effector accurately with the basket handle and achieve full closure around the handle before executing the lifting and flipping motion.
+  Summary: align gripper with handle; secure firm grip before flipping
+- **justification:** None
diff --git a/basket/ep046/labels.json b/basket/ep046/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..94f8f7f7a3aa6970e00a8f93eec18185e09c454d
--- /dev/null
+++ b/basket/ep046/labels.json
@@ -0,0 +1,109 @@
+{
+  "rollout_id": "t002__ep046__basket",
+  "scene_variant": "",
+  "task_id": 2,
+  "task_file": "",
+  "instruction": "Flip the basket upright",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T19:52:39+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.001988071570577,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching and grasping basket handle",
+      "description": "reaching and grasping basket handle"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.001988071570577,
+      "t_end_sec": 6.003976143141154,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "releasing handle and repositioning gripper",
+      "description": "releasing handle and repositioning gripper"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.003976143141154,
+      "t_end_sec": 9.00596421471173,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "attempting second grasp on handle",
+      "description": "attempting second grasp on handle"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.00596421471173,
+      "t_end_sec": 9.920477137176938,
+      "control_step_start": 453,
+      "control_step_end": 498,
+      "policy_step_start": 75,
+      "policy_step_end": 83,
+      "phrase": "holding position near basket",
+      "description": "holding position near basket"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 3.1013916500994037,
+    "vlm_failure_onset_timestamp": "00:03.1",
+    "vlm_failure_onset_frame": 156,
+    "vlm_failure_onset_step": 156,
+    "vlm_coarse_failure_onset_seconds": 2.5,
+    "vlm_failure_window_start_seconds": 2.0,
+    "vlm_failure_window_end_seconds": 3.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the handle of the basket to flip it upright, but failed to secure a firm grip, causing the handle to slip out during the attempt.",
+    "vlm_failure_reason_summary": "failed grasp on basket handle; handle slipping",
+    "vlm_recovery_action": "Realign the gripper directly over the handle, close the fingers tightly to establish a firm grasp, and lift and rotate the basket until it rests upright.",
+    "vlm_recovery_action_summary": "realign gripper; firmly grasp handle; flip basket upright",
+    "vlm_prevention_action": "Align the end-effector accurately with the basket handle and achieve full closure around the handle before executing the lifting and flipping motion.",
+    "vlm_prevention_action_summary": "align gripper with handle; secure firm grip before flipping",
+    "vlm_justification": null,
+    "vlm_temporal_justification": "At clip timestamp 02:36 (frame 156), the robot's end-effector closes its fingers around the basket's handle, but fails to establish a secure hold. This poor grasp alignment immediately results in the handle slipping out as the robot begins lifting.",
+    "vlm_video_fps": 50.3,
+    "vlm_video_num_frames": 499,
+    "vlm_video_duration_seconds": 9.900596421471173,
+    "vlm_usage": {
+      "input_tokens": 19991,
+      "output_tokens": 342,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 2.5,\n \"failure_window_start_seconds\": 2.0,\n \"failure_window_end_seconds\": 3.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the handle of the basket to flip it upright, but failed to secure a firm grip, causing the handle to slip out during the attempt.\",\n \"failure_reason_summary\": \"failed grasp on basket handle; handle slipping\",\n \"recovery_action\": \"Realign the gripper directly over the handle, close the fingers tightly to establish a firm grasp, and lift and rotate the basket until it rests upright.\",\n \"recovery_action_summary\": \"realign gripper; firmly grasp handle; flip basket upright\",\n \"prevention_action\": \"Align the end-effector accurately with the basket handle and achieve full closure around the handle before executing the lifting and flipping motion.\",\n \"prevention_action_summary\": \"align gripper with handle; secure firm grip before flipping\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 156,
+    "failure_sim_step": 156,
+    "failure_policy_step": 26,
+    "failure_chunk_index": 0,
+    "first_post_failure_policy_step": 27,
+    "failure_type": "grasp_failure",
+    "correction_action": "Realign the gripper directly over the handle, close the fingers tightly to establish a firm grasp, and lift and rotate the basket until it rests upright."
+  }
+}
\ No newline at end of file
diff --git a/basket/ep046/labels.npz b/basket/ep046/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..ad24541bec086ecc27bcd9262c03b1ce2ec14cf8
--- /dev/null
+++ b/basket/ep046/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9930aebb47fc8746719eef5179d892d7dc0ca1482ee51f17f586ea1dec5bfdcb
+size 3911
diff --git a/basket/ep047/example.md b/basket/ep047/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..edafe17b82c0aa6e9bb79a646cedd2d812abb4f0
--- /dev/null
+++ b/basket/ep047/example.md
@@ -0,0 +1,32 @@
+# t002__ep047__basket
+
+- **suite / variant:** `` / ``
+- **instruction:** Flip the basket upright
+- **success:** False
+- **control steps:** 289  (5.7s at 50 Hz)
+- **policy inferences:** 50  (replan_steps=6)
+- **video:** `basket/ep047/rollout.mp4`
+- **features:** `50` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | reaching for and grasping basket |
+| 3.0–5.7 | 151–288 | 25–48 | dropping basket and retracting gripper |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 2.678571428571429 s (00:02.7)
+- **frame / control_step:** 135 (policy_step=22, chunk_index=3)
+- **coarse onset:** 3.2 s (refined=True)
+- **mode:** object_displacement
+- **reason:** The robot attempted to lift and flip the red basket, but lost its grip mid-air, causing the basket to drop back onto the table without being flipped upright.
+  Summary: losing grip on basket; dropping basket back onto table
+- **recovery:** Re-approach the red basket, establish a secure grip on its rim or handle, and complete the flipping motion until it rests upright on the table.
+  Summary: re-grip basket; flip basket upright
+- **prevention:** Ensure a deep and stable grasp on the basket before applying lifting and rotational force so it does not slip out mid-motion.
+  Summary: secure firm grip; complete flipping trajectory cleanly
+- **justification:** Around 3.2 seconds, the robot attempts to lift and flip the red basket, but the basket slips out of the gripper and falls back down. The robot then retracts without completing the task.
diff --git a/basket/ep047/labels.json b/basket/ep047/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..49cc6c87a0aa881f82add93366ec9a60ac617a72
--- /dev/null
+++ b/basket/ep047/labels.json
@@ -0,0 +1,87 @@
+{
+  "rollout_id": "t002__ep047__basket",
+  "scene_variant": "",
+  "task_id": 2,
+  "task_file": "",
+  "instruction": "Flip the basket upright",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T19:53:17+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.996031746031746,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for and grasping basket",
+      "description": "reaching for and grasping basket"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.996031746031746,
+      "t_end_sec": 5.734126984126984,
+      "control_step_start": 151,
+      "control_step_end": 288,
+      "policy_step_start": 25,
+      "policy_step_end": 48,
+      "phrase": "dropping basket and retracting gripper",
+      "description": "dropping basket and retracting gripper"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "object_displacement",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 2.678571428571429,
+    "vlm_failure_onset_timestamp": "00:02.7",
+    "vlm_failure_onset_frame": 135,
+    "vlm_failure_onset_step": 135,
+    "vlm_coarse_failure_onset_seconds": 3.2,
+    "vlm_failure_window_start_seconds": 2.5,
+    "vlm_failure_window_end_seconds": 4.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to lift and flip the red basket, but lost its grip mid-air, causing the basket to drop back onto the table without being flipped upright.",
+    "vlm_failure_reason_summary": "losing grip on basket; dropping basket back onto table",
+    "vlm_recovery_action": "Re-approach the red basket, establish a secure grip on its rim or handle, and complete the flipping motion until it rests upright on the table.",
+    "vlm_recovery_action_summary": "re-grip basket; flip basket upright",
+    "vlm_prevention_action": "Ensure a deep and stable grasp on the basket before applying lifting and rotational force so it does not slip out mid-motion.",
+    "vlm_prevention_action_summary": "secure firm grip; complete flipping trajectory cleanly",
+    "vlm_justification": "Around 3.2 seconds, the robot attempts to lift and flip the red basket, but the basket slips out of the gripper and falls back down. The robot then retracts without completing the task.",
+    "vlm_temporal_justification": "At 02:05 (clip second 125), the robot attempts to twist and flip the basket in mid-air, but its grip on the handle begins to slip, causing the basket to drop back onto the table at 02:06 without flipping upright.",
+    "vlm_video_fps": 50.4,
+    "vlm_video_num_frames": 289,
+    "vlm_video_duration_seconds": 5.714285714285714,
+    "vlm_usage": {
+      "input_tokens": 19805,
+      "output_tokens": 388,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"object_displacement\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 3.2,\n \"failure_window_start_seconds\": 2.5,\n \"failure_window_end_seconds\": 4.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to lift and flip the red basket, but lost its grip mid-air, causing the basket to drop back onto the table without being flipped upright.\",\n \"failure_reason_summary\": \"losing grip on basket; dropping basket back onto table\",\n \"recovery_action\": \"Re-approach the red basket, establish a secure grip on its rim or handle, and complete the flipping motion until it rests upright on the table.\",\n \"recovery_action_summary\": \"re-grip basket; flip basket upright\",\n \"prevention_action\": \"Ensure a deep and stable grasp on the basket before applying lifting and rotational force so it does not slip out mid-motion.\",\n \"prevention_action_summary\": \"secure firm grip; complete flipping trajectory cleanly\",\n \"justification\": \"Around 3.2 seconds, the robot attempts to lift and flip the red basket, but the basket slips out of the gripper and falls back down. The robot then retracts without completing the task.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 135,
+    "failure_sim_step": 135,
+    "failure_policy_step": 22,
+    "failure_chunk_index": 3,
+    "first_post_failure_policy_step": 23,
+    "failure_type": "object_displacement",
+    "correction_action": "Re-approach the red basket, establish a secure grip on its rim or handle, and complete the flipping motion until it rests upright on the table."
+  }
+}
\ No newline at end of file
diff --git a/basket/ep047/labels.npz b/basket/ep047/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..f0aae94b774e61942fc9dfeb235e9252e1055d4b
--- /dev/null
+++ b/basket/ep047/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:664fce18a887410074e80a04039c3325446a15786c0ed8a8b32a84737ee5d2bf
+size 3780
diff --git a/basket/ep048/example.md b/basket/ep048/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..6476a7aaba35b3daf76bc447f0b719969e1168a9
--- /dev/null
+++ b/basket/ep048/example.md
@@ -0,0 +1,22 @@
+# t002__ep048__basket
+
+- **suite / variant:** `` / ``
+- **instruction:** Flip the basket upright
+- **success:** True
+- **control steps:** 489  (9.7s at 50 Hz)
+- **policy inferences:** 84  (replan_steps=6)
+- **video:** `basket/ep048/rollout.mp4`
+- **features:** `84` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | reaching for the basket |
+| 3.0–6.0 | 151–301 | 25–50 | grasping the basket handle |
+| 6.0–9.0 | 302–452 | 50–75 | flipping the basket upright |
+| 9.0–9.7 | 453–488 | 75–81 | releasing the basket |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/basket/ep048/labels.json b/basket/ep048/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..60cbfd3cd74777fa5ef640c1a80237095d1bba1d
--- /dev/null
+++ b/basket/ep048/labels.json
@@ -0,0 +1,79 @@
+{
+  "rollout_id": "t002__ep048__basket",
+  "scene_variant": "",
+  "task_id": 2,
+  "task_file": "",
+  "instruction": "Flip the basket upright",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T19:53:26+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00796812749004,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for the basket",
+      "description": "reaching for the basket"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00796812749004,
+      "t_end_sec": 6.01593625498008,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "grasping the basket handle",
+      "description": "grasping the basket handle"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01593625498008,
+      "t_end_sec": 9.02390438247012,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "flipping the basket upright",
+      "description": "flipping the basket upright"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.02390438247012,
+      "t_end_sec": 9.741035856573705,
+      "control_step_start": 453,
+      "control_step_end": 488,
+      "policy_step_start": 75,
+      "policy_step_end": 81,
+      "phrase": "releasing the basket",
+      "description": "releasing the basket"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/basket/ep048/labels.npz b/basket/ep048/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..54a1f65770b323500f3d7a619451f0299799d030
--- /dev/null
+++ b/basket/ep048/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:311bcce4a741338ac0faacb5437d2f7fb9934b0d54c747573241fa520452018e
+size 3100
diff --git a/basket/ep049/example.md b/basket/ep049/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..cb33d8ca348bc7bdd76afa4ac09eb39083e26e88
--- /dev/null
+++ b/basket/ep049/example.md
@@ -0,0 +1,22 @@
+# t002__ep049__basket
+
+- **suite / variant:** `` / ``
+- **instruction:** Flip the basket upright
+- **success:** True
+- **control steps:** 497  (9.9s at 50 Hz)
+- **policy inferences:** 86  (replan_steps=6)
+- **video:** `basket/ep049/rollout.mp4`
+- **features:** `86` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | reaching for basket handle |
+| 3.0–6.0 | 151–301 | 25–50 | grasping basket handle |
+| 6.0–9.0 | 302–452 | 50–75 | flipping basket upright |
+| 9.0–9.9 | 453–496 | 75–82 | retracting arm |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/basket/ep049/labels.json b/basket/ep049/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..167599599acf5a004327f8d0f43f10b203e2845e
--- /dev/null
+++ b/basket/ep049/labels.json
@@ -0,0 +1,79 @@
+{
+  "rollout_id": "t002__ep049__basket",
+  "scene_variant": "",
+  "task_id": 2,
+  "task_file": "",
+  "instruction": "Flip the basket upright",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T19:53:35+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.996031746031746,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for basket handle",
+      "description": "reaching for basket handle"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.996031746031746,
+      "t_end_sec": 5.992063492063492,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "grasping basket handle",
+      "description": "grasping basket handle"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.992063492063492,
+      "t_end_sec": 8.988095238095239,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "flipping basket upright",
+      "description": "flipping basket upright"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.988095238095239,
+      "t_end_sec": 9.86111111111111,
+      "control_step_start": 453,
+      "control_step_end": 496,
+      "policy_step_start": 75,
+      "policy_step_end": 82,
+      "phrase": "retracting arm",
+      "description": "retracting arm"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/basket/ep049/labels.npz b/basket/ep049/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..f2273033d8ccf65cc38aac7de61e4b8edcbc9a81
--- /dev/null
+++ b/basket/ep049/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bdd51b62eb04eed3b6f2f05ef049c3620ebdb5b3d8ddc9ba2fab0cb41be076bb
+size 3102
diff --git a/blocks/ep000/example.md b/blocks/ep000/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..1898a97ccce331484aa3f7eb90a9b548eea5f87e
--- /dev/null
+++ b/blocks/ep000/example.md
@@ -0,0 +1,37 @@
+# t000__ep000__blocks
+
+- **suite / variant:** `` / ``
+- **instruction:** Stack the yellow block on the wooden block
+- **success:** False
+- **control steps:** 929  (18.3s at 51 Hz)
+- **policy inferences:** 160  (replan_steps=6)
+- **video:** `blocks/ep000/rollout.mp4`
+- **features:** `160` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–152 | 0–25 | approaching yellow block |
+| 3.0–6.0 | 153–305 | 25–50 | grasping and lifting yellow block |
+| 6.0–9.0 | 306–458 | 51–76 | releasing yellow block onto target |
+| 9.0–12.0 | 459–611 | 76–101 | hovering gripper over yellow block |
+| 12.0–15.0 | 612–764 | 102–127 | pressing down on yellow block |
+| 15.0–18.0 | 765–917 | 127–152 | holding position near yellow block |
+| 18.0–18.3 | 918–928 | 153–154 | remaining static near yellow block |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 4.675834970530452 s (00:04.7)
+- **frame / control_step:** 238 (policy_step=39, chunk_index=4)
+- **coarse onset:** 7.5 s (refined=True)
+- **mode:** placement_or_insertion_failure
+- **reason:** The robot brought the yellow block over to the wooden block but released it near the edge without proper alignment, causing the yellow block to slip off onto the tabletop.
+  Summary: misaligned placement; yellow block slipped off
+- **recovery:** Re-grasp the yellow block from the table surface, align it directly over the center of the wooden block, and gently set it down.
+  Summary: re-grasp yellow block; stack on wooden block
+- **prevention:** Position the yellow block directly over the center of the wooden block and verify alignment before opening the gripper.
+  Summary: center yellow block over target; release gently
+- **justification:** The robot successfully picked up the yellow block and brought it to the wooden block, but released it off-center around 7.5 seconds, causing the block to slide off the wooden block and fail the stacking task.
diff --git a/blocks/ep000/labels.json b/blocks/ep000/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..2451f470c8db7e2234abd3e79902a70feafb75f8
--- /dev/null
+++ b/blocks/ep000/labels.json
@@ -0,0 +1,142 @@
+{
+  "rollout_id": "t000__ep000__blocks",
+  "scene_variant": "",
+  "task_id": 0,
+  "task_file": "",
+  "instruction": "Stack the yellow block on the wooden block",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T19:54:07+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.005893909626719,
+      "control_step_start": 0,
+      "control_step_end": 152,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "approaching yellow block",
+      "description": "approaching yellow block"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.005893909626719,
+      "t_end_sec": 6.011787819253438,
+      "control_step_start": 153,
+      "control_step_end": 305,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "grasping and lifting yellow block",
+      "description": "grasping and lifting yellow block"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.011787819253438,
+      "t_end_sec": 9.017681728880158,
+      "control_step_start": 306,
+      "control_step_end": 458,
+      "policy_step_start": 51,
+      "policy_step_end": 76,
+      "phrase": "releasing yellow block onto target",
+      "description": "releasing yellow block onto target"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.017681728880158,
+      "t_end_sec": 12.023575638506877,
+      "control_step_start": 459,
+      "control_step_end": 611,
+      "policy_step_start": 76,
+      "policy_step_end": 101,
+      "phrase": "hovering gripper over yellow block",
+      "description": "hovering gripper over yellow block"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 12.023575638506877,
+      "t_end_sec": 15.029469548133596,
+      "control_step_start": 612,
+      "control_step_end": 764,
+      "policy_step_start": 102,
+      "policy_step_end": 127,
+      "phrase": "pressing down on yellow block",
+      "description": "pressing down on yellow block"
+    },
+    {
+      "segment_index": 5,
+      "t_start_sec": 15.029469548133596,
+      "t_end_sec": 18.035363457760315,
+      "control_step_start": 765,
+      "control_step_end": 917,
+      "policy_step_start": 127,
+      "policy_step_end": 152,
+      "phrase": "holding position near yellow block",
+      "description": "holding position near yellow block"
+    },
+    {
+      "segment_index": 6,
+      "t_start_sec": 18.035363457760315,
+      "t_end_sec": 18.25147347740668,
+      "control_step_start": 918,
+      "control_step_end": 928,
+      "policy_step_start": 153,
+      "policy_step_end": 154,
+      "phrase": "remaining static near yellow block",
+      "description": "remaining static near yellow block"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "placement_or_insertion_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 4.675834970530452,
+    "vlm_failure_onset_timestamp": "00:04.7",
+    "vlm_failure_onset_frame": 238,
+    "vlm_failure_onset_step": 238,
+    "vlm_coarse_failure_onset_seconds": 7.5,
+    "vlm_failure_window_start_seconds": 6.8,
+    "vlm_failure_window_end_seconds": 8.8,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot brought the yellow block over to the wooden block but released it near the edge without proper alignment, causing the yellow block to slip off onto the tabletop.",
+    "vlm_failure_reason_summary": "misaligned placement; yellow block slipped off",
+    "vlm_recovery_action": "Re-grasp the yellow block from the table surface, align it directly over the center of the wooden block, and gently set it down.",
+    "vlm_recovery_action_summary": "re-grasp yellow block; stack on wooden block",
+    "vlm_prevention_action": "Position the yellow block directly over the center of the wooden block and verify alignment before opening the gripper.",
+    "vlm_prevention_action_summary": "center yellow block over target; release gently",
+    "vlm_justification": "The robot successfully picked up the yellow block and brought it to the wooden block, but released it off-center around 7.5 seconds, causing the block to slide off the wooden block and fail the stacking task.",
+    "vlm_temporal_justification": "At second 9, the robot opens its gripper to release the yellow block onto the wooden block, but because it is misaligned over the edge, the block immediately slips off onto the tabletop.",
+    "vlm_video_fps": 50.9,
+    "vlm_video_num_frames": 929,
+    "vlm_video_duration_seconds": 18.231827111984284,
+    "vlm_usage": {
+      "input_tokens": 22350,
+      "output_tokens": 364,
+      "cached_tokens": 1225
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"placement_or_insertion_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 7.5,\n \"failure_window_start_seconds\": 6.8,\n \"failure_window_end_seconds\": 8.8,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot brought the yellow block over to the wooden block but released it near the edge without proper alignment, causing the yellow block to slip off onto the tabletop.\",\n \"failure_reason_summary\": \"misaligned placement; yellow block slipped off\",\n \"recovery_action\": \"Re-grasp the yellow block from the table surface, align it directly over the center of the wooden block, and gently set it down.\",\n \"recovery_action_summary\": \"re-grasp yellow block; stack on wooden block\",\n \"prevention_action\": \"Position the yellow block directly over the center of the wooden block and verify alignment before opening the gripper.\",\n \"prevention_action_summary\": \"center yellow block over target; release gently\",\n \"justification\": \"The robot successfully picked up the yellow block and brought it to the wooden block, but released it off-center around 7.5 seconds, causing the block to slide off the wooden block and fail the stacking task.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 238,
+    "failure_sim_step": 238,
+    "failure_policy_step": 39,
+    "failure_chunk_index": 4,
+    "first_post_failure_policy_step": 40,
+    "failure_type": "placement_or_insertion_failure",
+    "correction_action": "Re-grasp the yellow block from the table surface, align it directly over the center of the wooden block, and gently set it down."
+  }
+}
\ No newline at end of file
diff --git a/blocks/ep000/labels.npz b/blocks/ep000/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..79ecc3bee6566ca0401edf70f94ca8304d948483
--- /dev/null
+++ b/blocks/ep000/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c1aba747d3bca4c5985c911852a35cecd9dc8476ed0453d9b3abc07221ee4828
+size 3965
diff --git a/blocks/ep001/example.md b/blocks/ep001/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..7822b95ad31f8dbdedcb311b7de620ece8f629fc
--- /dev/null
+++ b/blocks/ep001/example.md
@@ -0,0 +1,36 @@
+# t000__ep001__blocks
+
+- **suite / variant:** `` / ``
+- **instruction:** Stack the yellow block on the wooden block
+- **success:** False
+- **control steps:** 790  (15.5s at 51 Hz)
+- **policy inferences:** 136  (replan_steps=6)
+- **video:** `blocks/ep001/rollout.mp4`
+- **features:** `136` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–152 | 0–25 | approaching yellow block |
+| 3.0–6.0 | 153–305 | 25–50 | grasping and lifting yellow block |
+| 6.0–9.0 | 306–458 | 51–76 | lowering yellow block near target |
+| 9.0–12.0 | 459–611 | 76–101 | twisting arm with yellow block |
+| 12.0–15.0 | 612–764 | 102–127 | holding block in tilted pose |
+| 15.0–15.5 | 765–789 | 127–131 | remaining frozen near wooden block |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 5.893909626719057 s (00:05.9)
+- **frame / control_step:** 300 (policy_step=50, chunk_index=0)
+- **coarse onset:** 7.0 s (refined=True)
+- **mode:** placement_or_insertion_failure
+- **reason:** The robot brought the yellow block to the wooden block but rotated its wrist awkwardly, failing to stack the yellow block squarely on top of the wooden block.
+  Summary: awkward wrist pitch; failed to stack block
+- **recovery:** Re-grip the yellow block, lift it clear of the wooden block, realign it vertically, and place it onto the wooden block.
+  Summary: re-grip yellow block; align and place on wooden block
+- **prevention:** Maintain a vertical end-effector orientation while lowering the yellow block directly onto the center of the wooden block.
+  Summary: keep vertical orientation; lower directly onto wooden block
+- **justification:** At 00:07, as the robot reaches the wooden block, the wrist pitches forward and sideways awkwardly instead of maintaining a stable top-down pose, causing the yellow block to miss the top of the wooden block and fail the stacking task.
diff --git a/blocks/ep001/labels.json b/blocks/ep001/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..e889d31c23295c808e8760f98738b89b1f81d111
--- /dev/null
+++ b/blocks/ep001/labels.json
@@ -0,0 +1,131 @@
+{
+  "rollout_id": "t000__ep001__blocks",
+  "scene_variant": "",
+  "task_id": 0,
+  "task_file": "",
+  "instruction": "Stack the yellow block on the wooden block",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T19:54:40+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.005893909626719,
+      "control_step_start": 0,
+      "control_step_end": 152,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "approaching yellow block",
+      "description": "approaching yellow block"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.005893909626719,
+      "t_end_sec": 6.011787819253438,
+      "control_step_start": 153,
+      "control_step_end": 305,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "grasping and lifting yellow block",
+      "description": "grasping and lifting yellow block"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.011787819253438,
+      "t_end_sec": 9.017681728880158,
+      "control_step_start": 306,
+      "control_step_end": 458,
+      "policy_step_start": 51,
+      "policy_step_end": 76,
+      "phrase": "lowering yellow block near target",
+      "description": "lowering yellow block near target"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.017681728880158,
+      "t_end_sec": 12.023575638506877,
+      "control_step_start": 459,
+      "control_step_end": 611,
+      "policy_step_start": 76,
+      "policy_step_end": 101,
+      "phrase": "twisting arm with yellow block",
+      "description": "twisting arm with yellow block"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 12.023575638506877,
+      "t_end_sec": 15.029469548133596,
+      "control_step_start": 612,
+      "control_step_end": 764,
+      "policy_step_start": 102,
+      "policy_step_end": 127,
+      "phrase": "holding block in tilted pose",
+      "description": "holding block in tilted pose"
+    },
+    {
+      "segment_index": 5,
+      "t_start_sec": 15.029469548133596,
+      "t_end_sec": 15.520628683693516,
+      "control_step_start": 765,
+      "control_step_end": 789,
+      "policy_step_start": 127,
+      "policy_step_end": 131,
+      "phrase": "remaining frozen near wooden block",
+      "description": "remaining frozen near wooden block"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "placement_or_insertion_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 5.893909626719057,
+    "vlm_failure_onset_timestamp": "00:05.9",
+    "vlm_failure_onset_frame": 300,
+    "vlm_failure_onset_step": 300,
+    "vlm_coarse_failure_onset_seconds": 7.0,
+    "vlm_failure_window_start_seconds": 6.5,
+    "vlm_failure_window_end_seconds": 8.0,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot brought the yellow block to the wooden block but rotated its wrist awkwardly, failing to stack the yellow block squarely on top of the wooden block.",
+    "vlm_failure_reason_summary": "awkward wrist pitch; failed to stack block",
+    "vlm_recovery_action": "Re-grip the yellow block, lift it clear of the wooden block, realign it vertically, and place it onto the wooden block.",
+    "vlm_recovery_action_summary": "re-grip yellow block; align and place on wooden block",
+    "vlm_prevention_action": "Maintain a vertical end-effector orientation while lowering the yellow block directly onto the center of the wooden block.",
+    "vlm_prevention_action_summary": "keep vertical orientation; lower directly onto wooden block",
+    "vlm_justification": "At 00:07, as the robot reaches the wooden block, the wrist pitches forward and sideways awkwardly instead of maintaining a stable top-down pose, causing the yellow block to miss the top of the wooden block and fail the stacking task.",
+    "vlm_temporal_justification": "At timestamp 01:37, as the robot arm descends to stack the yellow block on the wooden block, it rotates its wrist into an awkward, misaligned angle, preventing it from placing the yellow block squarely on top of the wooden block.",
+    "vlm_video_fps": 50.9,
+    "vlm_video_num_frames": 790,
+    "vlm_video_duration_seconds": 15.50098231827112,
+    "vlm_usage": {
+      "input_tokens": 22222,
+      "output_tokens": 387,
+      "cached_tokens": 1089
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"placement_or_insertion_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 7.0,\n \"failure_window_start_seconds\": 6.5,\n \"failure_window_end_seconds\": 8.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot brought the yellow block to the wooden block but rotated its wrist awkwardly, failing to stack the yellow block squarely on top of the wooden block.\",\n \"failure_reason_summary\": \"awkward wrist pitch; failed to stack block\",\n \"recovery_action\": \"Re-grip the yellow block, lift it clear of the wooden block, realign it vertically, and place it onto the wooden block.\",\n \"recovery_action_summary\": \"re-grip yellow block; align and place on wooden block\",\n \"prevention_action\": \"Maintain a vertical end-effector orientation while lowering the yellow block directly onto the center of the wooden block.\",\n \"prevention_action_summary\": \"keep vertical orientation; lower directly onto wooden block\",\n \"justification\": \"At 00:07, as the robot reaches the wooden block, the wrist pitches forward and sideways awkwardly instead of maintaining a stable top-down pose, causing the yellow block to miss the top of the wooden block and fail the stacking task.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 300,
+    "failure_sim_step": 300,
+    "failure_policy_step": 50,
+    "failure_chunk_index": 0,
+    "first_post_failure_policy_step": 51,
+    "failure_type": "placement_or_insertion_failure",
+    "correction_action": "Re-grip the yellow block, lift it clear of the wooden block, realign it vertically, and place it onto the wooden block."
+  }
+}
\ No newline at end of file
diff --git a/blocks/ep001/labels.npz b/blocks/ep001/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..d0bb7e2188b7b423c20ef6d5b4c351d540dbefee
--- /dev/null
+++ b/blocks/ep001/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:be43ebbf9f2ccbdbf5c5cb042188cf136f8c828467acff141ca6e555049bb938
+size 3929
diff --git a/blocks/ep002/example.md b/blocks/ep002/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..9993b3598c82f1df8829f1a5275d018254139841
--- /dev/null
+++ b/blocks/ep002/example.md
@@ -0,0 +1,33 @@
+# t000__ep002__blocks
+
+- **suite / variant:** `` / ``
+- **instruction:** Stack the yellow block on the wooden block
+- **success:** False
+- **control steps:** 316  (6.3s at 50 Hz)
+- **policy inferences:** 55  (replan_steps=6)
+- **video:** `blocks/ep002/rollout.mp4`
+- **features:** `55` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | reaching down toward yellow block |
+| 3.0–6.0 | 151–301 | 25–50 | closing gripper and lifting empty |
+| 6.0–6.3 | 302–315 | 50–52 | hovering above blocks |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 4.183266932270916 s (00:04.2)
+- **frame / control_step:** 210 (policy_step=35, chunk_index=0)
+- **coarse onset:** 3.2 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot aligned with the yellow block and closed its end-effector, but failed to establish a secure grip, lifting up empty while leaving the yellow block on the table.
+  Summary: failed grasp on yellow block; lifting empty gripper
+- **recovery:** Lower the gripper back down, align the fingers squarely around the yellow block, close securely, and lift to proceed with stacking.
+  Summary: re-align gripper; grasp yellow block securely
+- **prevention:** Ensure the end-effector descends to the correct height and centers over the yellow block before closing the gripper to guarantee a firm hold.
+  Summary: align gripper centered on yellow block; secure firm grip
+- **justification:** Around 3.2 seconds into the rollout, the gripper attempts to grasp the yellow block but fails to pinch it properly, lifting back into the air without the object by 4.0 seconds.
diff --git a/blocks/ep002/labels.json b/blocks/ep002/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..576655be3261a1a1f9e91b2e49261616477d136c
--- /dev/null
+++ b/blocks/ep002/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t000__ep002__blocks",
+  "scene_variant": "",
+  "task_id": 0,
+  "task_file": "",
+  "instruction": "Stack the yellow block on the wooden block",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T19:55:11+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00796812749004,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching down toward yellow block",
+      "description": "reaching down toward yellow block"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00796812749004,
+      "t_end_sec": 6.01593625498008,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "closing gripper and lifting empty",
+      "description": "closing gripper and lifting empty"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01593625498008,
+      "t_end_sec": 6.294820717131474,
+      "control_step_start": 302,
+      "control_step_end": 315,
+      "policy_step_start": 50,
+      "policy_step_end": 52,
+      "phrase": "hovering above blocks",
+      "description": "hovering above blocks"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 4.183266932270916,
+    "vlm_failure_onset_timestamp": "00:04.2",
+    "vlm_failure_onset_frame": 210,
+    "vlm_failure_onset_step": 210,
+    "vlm_coarse_failure_onset_seconds": 3.2,
+    "vlm_failure_window_start_seconds": 3.0,
+    "vlm_failure_window_end_seconds": 4.0,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot aligned with the yellow block and closed its end-effector, but failed to establish a secure grip, lifting up empty while leaving the yellow block on the table.",
+    "vlm_failure_reason_summary": "failed grasp on yellow block; lifting empty gripper",
+    "vlm_recovery_action": "Lower the gripper back down, align the fingers squarely around the yellow block, close securely, and lift to proceed with stacking.",
+    "vlm_recovery_action_summary": "re-align gripper; grasp yellow block securely",
+    "vlm_prevention_action": "Ensure the end-effector descends to the correct height and centers over the yellow block before closing the gripper to guarantee a firm hold.",
+    "vlm_prevention_action_summary": "align gripper centered on yellow block; secure firm grip",
+    "vlm_justification": "Around 3.2 seconds into the rollout, the gripper attempts to grasp the yellow block but fails to pinch it properly, lifting back into the air without the object by 4.0 seconds.",
+    "vlm_temporal_justification": "At clip timestamp 03:20 (second 200), the robot's gripper fingers close completely without grasping the yellow block, failing to establish a grip as the block remains untouched on the table.",
+    "vlm_video_fps": 50.2,
+    "vlm_video_num_frames": 316,
+    "vlm_video_duration_seconds": 6.274900398406374,
+    "vlm_usage": {
+      "input_tokens": 21334,
+      "output_tokens": 370,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 3.2,\n \"failure_window_start_seconds\": 3.0,\n \"failure_window_end_seconds\": 4.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot aligned with the yellow block and closed its end-effector, but failed to establish a secure grip, lifting up empty while leaving the yellow block on the table.\",\n \"failure_reason_summary\": \"failed grasp on yellow block; lifting empty gripper\",\n \"recovery_action\": \"Lower the gripper back down, align the fingers squarely around the yellow block, close securely, and lift to proceed with stacking.\",\n \"recovery_action_summary\": \"re-align gripper; grasp yellow block securely\",\n \"prevention_action\": \"Ensure the end-effector descends to the correct height and centers over the yellow block before closing the gripper to guarantee a firm hold.\",\n \"prevention_action_summary\": \"align gripper centered on yellow block; secure firm grip\",\n \"justification\": \"Around 3.2 seconds into the rollout, the gripper attempts to grasp the yellow block but fails to pinch it properly, lifting back into the air without the object by 4.0 seconds.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 210,
+    "failure_sim_step": 210,
+    "failure_policy_step": 35,
+    "failure_chunk_index": 0,
+    "first_post_failure_policy_step": 36,
+    "failure_type": "grasp_failure",
+    "correction_action": "Lower the gripper back down, align the fingers squarely around the yellow block, close securely, and lift to proceed with stacking."
+  }
+}
\ No newline at end of file
diff --git a/blocks/ep002/labels.npz b/blocks/ep002/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..274e690071ab0efd14f907f2c93f8ccfa8d1609b
--- /dev/null
+++ b/blocks/ep002/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2acc8af12a2f98c6dc2cebcec34daea0a4669e11c279f8cf41a3804291203272
+size 3859
diff --git a/blocks/ep003/example.md b/blocks/ep003/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..68ed3e3aeb55fbce15c2f0b1f5ca9c9231f1531f
--- /dev/null
+++ b/blocks/ep003/example.md
@@ -0,0 +1,43 @@
+# t000__ep003__blocks
+
+- **suite / variant:** `` / ``
+- **instruction:** Stack the yellow block on the wooden block
+- **success:** False
+- **control steps:** 1955  (38.8s at 50 Hz)
+- **policy inferences:** 337  (replan_steps=6)
+- **video:** `blocks/ep003/rollout.mp4`
+- **features:** `337` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | reaching toward yellow block |
+| 3.0–6.0 | 151–301 | 25–50 | grasping yellow block |
+| 6.0–9.0 | 302–452 | 50–75 | lifting yellow block |
+| 9.0–12.0 | 453–603 | 75–100 | moving yellow block near target |
+| 12.0–15.0 | 604–754 | 100–125 | lowering yellow block onto target |
+| 15.0–18.0 | 755–905 | 125–150 | placing yellow block on wooden block |
+| 18.0–21.0 | 906–1056 | 151–176 | resting gripper on tilted block |
+| 21.0–24.0 | 1057–1207 | 176–201 | remaining stationary on block |
+| 24.0–27.0 | 1208–1358 | 201–226 | idle with end-effector over block |
+| 27.0–30.0 | 1359–1509 | 226–251 | holding position near block |
+| 30.0–33.0 | 1510–1660 | 251–276 | motionless gripper over wooden block |
+| 33.0–36.0 | 1661–1811 | 276–301 | remaining stationary near block |
+| 36.0–38.8 | 1812–1954 | 302–325 | idle / no progress |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 15.0 s (00:15.0)
+- **frame / control_step:** 756 (policy_step=126, chunk_index=0)
+- **coarse onset:** 18.0 s (refined=True)
+- **mode:** placement_or_insertion_failure
+- **reason:** The robot brought the yellow block to the wooden block but misaligned it during placement, leaving the yellow block tilted sideways on top of the wooden block with the gripper stuck against it.
+  Summary: misaligned block placement; block left tilted on wooden block
+- **recovery:** Re-grasp the yellow block, raise it above the wooden block, align its orientation horizontally, and place it squarely on top of the wooden block.
+  Summary: re-grasp yellow block; align and stack on wooden block
+- **prevention:** Center the yellow block directly over the top face of the wooden block in a level orientation before lowering and releasing the gripper.
+  Summary: align yellow block centered over wooden block; release smoothly
+- **justification:** Around 18.0 seconds, the robot attempts to place the yellow block on the wooden block, but releases it at a skewed angle, leaving it tilted unsteadily on top of the wooden block while the gripper remains wedged over it.
diff --git a/blocks/ep003/labels.json b/blocks/ep003/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..59ff9de400be01013dc4c3379a64bc40c69a01d8
--- /dev/null
+++ b/blocks/ep003/labels.json
@@ -0,0 +1,208 @@
+{
+  "rollout_id": "t000__ep003__blocks",
+  "scene_variant": "",
+  "task_id": 0,
+  "task_file": "",
+  "instruction": "Stack the yellow block on the wooden block",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T19:55:52+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.996031746031746,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching toward yellow block",
+      "description": "reaching toward yellow block"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.996031746031746,
+      "t_end_sec": 5.992063492063492,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "grasping yellow block",
+      "description": "grasping yellow block"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.992063492063492,
+      "t_end_sec": 8.988095238095239,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "lifting yellow block",
+      "description": "lifting yellow block"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.988095238095239,
+      "t_end_sec": 11.984126984126984,
+      "control_step_start": 453,
+      "control_step_end": 603,
+      "policy_step_start": 75,
+      "policy_step_end": 100,
+      "phrase": "moving yellow block near target",
+      "description": "moving yellow block near target"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 11.984126984126984,
+      "t_end_sec": 14.980158730158731,
+      "control_step_start": 604,
+      "control_step_end": 754,
+      "policy_step_start": 100,
+      "policy_step_end": 125,
+      "phrase": "lowering yellow block onto target",
+      "description": "lowering yellow block onto target"
+    },
+    {
+      "segment_index": 5,
+      "t_start_sec": 14.980158730158731,
+      "t_end_sec": 17.976190476190478,
+      "control_step_start": 755,
+      "control_step_end": 905,
+      "policy_step_start": 125,
+      "policy_step_end": 150,
+      "phrase": "placing yellow block on wooden block",
+      "description": "placing yellow block on wooden block"
+    },
+    {
+      "segment_index": 6,
+      "t_start_sec": 17.976190476190478,
+      "t_end_sec": 20.97222222222222,
+      "control_step_start": 906,
+      "control_step_end": 1056,
+      "policy_step_start": 151,
+      "policy_step_end": 176,
+      "phrase": "resting gripper on tilted block",
+      "description": "resting gripper on tilted block"
+    },
+    {
+      "segment_index": 7,
+      "t_start_sec": 20.97222222222222,
+      "t_end_sec": 23.96825396825397,
+      "control_step_start": 1057,
+      "control_step_end": 1207,
+      "policy_step_start": 176,
+      "policy_step_end": 201,
+      "phrase": "remaining stationary on block",
+      "description": "remaining stationary on block"
+    },
+    {
+      "segment_index": 8,
+      "t_start_sec": 23.96825396825397,
+      "t_end_sec": 26.964285714285715,
+      "control_step_start": 1208,
+      "control_step_end": 1358,
+      "policy_step_start": 201,
+      "policy_step_end": 226,
+      "phrase": "idle with end-effector over block",
+      "description": "idle with end-effector over block"
+    },
+    {
+      "segment_index": 9,
+      "t_start_sec": 26.964285714285715,
+      "t_end_sec": 29.960317460317462,
+      "control_step_start": 1359,
+      "control_step_end": 1509,
+      "policy_step_start": 226,
+      "policy_step_end": 251,
+      "phrase": "holding position near block",
+      "description": "holding position near block"
+    },
+    {
+      "segment_index": 10,
+      "t_start_sec": 29.960317460317462,
+      "t_end_sec": 32.95634920634921,
+      "control_step_start": 1510,
+      "control_step_end": 1660,
+      "policy_step_start": 251,
+      "policy_step_end": 276,
+      "phrase": "motionless gripper over wooden block",
+      "description": "motionless gripper over wooden block"
+    },
+    {
+      "segment_index": 11,
+      "t_start_sec": 32.95634920634921,
+      "t_end_sec": 35.952380952380956,
+      "control_step_start": 1661,
+      "control_step_end": 1811,
+      "policy_step_start": 276,
+      "policy_step_end": 301,
+      "phrase": "remaining stationary near block",
+      "description": "remaining stationary near block"
+    },
+    {
+      "segment_index": 12,
+      "t_start_sec": 35.952380952380956,
+      "t_end_sec": 38.78968253968254,
+      "control_step_start": 1812,
+      "control_step_end": 1954,
+      "policy_step_start": 302,
+      "policy_step_end": 325,
+      "phrase": "idle / no progress",
+      "description": "idle / no progress"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "placement_or_insertion_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 15.0,
+    "vlm_failure_onset_timestamp": "00:15.0",
+    "vlm_failure_onset_frame": 756,
+    "vlm_failure_onset_step": 756,
+    "vlm_coarse_failure_onset_seconds": 18.0,
+    "vlm_failure_window_start_seconds": 17.0,
+    "vlm_failure_window_end_seconds": 19.0,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot brought the yellow block to the wooden block but misaligned it during placement, leaving the yellow block tilted sideways on top of the wooden block with the gripper stuck against it.",
+    "vlm_failure_reason_summary": "misaligned block placement; block left tilted on wooden block",
+    "vlm_recovery_action": "Re-grasp the yellow block, raise it above the wooden block, align its orientation horizontally, and place it squarely on top of the wooden block.",
+    "vlm_recovery_action_summary": "re-grasp yellow block; align and stack on wooden block",
+    "vlm_prevention_action": "Center the yellow block directly over the top face of the wooden block in a level orientation before lowering and releasing the gripper.",
+    "vlm_prevention_action_summary": "align yellow block centered over wooden block; release smoothly",
+    "vlm_justification": "Around 18.0 seconds, the robot attempts to place the yellow block on the wooden block, but releases it at a skewed angle, leaving it tilted unsteadily on top of the wooden block while the gripper remains wedged over it.",
+    "vlm_temporal_justification": "At the start of the clip (00:00), the yellow block is already misaligned and resting tilted on top of the wooden block under the robot gripper.",
+    "vlm_video_fps": 50.4,
+    "vlm_video_num_frames": 1955,
+    "vlm_video_duration_seconds": 38.76984126984127,
+    "vlm_usage": {
+      "input_tokens": 23426,
+      "output_tokens": 385,
+      "cached_tokens": 2653
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"placement_or_insertion_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 18.0,\n \"failure_window_start_seconds\": 17.0,\n \"failure_window_end_seconds\": 19.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot brought the yellow block to the wooden block but misaligned it during placement, leaving the yellow block tilted sideways on top of the wooden block with the gripper stuck against it.\",\n \"failure_reason_summary\": \"misaligned block placement; block left tilted on wooden block\",\n \"recovery_action\": \"Re-grasp the yellow block, raise it above the wooden block, align its orientation horizontally, and place it squarely on top of the wooden block.\",\n \"recovery_action_summary\": \"re-grasp yellow block; align and stack on wooden block\",\n \"prevention_action\": \"Center the yellow block directly over the top face of the wooden block in a level orientation before lowering and releasing the gripper.\",\n \"prevention_action_summary\": \"align yellow block centered over wooden block; release smoothly\",\n \"justification\": \"Around 18.0 seconds, the robot attempts to place the yellow block on the wooden block, but releases it at a skewed angle, leaving it tilted unsteadily on top of the wooden block while the gripper remains wedged over it.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 756,
+    "failure_sim_step": 756,
+    "failure_policy_step": 126,
+    "failure_chunk_index": 0,
+    "first_post_failure_policy_step": 127,
+    "failure_type": "placement_or_insertion_failure",
+    "correction_action": "Re-grasp the yellow block, raise it above the wooden block, align its orientation horizontally, and place it squarely on top of the wooden block."
+  }
+}
\ No newline at end of file
diff --git a/blocks/ep003/labels.npz b/blocks/ep003/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..6fd07e1c7b6b8137d71b619f41a163e20259d184
--- /dev/null
+++ b/blocks/ep003/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:46a01c736361732dd92188354d149256a6f11d6ab6090963961b6925d1774394
+size 4272
diff --git a/blocks/ep004/example.md b/blocks/ep004/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..cdd1828540b94b1710e9b78fb62967e1ddfc8eaf
--- /dev/null
+++ b/blocks/ep004/example.md
@@ -0,0 +1,33 @@
+# t000__ep004__blocks
+
+- **suite / variant:** `` / ``
+- **instruction:** Stack the yellow block on the wooden block
+- **success:** False
+- **control steps:** 349  (6.9s at 51 Hz)
+- **policy inferences:** 60  (replan_steps=6)
+- **video:** `blocks/ep004/rollout.mp4`
+- **features:** `60` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | descending toward yellow block |
+| 3.0–6.0 | 152–303 | 25–50 | attempting grasp and lifting empty |
+| 6.0–6.9 | 304–348 | 50–58 | hovering empty over wooden block |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 4.251968503937008 s (00:04.3)
+- **frame / control_step:** 216 (policy_step=36, chunk_index=0)
+- **coarse onset:** 3.8 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the yellow block, but the gripper closed improperly without securing the block, lifting up empty.
+  Summary: failed grasp on yellow block; empty gripper lift
+- **recovery:** Re-align the gripper over the yellow block, descend, securely grasp the yellow block, and then move to stack it on the wooden block.
+  Summary: re-align gripper over yellow block; grasp yellow block
+- **prevention:** Ensure precise alignment and sufficient finger contact around the yellow block before closing the gripper and lifting.
+  Summary: align gripper precisely; secure grasp before lifting
+- **justification:** At around 3.8 seconds, the robot gripper closes around the yellow block but fails to grasp it securely. By 5.0 seconds, the arm lifts up completely empty, leaving the yellow block on the table while proceeding to move towards the wooden block without the object.
diff --git a/blocks/ep004/labels.json b/blocks/ep004/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..6606744cbd1d5fa1c482bea0dad41afbbbf68943
--- /dev/null
+++ b/blocks/ep004/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t000__ep004__blocks",
+  "scene_variant": "",
+  "task_id": 0,
+  "task_file": "",
+  "instruction": "Stack the yellow block on the wooden block",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T19:56:31+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9921259842519685,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "descending toward yellow block",
+      "description": "descending toward yellow block"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9921259842519685,
+      "t_end_sec": 5.984251968503937,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "attempting grasp and lifting empty",
+      "description": "attempting grasp and lifting empty"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.984251968503937,
+      "t_end_sec": 6.8700787401574805,
+      "control_step_start": 304,
+      "control_step_end": 348,
+      "policy_step_start": 50,
+      "policy_step_end": 58,
+      "phrase": "hovering empty over wooden block",
+      "description": "hovering empty over wooden block"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 4.251968503937008,
+    "vlm_failure_onset_timestamp": "00:04.3",
+    "vlm_failure_onset_frame": 216,
+    "vlm_failure_onset_step": 216,
+    "vlm_coarse_failure_onset_seconds": 3.8,
+    "vlm_failure_window_start_seconds": 3.3,
+    "vlm_failure_window_end_seconds": 4.3,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the yellow block, but the gripper closed improperly without securing the block, lifting up empty.",
+    "vlm_failure_reason_summary": "failed grasp on yellow block; empty gripper lift",
+    "vlm_recovery_action": "Re-align the gripper over the yellow block, descend, securely grasp the yellow block, and then move to stack it on the wooden block.",
+    "vlm_recovery_action_summary": "re-align gripper over yellow block; grasp yellow block",
+    "vlm_prevention_action": "Ensure precise alignment and sufficient finger contact around the yellow block before closing the gripper and lifting.",
+    "vlm_prevention_action_summary": "align gripper precisely; secure grasp before lifting",
+    "vlm_justification": "At around 3.8 seconds, the robot gripper closes around the yellow block but fails to grasp it securely. By 5.0 seconds, the arm lifts up completely empty, leaving the yellow block on the table while proceeding to move towards the wooden block without the object.",
+    "vlm_temporal_justification": "At 02:55 in the clip (175 seconds), the robot's gripper finishes closing completely shut without pinching or securing the yellow block, failing the grasp attempt before lifting up empty.",
+    "vlm_video_fps": 50.8,
+    "vlm_video_num_frames": 349,
+    "vlm_video_duration_seconds": 6.850393700787402,
+    "vlm_usage": {
+      "input_tokens": 21512,
+      "output_tokens": 369,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 3.8,\n \"failure_window_start_seconds\": 3.3,\n \"failure_window_end_seconds\": 4.3,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the yellow block, but the gripper closed improperly without securing the block, lifting up empty.\",\n \"failure_reason_summary\": \"failed grasp on yellow block; empty gripper lift\",\n \"recovery_action\": \"Re-align the gripper over the yellow block, descend, securely grasp the yellow block, and then move to stack it on the wooden block.\",\n \"recovery_action_summary\": \"re-align gripper over yellow block; grasp yellow block\",\n \"prevention_action\": \"Ensure precise alignment and sufficient finger contact around the yellow block before closing the gripper and lifting.\",\n \"prevention_action_summary\": \"align gripper precisely; secure grasp before lifting\",\n \"justification\": \"At around 3.8 seconds, the robot gripper closes around the yellow block but fails to grasp it securely. By 5.0 seconds, the arm lifts up completely empty, leaving the yellow block on the table while proceeding to move towards the wooden block without the object.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 216,
+    "failure_sim_step": 216,
+    "failure_policy_step": 36,
+    "failure_chunk_index": 0,
+    "first_post_failure_policy_step": 37,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-align the gripper over the yellow block, descend, securely grasp the yellow block, and then move to stack it on the wooden block."
+  }
+}
\ No newline at end of file
diff --git a/blocks/ep004/labels.npz b/blocks/ep004/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..72970b74fec606869729f80e1a18bf3a1f9609e9
--- /dev/null
+++ b/blocks/ep004/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:29a3f9983a8c80dcbdf3b326ccd4c25d93fe2d01b70b4536e81ec4689f803ac3
+size 3761
diff --git a/blocks/ep005/example.md b/blocks/ep005/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..2ddb2e0f14ba5915d105602e87a2681a06a16449
--- /dev/null
+++ b/blocks/ep005/example.md
@@ -0,0 +1,35 @@
+# t000__ep005__blocks
+
+- **suite / variant:** `` / ``
+- **instruction:** Stack the yellow block on the wooden block
+- **success:** False
+- **control steps:** 622  (12.4s at 50 Hz)
+- **policy inferences:** 108  (replan_steps=6)
+- **video:** `blocks/ep005/rollout.mp4`
+- **features:** `108` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | approaching yellow block |
+| 3.0–6.0 | 151–301 | 25–50 | grasping and lifting yellow block |
+| 6.0–9.0 | 302–452 | 50–75 | carrying yellow block to target |
+| 9.0–12.0 | 453–603 | 75–100 | releasing yellow block off target |
+| 12.0–12.4 | 604–621 | 100–103 | retracting arm and remaining idle |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 10.735586481113321 s (00:10.7)
+- **frame / control_step:** 540 (policy_step=90, chunk_index=0)
+- **coarse onset:** 10.0 s (refined=True)
+- **mode:** placement_or_insertion_failure
+- **reason:** The robot successfully picked up the yellow block and brought it over the target area, but released it beside the wooden block on the tabletop rather than stacking it on top.
+  Summary: misaligned placement; releasing object off-target
+- **recovery:** Re-grasp the yellow block from the table surface, center it directly over the wooden block, and stack it on top.
+  Summary: re-grasp yellow block; stack on wooden block
+- **prevention:** Ensure the yellow block is centered directly above the top surface of the wooden block before opening the gripper.
+  Summary: center yellow block over target; release carefully
+- **justification:** The robot successfully picked up the yellow block and moved toward the target, but at around 10.0 seconds it released the block beside the wooden block onto the tabletop instead of placing it on top.
diff --git a/blocks/ep005/labels.json b/blocks/ep005/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..84d4c5ad7fa84bad046a0171760501fe27d421b1
--- /dev/null
+++ b/blocks/ep005/labels.json
@@ -0,0 +1,120 @@
+{
+  "rollout_id": "t000__ep005__blocks",
+  "scene_variant": "",
+  "task_id": 0,
+  "task_file": "",
+  "instruction": "Stack the yellow block on the wooden block",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T19:57:01+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.001988071570577,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "approaching yellow block",
+      "description": "approaching yellow block"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.001988071570577,
+      "t_end_sec": 6.003976143141154,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "grasping and lifting yellow block",
+      "description": "grasping and lifting yellow block"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.003976143141154,
+      "t_end_sec": 9.00596421471173,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "carrying yellow block to target",
+      "description": "carrying yellow block to target"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.00596421471173,
+      "t_end_sec": 12.007952286282308,
+      "control_step_start": 453,
+      "control_step_end": 603,
+      "policy_step_start": 75,
+      "policy_step_end": 100,
+      "phrase": "releasing yellow block off target",
+      "description": "releasing yellow block off target"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 12.007952286282308,
+      "t_end_sec": 12.365805168986084,
+      "control_step_start": 604,
+      "control_step_end": 621,
+      "policy_step_start": 100,
+      "policy_step_end": 103,
+      "phrase": "retracting arm and remaining idle",
+      "description": "retracting arm and remaining idle"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "placement_or_insertion_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 10.735586481113321,
+    "vlm_failure_onset_timestamp": "00:10.7",
+    "vlm_failure_onset_frame": 540,
+    "vlm_failure_onset_step": 540,
+    "vlm_coarse_failure_onset_seconds": 10.0,
+    "vlm_failure_window_start_seconds": 9.5,
+    "vlm_failure_window_end_seconds": 10.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot successfully picked up the yellow block and brought it over the target area, but released it beside the wooden block on the tabletop rather than stacking it on top.",
+    "vlm_failure_reason_summary": "misaligned placement; releasing object off-target",
+    "vlm_recovery_action": "Re-grasp the yellow block from the table surface, center it directly over the wooden block, and stack it on top.",
+    "vlm_recovery_action_summary": "re-grasp yellow block; stack on wooden block",
+    "vlm_prevention_action": "Ensure the yellow block is centered directly above the top surface of the wooden block before opening the gripper.",
+    "vlm_prevention_action_summary": "center yellow block over target; release carefully",
+    "vlm_justification": "The robot successfully picked up the yellow block and moved toward the target, but at around 10.0 seconds it released the block beside the wooden block onto the tabletop instead of placing it on top.",
+    "vlm_temporal_justification": "At clip second 03:08 (frame 188 in the clip sequence), the robot gripper opens and releases the yellow block directly onto the tabletop surface beside the wooden block, failing to stack it on top.",
+    "vlm_video_fps": 50.3,
+    "vlm_video_num_frames": 622,
+    "vlm_video_duration_seconds": 12.345924453280318,
+    "vlm_usage": {
+      "input_tokens": 19642,
+      "output_tokens": 371,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"placement_or_insertion_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 10.0,\n \"failure_window_start_seconds\": 9.5,\n \"failure_window_end_seconds\": 10.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot successfully picked up the yellow block and brought it over the target area, but released it beside the wooden block on the tabletop rather than stacking it on top.\",\n \"failure_reason_summary\": \"misaligned placement; releasing object off-target\",\n \"recovery_action\": \"Re-grasp the yellow block from the table surface, center it directly over the wooden block, and stack it on top.\",\n \"recovery_action_summary\": \"re-grasp yellow block; stack on wooden block\",\n \"prevention_action\": \"Ensure the yellow block is centered directly above the top surface of the wooden block before opening the gripper.\",\n \"prevention_action_summary\": \"center yellow block over target; release carefully\",\n \"justification\": \"The robot successfully picked up the yellow block and moved toward the target, but at around 10.0 seconds it released the block beside the wooden block onto the tabletop instead of placing it on top.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 540,
+    "failure_sim_step": 540,
+    "failure_policy_step": 90,
+    "failure_chunk_index": 0,
+    "first_post_failure_policy_step": 91,
+    "failure_type": "placement_or_insertion_failure",
+    "correction_action": "Re-grasp the yellow block from the table surface, center it directly over the wooden block, and stack it on top."
+  }
+}
\ No newline at end of file
diff --git a/blocks/ep005/labels.npz b/blocks/ep005/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..2ae96d0cfa1395be57169ff8807ae6a86954554f
--- /dev/null
+++ b/blocks/ep005/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5afed9ad4df870db65e0eb4d8e8ab01e29b41a28875641314f7d49b29d606089
+size 3894
diff --git a/blocks/ep006/example.md b/blocks/ep006/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..ff593c0a77cbeff6c21430fa4dbb794749ab6599
--- /dev/null
+++ b/blocks/ep006/example.md
@@ -0,0 +1,23 @@
+# t000__ep006__blocks
+
+- **suite / variant:** `` / ``
+- **instruction:** Stack the yellow block on the wooden block
+- **success:** True
+- **control steps:** 760  (15.0s at 51 Hz)
+- **policy inferences:** 130  (replan_steps=6)
+- **video:** `blocks/ep006/rollout.mp4`
+- **features:** `130` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching for yellow block |
+| 3.0–6.0 | 152–303 | 25–50 | grasping yellow block |
+| 6.0–9.0 | 304–455 | 50–75 | stacking yellow block on wooden block |
+| 9.0–12.0 | 456–607 | 76–101 | releasing yellow block |
+| 12.0–15.0 | 608–759 | 101–126 | retracting arm to home position |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/blocks/ep006/labels.json b/blocks/ep006/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..fd1790611f382baf9371b9f21ffc79d4bc1d94a0
--- /dev/null
+++ b/blocks/ep006/labels.json
@@ -0,0 +1,90 @@
+{
+  "rollout_id": "t000__ep006__blocks",
+  "scene_variant": "",
+  "task_id": 0,
+  "task_file": "",
+  "instruction": "Stack the yellow block on the wooden block",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T19:57:08+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0039525691699605,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for yellow block",
+      "description": "reaching for yellow block"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0039525691699605,
+      "t_end_sec": 6.007905138339921,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "grasping yellow block",
+      "description": "grasping yellow block"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.007905138339921,
+      "t_end_sec": 9.011857707509881,
+      "control_step_start": 304,
+      "control_step_end": 455,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "stacking yellow block on wooden block",
+      "description": "stacking yellow block on wooden block"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.011857707509881,
+      "t_end_sec": 12.015810276679842,
+      "control_step_start": 456,
+      "control_step_end": 607,
+      "policy_step_start": 76,
+      "policy_step_end": 101,
+      "phrase": "releasing yellow block",
+      "description": "releasing yellow block"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 12.015810276679842,
+      "t_end_sec": 15.019762845849803,
+      "control_step_start": 608,
+      "control_step_end": 759,
+      "policy_step_start": 101,
+      "policy_step_end": 126,
+      "phrase": "retracting arm to home position",
+      "description": "retracting arm to home position"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/blocks/ep006/labels.npz b/blocks/ep006/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..32aaf4673fdfa0fb563bd8da7cb32ed55bcd7f9b
--- /dev/null
+++ b/blocks/ep006/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e26c94e053f2771256570f98dafd77de4f6b95cfc13f3a17d17ba3c1c579ef32
+size 3152
diff --git a/blocks/ep007/example.md b/blocks/ep007/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..6f3b0e53142b02902995fdc7aad2f4d36f0d80ac
--- /dev/null
+++ b/blocks/ep007/example.md
@@ -0,0 +1,33 @@
+# t000__ep007__blocks
+
+- **suite / variant:** `` / ``
+- **instruction:** Stack the yellow block on the wooden block
+- **success:** False
+- **control steps:** 342  (6.8s at 50 Hz)
+- **policy inferences:** 60  (replan_steps=6)
+- **video:** `blocks/ep007/rollout.mp4`
+- **features:** `60` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching for yellow block |
+| 3.0–6.0 | 152–303 | 25–50 | attempting grasp and lifting empty |
+| 6.0–6.8 | 304–341 | 50–56 | hovering empty gripper |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 3.5841584158415842 s (00:03.6)
+- **frame / control_step:** 181 (policy_step=30, chunk_index=1)
+- **coarse onset:** 3.5 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the yellow block, but the gripper closed off-center, causing the block to slip out and remain on the table.
+  Summary: failed grasp on yellow block; object slipped
+- **recovery:** Re-align the gripper directly over the yellow block, close the fingers firmly to secure the grasp, and lift it to stack on top of the wooden block.
+  Summary: re-align gripper; regrasp yellow block; stack on wooden block
+- **prevention:** Position the gripper accurately over the center of the yellow block before closing the fingers to establish a stable hold.
+  Summary: center gripper; secure hold on block
+- **justification:** Around 3.5 seconds into the rollout, the robot closes its gripper on the yellow block, but fails to establish a grasp. The block slips out as the arm rises, leaving the gripper empty for the remainder of the trajectory.
diff --git a/blocks/ep007/labels.json b/blocks/ep007/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..36c7915642027bd095d2ca8d11040311a74a5b05
--- /dev/null
+++ b/blocks/ep007/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t000__ep007__blocks",
+  "scene_variant": "",
+  "task_id": 0,
+  "task_file": "",
+  "instruction": "Stack the yellow block on the wooden block",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T19:57:43+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00990099009901,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for yellow block",
+      "description": "reaching for yellow block"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00990099009901,
+      "t_end_sec": 6.01980198019802,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "attempting grasp and lifting empty",
+      "description": "attempting grasp and lifting empty"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01980198019802,
+      "t_end_sec": 6.772277227722772,
+      "control_step_start": 304,
+      "control_step_end": 341,
+      "policy_step_start": 50,
+      "policy_step_end": 56,
+      "phrase": "hovering empty gripper",
+      "description": "hovering empty gripper"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 3.5841584158415842,
+    "vlm_failure_onset_timestamp": "00:03.6",
+    "vlm_failure_onset_frame": 181,
+    "vlm_failure_onset_step": 181,
+    "vlm_coarse_failure_onset_seconds": 3.5,
+    "vlm_failure_window_start_seconds": 3.0,
+    "vlm_failure_window_end_seconds": 4.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the yellow block, but the gripper closed off-center, causing the block to slip out and remain on the table.",
+    "vlm_failure_reason_summary": "failed grasp on yellow block; object slipped",
+    "vlm_recovery_action": "Re-align the gripper directly over the yellow block, close the fingers firmly to secure the grasp, and lift it to stack on top of the wooden block.",
+    "vlm_recovery_action_summary": "re-align gripper; regrasp yellow block; stack on wooden block",
+    "vlm_prevention_action": "Position the gripper accurately over the center of the yellow block before closing the fingers to establish a stable hold.",
+    "vlm_prevention_action_summary": "center gripper; secure hold on block",
+    "vlm_justification": "Around 3.5 seconds into the rollout, the robot closes its gripper on the yellow block, but fails to establish a grasp. The block slips out as the arm rises, leaving the gripper empty for the remainder of the trajectory.",
+    "vlm_temporal_justification": "At 02:36 in the clip, the robot gripper begins closing its fingers around the yellow block while positioned off-center. This results in a failed grasp, causing the block to slip out and remain on the table when the arm lifts.",
+    "vlm_video_fps": 50.5,
+    "vlm_video_num_frames": 342,
+    "vlm_video_duration_seconds": 6.752475247524752,
+    "vlm_usage": {
+      "input_tokens": 21518,
+      "output_tokens": 382,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 3.5,\n \"failure_window_start_seconds\": 3.0,\n \"failure_window_end_seconds\": 4.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the yellow block, but the gripper closed off-center, causing the block to slip out and remain on the table.\",\n \"failure_reason_summary\": \"failed grasp on yellow block; object slipped\",\n \"recovery_action\": \"Re-align the gripper directly over the yellow block, close the fingers firmly to secure the grasp, and lift it to stack on top of the wooden block.\",\n \"recovery_action_summary\": \"re-align gripper; regrasp yellow block; stack on wooden block\",\n \"prevention_action\": \"Position the gripper accurately over the center of the yellow block before closing the fingers to establish a stable hold.\",\n \"prevention_action_summary\": \"center gripper; secure hold on block\",\n \"justification\": \"Around 3.5 seconds into the rollout, the robot closes its gripper on the yellow block, but fails to establish a grasp. The block slips out as the arm rises, leaving the gripper empty for the remainder of the trajectory.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 181,
+    "failure_sim_step": 181,
+    "failure_policy_step": 30,
+    "failure_chunk_index": 1,
+    "first_post_failure_policy_step": 31,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-align the gripper directly over the yellow block, close the fingers firmly to secure the grasp, and lift it to stack on top of the wooden block."
+  }
+}
\ No newline at end of file
diff --git a/blocks/ep007/labels.npz b/blocks/ep007/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..5ec9901d0c1647e8e744f8db56fbece2a852fb0a
--- /dev/null
+++ b/blocks/ep007/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:481a03d6a8b17e5309941c8a4f9daa26320dd21577795d9e8ebf68a21ab50ca8
+size 3797
diff --git a/blocks/ep008/example.md b/blocks/ep008/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..bbd1b3b3e5d34dc6d61bd2531af7c6f6a40049b2
--- /dev/null
+++ b/blocks/ep008/example.md
@@ -0,0 +1,33 @@
+# t000__ep008__blocks
+
+- **suite / variant:** `` / ``
+- **instruction:** Stack the yellow block on the wooden block
+- **success:** False
+- **control steps:** 381  (7.4s at 52 Hz)
+- **policy inferences:** 66  (replan_steps=6)
+- **video:** `blocks/ep008/rollout.mp4`
+- **features:** `66` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–154 | 0–25 | approaching yellow block |
+| 3.0–6.0 | 155–309 | 25–51 | closing gripper and missing grasp |
+| 6.0–7.4 | 310–380 | 51–63 | retracting empty gripper |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 4.050387596899225 s (00:04.1)
+- **frame / control_step:** 209 (policy_step=34, chunk_index=5)
+- **coarse onset:** 4.0 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot closed its gripper on the yellow block, but failed to establish a secure grip, causing the block to slip out as the arm lifted.
+  Summary: improper grasp alignment; failed object pick-up
+- **recovery:** Re-align the gripper directly over the yellow block, close the fingers securely around it, lift it, and stack it onto the wooden block.
+  Summary: re-align gripper; regrasp yellow block; stack on wooden block
+- **prevention:** Ensure the gripper is properly centered and at the correct height over the yellow block before closing the fingers to guarantee a solid grasp.
+  Summary: center gripper on block; close gripper securely
+- **justification:** At 4.0 seconds, the robot attempts to close its gripper around the yellow block, but the block pops out of the fingers as the arm begins lifting, leaving the block on the table.
diff --git a/blocks/ep008/labels.json b/blocks/ep008/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..e377b120cdbe92aeddfe294a9ca92f7a2e02d3bd
--- /dev/null
+++ b/blocks/ep008/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t000__ep008__blocks",
+  "scene_variant": "",
+  "task_id": 0,
+  "task_file": "",
+  "instruction": "Stack the yellow block on the wooden block",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T19:58:20+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.003875968992248,
+      "control_step_start": 0,
+      "control_step_end": 154,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "approaching yellow block",
+      "description": "approaching yellow block"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.003875968992248,
+      "t_end_sec": 6.007751937984496,
+      "control_step_start": 155,
+      "control_step_end": 309,
+      "policy_step_start": 25,
+      "policy_step_end": 51,
+      "phrase": "closing gripper and missing grasp",
+      "description": "closing gripper and missing grasp"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.007751937984496,
+      "t_end_sec": 7.383720930232558,
+      "control_step_start": 310,
+      "control_step_end": 380,
+      "policy_step_start": 51,
+      "policy_step_end": 63,
+      "phrase": "retracting empty gripper",
+      "description": "retracting empty gripper"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 4.050387596899225,
+    "vlm_failure_onset_timestamp": "00:04.1",
+    "vlm_failure_onset_frame": 209,
+    "vlm_failure_onset_step": 209,
+    "vlm_coarse_failure_onset_seconds": 4.0,
+    "vlm_failure_window_start_seconds": 3.5,
+    "vlm_failure_window_end_seconds": 5.0,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot closed its gripper on the yellow block, but failed to establish a secure grip, causing the block to slip out as the arm lifted.",
+    "vlm_failure_reason_summary": "improper grasp alignment; failed object pick-up",
+    "vlm_recovery_action": "Re-align the gripper directly over the yellow block, close the fingers securely around it, lift it, and stack it onto the wooden block.",
+    "vlm_recovery_action_summary": "re-align gripper; regrasp yellow block; stack on wooden block",
+    "vlm_prevention_action": "Ensure the gripper is properly centered and at the correct height over the yellow block before closing the fingers to guarantee a solid grasp.",
+    "vlm_prevention_action_summary": "center gripper on block; close gripper securely",
+    "vlm_justification": "At 4.0 seconds, the robot attempts to close its gripper around the yellow block, but the block pops out of the fingers as the arm begins lifting, leaving the block on the table.",
+    "vlm_temporal_justification": "At 02:38 (second 158), the gripper fully closes its fingers without establishing a proper grip on the yellow block, leaving the block sitting ungrasped on the table as the arm prepares to lift.",
+    "vlm_video_fps": 51.6,
+    "vlm_video_num_frames": 381,
+    "vlm_video_duration_seconds": 7.364341085271318,
+    "vlm_usage": {
+      "input_tokens": 21896,
+      "output_tokens": 374,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 4.0,\n \"failure_window_start_seconds\": 3.5,\n \"failure_window_end_seconds\": 5.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot closed its gripper on the yellow block, but failed to establish a secure grip, causing the block to slip out as the arm lifted.\",\n \"failure_reason_summary\": \"improper grasp alignment; failed object pick-up\",\n \"recovery_action\": \"Re-align the gripper directly over the yellow block, close the fingers securely around it, lift it, and stack it onto the wooden block.\",\n \"recovery_action_summary\": \"re-align gripper; regrasp yellow block; stack on wooden block\",\n \"prevention_action\": \"Ensure the gripper is properly centered and at the correct height over the yellow block before closing the fingers to guarantee a solid grasp.\",\n \"prevention_action_summary\": \"center gripper on block; close gripper securely\",\n \"justification\": \"At 4.0 seconds, the robot attempts to close its gripper around the yellow block, but the block pops out of the fingers as the arm begins lifting, leaving the block on the table.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 209,
+    "failure_sim_step": 209,
+    "failure_policy_step": 34,
+    "failure_chunk_index": 5,
+    "first_post_failure_policy_step": 35,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-align the gripper directly over the yellow block, close the fingers securely around it, lift it, and stack it onto the wooden block."
+  }
+}
\ No newline at end of file
diff --git a/blocks/ep008/labels.npz b/blocks/ep008/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..b85d28be6add82e16fbb780af0ccfabb2107c0ae
--- /dev/null
+++ b/blocks/ep008/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9df7ac8737676cf951a148860a73bcb12513d6a37354c3105fa5a922990f6745
+size 3812
diff --git a/blocks/ep009/example.md b/blocks/ep009/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..dbc9299db6858d5a3dd07f5c9e40ead053a79428
--- /dev/null
+++ b/blocks/ep009/example.md
@@ -0,0 +1,23 @@
+# t000__ep009__blocks
+
+- **suite / variant:** `` / ``
+- **instruction:** Stack the yellow block on the wooden block
+- **success:** True
+- **control steps:** 670  (13.1s at 51 Hz)
+- **policy inferences:** 115  (replan_steps=6)
+- **video:** `blocks/ep009/rollout.mp4`
+- **features:** `115` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–152 | 0–25 | reaching for yellow block |
+| 3.0–6.0 | 153–305 | 25–50 | grasping and lifting yellow block |
+| 6.0–9.0 | 306–458 | 51–76 | carrying yellow block to wooden block |
+| 9.0–12.0 | 459–611 | 76–101 | stacking yellow block on wooden block |
+| 12.0–13.1 | 612–669 | 102–111 | releasing block and retracting arm |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/blocks/ep009/labels.json b/blocks/ep009/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..50ac3d92f74585053117d72d05676416681caa28
--- /dev/null
+++ b/blocks/ep009/labels.json
@@ -0,0 +1,90 @@
+{
+  "rollout_id": "t000__ep009__blocks",
+  "scene_variant": "",
+  "task_id": 0,
+  "task_file": "",
+  "instruction": "Stack the yellow block on the wooden block",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T19:58:28+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.99412915851272,
+      "control_step_start": 0,
+      "control_step_end": 152,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for yellow block",
+      "description": "reaching for yellow block"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.99412915851272,
+      "t_end_sec": 5.98825831702544,
+      "control_step_start": 153,
+      "control_step_end": 305,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "grasping and lifting yellow block",
+      "description": "grasping and lifting yellow block"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.98825831702544,
+      "t_end_sec": 8.982387475538161,
+      "control_step_start": 306,
+      "control_step_end": 458,
+      "policy_step_start": 51,
+      "policy_step_end": 76,
+      "phrase": "carrying yellow block to wooden block",
+      "description": "carrying yellow block to wooden block"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.982387475538161,
+      "t_end_sec": 11.97651663405088,
+      "control_step_start": 459,
+      "control_step_end": 611,
+      "policy_step_start": 76,
+      "policy_step_end": 101,
+      "phrase": "stacking yellow block on wooden block",
+      "description": "stacking yellow block on wooden block"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 11.97651663405088,
+      "t_end_sec": 13.111545988258317,
+      "control_step_start": 612,
+      "control_step_end": 669,
+      "policy_step_start": 102,
+      "policy_step_end": 111,
+      "phrase": "releasing block and retracting arm",
+      "description": "releasing block and retracting arm"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/blocks/ep009/labels.npz b/blocks/ep009/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..9ad06bbd68629eebb62b90fa4549d0996b000fac
--- /dev/null
+++ b/blocks/ep009/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fd2950b8cfe1c502063f0f4f1213e6b04e135c6e5b5f77babcd0d1b1acd511f5
+size 3171
diff --git a/blocks/ep010/example.md b/blocks/ep010/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..dac0f852f6a73480703f1513a8b43b272165488f
--- /dev/null
+++ b/blocks/ep010/example.md
@@ -0,0 +1,33 @@
+# t000__ep010__blocks
+
+- **suite / variant:** `` / ``
+- **instruction:** Stack the yellow block on the wooden block
+- **success:** False
+- **control steps:** 359  (7.1s at 51 Hz)
+- **policy inferences:** 61  (replan_steps=6)
+- **video:** `blocks/ep010/rollout.mp4`
+- **features:** `61` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | positioning gripper above yellow block |
+| 3.0–6.0 | 152–303 | 25–50 | attempting grasp on yellow block |
+| 6.0–7.1 | 304–358 | 50–59 | moving empty gripper to wooden block |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 4.547244094488189 s (00:04.5)
+- **frame / control_step:** 231 (policy_step=38, chunk_index=3)
+- **coarse onset:** 4.0 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the yellow block but failed to establish a secure grip, leaving the block on the table while moving its empty gripper toward the wooden block.
+  Summary: failed grasp on yellow block; moving empty gripper
+- **recovery:** Move the arm back to the yellow block, center the end-effector, securely grasp the block, and proceed to stack it on the wooden block.
+  Summary: re-align gripper; grasp yellow block; stack on wooden block
+- **prevention:** Ensure the gripper is properly aligned and centered over the yellow block before closing the fingers so that the object is securely held prior to moving toward the target block.
+  Summary: center gripper over yellow block; secure grasp before moving
+- **justification:** At around 4.0 seconds, the robot closes its fingers around the yellow block but slips off without securing it. The end-effector lifts away empty at 5.0 seconds and moves toward the wooden block as if carrying the block.
diff --git a/blocks/ep010/labels.json b/blocks/ep010/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..8998bf5733fa2edcf0e0c7b4c6cf3e9418289e77
--- /dev/null
+++ b/blocks/ep010/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t000__ep010__blocks",
+  "scene_variant": "",
+  "task_id": 0,
+  "task_file": "",
+  "instruction": "Stack the yellow block on the wooden block",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T19:59:01+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9921259842519685,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "positioning gripper above yellow block",
+      "description": "positioning gripper above yellow block"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9921259842519685,
+      "t_end_sec": 5.984251968503937,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "attempting grasp on yellow block",
+      "description": "attempting grasp on yellow block"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.984251968503937,
+      "t_end_sec": 7.0669291338582685,
+      "control_step_start": 304,
+      "control_step_end": 358,
+      "policy_step_start": 50,
+      "policy_step_end": 59,
+      "phrase": "moving empty gripper to wooden block",
+      "description": "moving empty gripper to wooden block"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 4.547244094488189,
+    "vlm_failure_onset_timestamp": "00:04.5",
+    "vlm_failure_onset_frame": 231,
+    "vlm_failure_onset_step": 231,
+    "vlm_coarse_failure_onset_seconds": 4.0,
+    "vlm_failure_window_start_seconds": 3.5,
+    "vlm_failure_window_end_seconds": 4.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the yellow block but failed to establish a secure grip, leaving the block on the table while moving its empty gripper toward the wooden block.",
+    "vlm_failure_reason_summary": "failed grasp on yellow block; moving empty gripper",
+    "vlm_recovery_action": "Move the arm back to the yellow block, center the end-effector, securely grasp the block, and proceed to stack it on the wooden block.",
+    "vlm_recovery_action_summary": "re-align gripper; grasp yellow block; stack on wooden block",
+    "vlm_prevention_action": "Ensure the gripper is properly aligned and centered over the yellow block before closing the fingers so that the object is securely held prior to moving toward the target block.",
+    "vlm_prevention_action_summary": "center gripper over yellow block; secure grasp before moving",
+    "vlm_justification": "At around 4.0 seconds, the robot closes its fingers around the yellow block but slips off without securing it. The end-effector lifts away empty at 5.0 seconds and moves toward the wooden block as if carrying the block.",
+    "vlm_temporal_justification": "At 03:00 (180 seconds into the clip), the gripper finishes closing its fingers without securing the yellow block, leaving it sitting on the table as the arm prepares to lift away.",
+    "vlm_video_fps": 50.8,
+    "vlm_video_num_frames": 359,
+    "vlm_video_duration_seconds": 7.04724409448819,
+    "vlm_usage": {
+      "input_tokens": 21521,
+      "output_tokens": 389,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 4.0,\n \"failure_window_start_seconds\": 3.5,\n \"failure_window_end_seconds\": 4.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the yellow block but failed to establish a secure grip, leaving the block on the table while moving its empty gripper toward the wooden block.\",\n \"failure_reason_summary\": \"failed grasp on yellow block; moving empty gripper\",\n \"recovery_action\": \"Move the arm back to the yellow block, center the end-effector, securely grasp the block, and proceed to stack it on the wooden block.\",\n \"recovery_action_summary\": \"re-align gripper; grasp yellow block; stack on wooden block\",\n \"prevention_action\": \"Ensure the gripper is properly aligned and centered over the yellow block before closing the fingers so that the object is securely held prior to moving toward the target block.\",\n \"prevention_action_summary\": \"center gripper over yellow block; secure grasp before moving\",\n \"justification\": \"At around 4.0 seconds, the robot closes its fingers around the yellow block but slips off without securing it. The end-effector lifts away empty at 5.0 seconds and moves toward the wooden block as if carrying the block.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 231,
+    "failure_sim_step": 231,
+    "failure_policy_step": 38,
+    "failure_chunk_index": 3,
+    "first_post_failure_policy_step": 39,
+    "failure_type": "grasp_failure",
+    "correction_action": "Move the arm back to the yellow block, center the end-effector, securely grasp the block, and proceed to stack it on the wooden block."
+  }
+}
\ No newline at end of file
diff --git a/blocks/ep010/labels.npz b/blocks/ep010/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..095b750931d66cd66291df39162ae89f0b8db0d5
--- /dev/null
+++ b/blocks/ep010/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:95fcd8c62b6d636dec4348edd246555802d4f038874113ea17c1dd814170e2d5
+size 3877
diff --git a/blocks/ep011/example.md b/blocks/ep011/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..83a29f14f47f99aa6527ef06489cce3b77038e0a
--- /dev/null
+++ b/blocks/ep011/example.md
@@ -0,0 +1,33 @@
+# t000__ep011__blocks
+
+- **suite / variant:** `` / ``
+- **instruction:** Stack the yellow block on the wooden block
+- **success:** False
+- **control steps:** 417  (8.1s at 51 Hz)
+- **policy inferences:** 73  (replan_steps=6)
+- **video:** `blocks/ep011/rollout.mp4`
+- **features:** `73` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–153 | 0–25 | adjusting gripper orientation above table |
+| 3.0–6.0 | 154–307 | 25–51 | hovering stationary in mid air |
+| 6.0–8.1 | 308–416 | 51–69 | remaining still above workspace |
+
+## Failure
+
+- **yes**
+- **onset type:** operator_intervention
+- **onset:** 0.05859375 s (00:00.1)
+- **frame / control_step:** 3 (policy_step=0, chunk_index=3)
+- **coarse onset:** 3.0 s (refined=True)
+- **mode:** stuck_or_no_progress
+- **reason:** The robot arm ceases making useful downward progress, hovering suspended high above the workspace without attempting to reach or grasp the yellow block.
+  Summary: arm stuck hovering above table; no trajectory toward block
+- **recovery:** Manually command or re-initialize the robot arm to descend directly to the yellow block, close the end effector to grasp it, and move it onto the wooden block.
+  Summary: descend to yellow block; grasp and stack on wooden block
+- **prevention:** Ensure the motion policy continuously advances downward toward the target block rather than halting in mid-air.
+  Summary: descend toward yellow block; execute pick and place
+- **justification:** The robot arm moves into a position above the table around 3.0 seconds into the rollout, but fails to descend toward the objects, remaining stuck hovering in mid-air until the episode expires.
diff --git a/blocks/ep011/labels.json b/blocks/ep011/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..cdec3310c6a2297fd16a242ae6571943a9b9a0b2
--- /dev/null
+++ b/blocks/ep011/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t000__ep011__blocks",
+  "scene_variant": "",
+  "task_id": 0,
+  "task_file": "",
+  "instruction": "Stack the yellow block on the wooden block",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T19:59:33+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0078125,
+      "control_step_start": 0,
+      "control_step_end": 153,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "adjusting gripper orientation above table",
+      "description": "adjusting gripper orientation above table"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0078125,
+      "t_end_sec": 6.015625,
+      "control_step_start": 154,
+      "control_step_end": 307,
+      "policy_step_start": 25,
+      "policy_step_end": 51,
+      "phrase": "hovering stationary in mid air",
+      "description": "hovering stationary in mid air"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.015625,
+      "t_end_sec": 8.14453125,
+      "control_step_start": 308,
+      "control_step_end": 416,
+      "policy_step_start": 51,
+      "policy_step_end": 69,
+      "phrase": "remaining still above workspace",
+      "description": "remaining still above workspace"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "stuck_or_no_progress",
+    "vlm_failure_onset_type": "operator_intervention",
+    "vlm_failure_onset_seconds": 0.05859375,
+    "vlm_failure_onset_timestamp": "00:00.1",
+    "vlm_failure_onset_frame": 3,
+    "vlm_failure_onset_step": 3,
+    "vlm_coarse_failure_onset_seconds": 3.0,
+    "vlm_failure_window_start_seconds": 2.5,
+    "vlm_failure_window_end_seconds": 3.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot arm ceases making useful downward progress, hovering suspended high above the workspace without attempting to reach or grasp the yellow block.",
+    "vlm_failure_reason_summary": "arm stuck hovering above table; no trajectory toward block",
+    "vlm_recovery_action": "Manually command or re-initialize the robot arm to descend directly to the yellow block, close the end effector to grasp it, and move it onto the wooden block.",
+    "vlm_recovery_action_summary": "descend to yellow block; grasp and stack on wooden block",
+    "vlm_prevention_action": "Ensure the motion policy continuously advances downward toward the target block rather than halting in mid-air.",
+    "vlm_prevention_action_summary": "descend toward yellow block; execute pick and place",
+    "vlm_justification": "The robot arm moves into a position above the table around 3.0 seconds into the rollout, but fails to descend toward the objects, remaining stuck hovering in mid-air until the episode expires.",
+    "vlm_temporal_justification": "At clip second 3 (frame 3 of the original rollout), the robot arm stops making any useful downward progress toward the yellow block and remains hovering suspended high above the workspace.",
+    "vlm_video_fps": 51.2,
+    "vlm_video_num_frames": 417,
+    "vlm_video_duration_seconds": 8.125,
+    "vlm_usage": {
+      "input_tokens": 21829,
+      "output_tokens": 360,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"stuck_or_no_progress\",\n \"onset_type\": \"operator_intervention\",\n \"failure_onset_seconds\": 3.0,\n \"failure_window_start_seconds\": 2.5,\n \"failure_window_end_seconds\": 3.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot arm ceases making useful downward progress, hovering suspended high above the workspace without attempting to reach or grasp the yellow block.\",\n \"failure_reason_summary\": \"arm stuck hovering above table; no trajectory toward block\",\n \"recovery_action\": \"Manually command or re-initialize the robot arm to descend directly to the yellow block, close the end effector to grasp it, and move it onto the wooden block.\",\n \"recovery_action_summary\": \"descend to yellow block; grasp and stack on wooden block\",\n \"prevention_action\": \"Ensure the motion policy continuously advances downward toward the target block rather than halting in mid-air.\",\n \"prevention_action_summary\": \"descend toward yellow block; execute pick and place\",\n \"justification\": \"The robot arm moves into a position above the table around 3.0 seconds into the rollout, but fails to descend toward the objects, remaining stuck hovering in mid-air until the episode expires.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 3,
+    "failure_sim_step": 3,
+    "failure_policy_step": 0,
+    "failure_chunk_index": 3,
+    "first_post_failure_policy_step": 1,
+    "failure_type": "stuck_or_no_progress",
+    "correction_action": "Manually command or re-initialize the robot arm to descend directly to the yellow block, close the end effector to grasp it, and move it onto the wooden block."
+  }
+}
\ No newline at end of file
diff --git a/blocks/ep011/labels.npz b/blocks/ep011/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..44a08189ba8866f83f66a2f5127b8baf097db5db
--- /dev/null
+++ b/blocks/ep011/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:001aa862ac2d11cda5ec43fce01bf2b1d24082ed52c12348048a7a3f78ff6d91
+size 3885
diff --git a/blocks/ep012/example.md b/blocks/ep012/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..a9c1e88b144218bdfcd934bbe7d5a7432f34148a
--- /dev/null
+++ b/blocks/ep012/example.md
@@ -0,0 +1,32 @@
+# t000__ep012__blocks
+
+- **suite / variant:** `` / ``
+- **instruction:** Stack the yellow block on the wooden block
+- **success:** False
+- **control steps:** 297  (5.9s at 50 Hz)
+- **policy inferences:** 52  (replan_steps=5)
+- **video:** `blocks/ep012/rollout.mp4`
+- **features:** `52` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–149 | 0–29 | reaching toward yellow block |
+| 3.0–5.9 | 150–296 | 30–59 | closing gripper and lifting empty |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 5.02 s (00:05.0)
+- **frame / control_step:** 251 (policy_step=50, chunk_index=1)
+- **coarse onset:** 3.8 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the yellow block, but failed to securely grasp it, causing the end-effector to lift up empty while the yellow block remained on the table.
+  Summary: failed grasp on yellow block; lifting empty gripper
+- **recovery:** Re-align the end-effector above the yellow block, lower it to surround the block, close the fingers firmly, and proceed to stack it on the wooden block.
+  Summary: re-align gripper; regrasp yellow block; stack on wooden block
+- **prevention:** Ensure the gripper descends sufficiently low and centers around the yellow block before closing the fingers to guarantee a secure grasp.
+  Summary: position lower over yellow block; secure firm grip before lifting
+- **justification:** Around 3.8 seconds into the rollout, the robot closes its gripper on the yellow block but fails to establish a secure hold. It subsequently lifts away empty, leaving the block on the table.
diff --git a/blocks/ep012/labels.json b/blocks/ep012/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..c6d94dad0aad56b01c6fdc0167fb5f5047dbb55a
--- /dev/null
+++ b/blocks/ep012/labels.json
@@ -0,0 +1,87 @@
+{
+  "rollout_id": "t000__ep012__blocks",
+  "scene_variant": "",
+  "task_id": 0,
+  "task_file": "",
+  "instruction": "Stack the yellow block on the wooden block",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:00:05+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0,
+      "control_step_start": 0,
+      "control_step_end": 149,
+      "policy_step_start": 0,
+      "policy_step_end": 29,
+      "phrase": "reaching toward yellow block",
+      "description": "reaching toward yellow block"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0,
+      "t_end_sec": 5.94,
+      "control_step_start": 150,
+      "control_step_end": 296,
+      "policy_step_start": 30,
+      "policy_step_end": 59,
+      "phrase": "closing gripper and lifting empty",
+      "description": "closing gripper and lifting empty"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 5.02,
+    "vlm_failure_onset_timestamp": "00:05.0",
+    "vlm_failure_onset_frame": 251,
+    "vlm_failure_onset_step": 251,
+    "vlm_coarse_failure_onset_seconds": 3.8,
+    "vlm_failure_window_start_seconds": 3.3,
+    "vlm_failure_window_end_seconds": 4.3,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the yellow block, but failed to securely grasp it, causing the end-effector to lift up empty while the yellow block remained on the table.",
+    "vlm_failure_reason_summary": "failed grasp on yellow block; lifting empty gripper",
+    "vlm_recovery_action": "Re-align the end-effector above the yellow block, lower it to surround the block, close the fingers firmly, and proceed to stack it on the wooden block.",
+    "vlm_recovery_action_summary": "re-align gripper; regrasp yellow block; stack on wooden block",
+    "vlm_prevention_action": "Ensure the gripper descends sufficiently low and centers around the yellow block before closing the fingers to guarantee a secure grasp.",
+    "vlm_prevention_action_summary": "position lower over yellow block; secure firm grip before lifting",
+    "vlm_justification": "Around 3.8 seconds into the rollout, the robot closes its gripper on the yellow block but fails to establish a secure hold. It subsequently lifts away empty, leaving the block on the table.",
+    "vlm_temporal_justification": "At 03:31 in the clip, the robot arm begins moving upward after closing its gripper, but the yellow block remains on the table surface, making the grasp failure clearly visible.",
+    "vlm_video_fps": 50.0,
+    "vlm_video_num_frames": 297,
+    "vlm_video_duration_seconds": 5.92,
+    "vlm_usage": {
+      "input_tokens": 18436,
+      "output_tokens": 376,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 3.8,\n \"failure_window_start_seconds\": 3.3,\n \"failure_window_end_seconds\": 4.3,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the yellow block, but failed to securely grasp it, causing the end-effector to lift up empty while the yellow block remained on the table.\",\n \"failure_reason_summary\": \"failed grasp on yellow block; lifting empty gripper\",\n \"recovery_action\": \"Re-align the end-effector above the yellow block, lower it to surround the block, close the fingers firmly, and proceed to stack it on the wooden block.\",\n \"recovery_action_summary\": \"re-align gripper; regrasp yellow block; stack on wooden block\",\n \"prevention_action\": \"Ensure the gripper descends sufficiently low and centers around the yellow block before closing the fingers to guarantee a secure grasp.\",\n \"prevention_action_summary\": \"position lower over yellow block; secure firm grip before lifting\",\n \"justification\": \"Around 3.8 seconds into the rollout, the robot closes its gripper on the yellow block but fails to establish a secure hold. It subsequently lifts away empty, leaving the block on the table.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 251,
+    "failure_sim_step": 251,
+    "failure_policy_step": 50,
+    "failure_chunk_index": 1,
+    "first_post_failure_policy_step": 51,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-align the end-effector above the yellow block, lower it to surround the block, close the fingers firmly, and proceed to stack it on the wooden block."
+  }
+}
\ No newline at end of file
diff --git a/blocks/ep012/labels.npz b/blocks/ep012/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..fadaf275667f87ec0e38332a9797d53611ab9c4a
--- /dev/null
+++ b/blocks/ep012/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f52d3a626c61011b8a3991876cd4e351b3245aef7c6f7589b4ff1d0db71cd3c4
+size 3834
diff --git a/blocks/ep013/example.md b/blocks/ep013/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..f1b6cc21e2ef86e40061d713e8562e9cfd425bda
--- /dev/null
+++ b/blocks/ep013/example.md
@@ -0,0 +1,41 @@
+# t000__ep013__blocks
+
+- **suite / variant:** `` / ``
+- **instruction:** Stack the yellow block on the wooden block
+- **success:** False
+- **control steps:** 1610  (31.9s at 50 Hz)
+- **policy inferences:** 275  (replan_steps=6)
+- **video:** `blocks/ep013/rollout.mp4`
+- **features:** `275` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | approaching yellow block |
+| 3.0–6.0 | 151–301 | 25–50 | grasping and lifting yellow block |
+| 6.0–9.0 | 302–452 | 50–75 | moving yellow block over wooden block |
+| 9.0–12.0 | 453–603 | 75–100 | lowering yellow block onto wooden block |
+| 12.0–15.0 | 604–754 | 100–125 | holding yellow block on wooden block |
+| 15.0–18.0 | 755–905 | 125–150 | adjusting gripper on yellow block |
+| 18.0–21.0 | 906–1056 | 151–176 | attempting release on wooden block |
+| 21.0–24.0 | 1057–1207 | 176–201 | shifting position over wooden block |
+| 24.0–27.0 | 1208–1358 | 201–226 | knocking yellow block off wooden block |
+| 27.0–30.0 | 1359–1509 | 226–251 | idle near displaced yellow block |
+| 30.0–31.9 | 1510–1609 | 251–268 | idle near objects |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 22.400793650793652 s (00:22.4)
+- **frame / control_step:** 1129 (policy_step=188, chunk_index=1)
+- **coarse onset:** 25.0 s (refined=True)
+- **mode:** placement_or_insertion_failure
+- **reason:** While attempting to place the yellow block on the wooden block, the robot failed to release the block cleanly and pushed it off the wooden block onto the table.
+  Summary: improper release alignment; block knocked off target
+- **recovery:** Re-grasp the yellow block from the tabletop, align it over the wooden block, and place it stably on top.
+  Summary: re-grasp yellow block; place on wooden block
+- **prevention:** Align the yellow block directly over the center of the wooden block and fully open the gripper before lifting away to prevent dragging or knocking the block.
+  Summary: align yellow block squarely; release cleanly before retracting
+- **justification:** The robot successfully picked up the yellow block and brought it over the wooden block. However, around 25.0 seconds, during the release phase, the gripper dragged and knocked the yellow block off the wooden block onto the table surface, resulting in a placement failure.
diff --git a/blocks/ep013/labels.json b/blocks/ep013/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..1a0650a057e3b70f3e8b021b548361efb37f6918
--- /dev/null
+++ b/blocks/ep013/labels.json
@@ -0,0 +1,186 @@
+{
+  "rollout_id": "t000__ep013__blocks",
+  "scene_variant": "",
+  "task_id": 0,
+  "task_file": "",
+  "instruction": "Stack the yellow block on the wooden block",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:00:42+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.996031746031746,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "approaching yellow block",
+      "description": "approaching yellow block"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.996031746031746,
+      "t_end_sec": 5.992063492063492,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "grasping and lifting yellow block",
+      "description": "grasping and lifting yellow block"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.992063492063492,
+      "t_end_sec": 8.988095238095239,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "moving yellow block over wooden block",
+      "description": "moving yellow block over wooden block"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.988095238095239,
+      "t_end_sec": 11.984126984126984,
+      "control_step_start": 453,
+      "control_step_end": 603,
+      "policy_step_start": 75,
+      "policy_step_end": 100,
+      "phrase": "lowering yellow block onto wooden block",
+      "description": "lowering yellow block onto wooden block"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 11.984126984126984,
+      "t_end_sec": 14.980158730158731,
+      "control_step_start": 604,
+      "control_step_end": 754,
+      "policy_step_start": 100,
+      "policy_step_end": 125,
+      "phrase": "holding yellow block on wooden block",
+      "description": "holding yellow block on wooden block"
+    },
+    {
+      "segment_index": 5,
+      "t_start_sec": 14.980158730158731,
+      "t_end_sec": 17.976190476190478,
+      "control_step_start": 755,
+      "control_step_end": 905,
+      "policy_step_start": 125,
+      "policy_step_end": 150,
+      "phrase": "adjusting gripper on yellow block",
+      "description": "adjusting gripper on yellow block"
+    },
+    {
+      "segment_index": 6,
+      "t_start_sec": 17.976190476190478,
+      "t_end_sec": 20.97222222222222,
+      "control_step_start": 906,
+      "control_step_end": 1056,
+      "policy_step_start": 151,
+      "policy_step_end": 176,
+      "phrase": "attempting release on wooden block",
+      "description": "attempting release on wooden block"
+    },
+    {
+      "segment_index": 7,
+      "t_start_sec": 20.97222222222222,
+      "t_end_sec": 23.96825396825397,
+      "control_step_start": 1057,
+      "control_step_end": 1207,
+      "policy_step_start": 176,
+      "policy_step_end": 201,
+      "phrase": "shifting position over wooden block",
+      "description": "shifting position over wooden block"
+    },
+    {
+      "segment_index": 8,
+      "t_start_sec": 23.96825396825397,
+      "t_end_sec": 26.964285714285715,
+      "control_step_start": 1208,
+      "control_step_end": 1358,
+      "policy_step_start": 201,
+      "policy_step_end": 226,
+      "phrase": "knocking yellow block off wooden block",
+      "description": "knocking yellow block off wooden block"
+    },
+    {
+      "segment_index": 9,
+      "t_start_sec": 26.964285714285715,
+      "t_end_sec": 29.960317460317462,
+      "control_step_start": 1359,
+      "control_step_end": 1509,
+      "policy_step_start": 226,
+      "policy_step_end": 251,
+      "phrase": "idle near displaced yellow block",
+      "description": "idle near displaced yellow block"
+    },
+    {
+      "segment_index": 10,
+      "t_start_sec": 29.960317460317462,
+      "t_end_sec": 31.944444444444446,
+      "control_step_start": 1510,
+      "control_step_end": 1609,
+      "policy_step_start": 251,
+      "policy_step_end": 268,
+      "phrase": "idle near objects",
+      "description": "idle near objects"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "placement_or_insertion_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 22.400793650793652,
+    "vlm_failure_onset_timestamp": "00:22.4",
+    "vlm_failure_onset_frame": 1129,
+    "vlm_failure_onset_step": 1129,
+    "vlm_coarse_failure_onset_seconds": 25.0,
+    "vlm_failure_window_start_seconds": 24.5,
+    "vlm_failure_window_end_seconds": 27.0,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "While attempting to place the yellow block on the wooden block, the robot failed to release the block cleanly and pushed it off the wooden block onto the table.",
+    "vlm_failure_reason_summary": "improper release alignment; block knocked off target",
+    "vlm_recovery_action": "Re-grasp the yellow block from the tabletop, align it over the wooden block, and place it stably on top.",
+    "vlm_recovery_action_summary": "re-grasp yellow block; place on wooden block",
+    "vlm_prevention_action": "Align the yellow block directly over the center of the wooden block and fully open the gripper before lifting away to prevent dragging or knocking the block.",
+    "vlm_prevention_action_summary": "align yellow block squarely; release cleanly before retracting",
+    "vlm_justification": "The robot successfully picked up the yellow block and brought it over the wooden block. However, around 25.0 seconds, during the release phase, the gripper dragged and knocked the yellow block off the wooden block onto the table surface, resulting in a placement failure.",
+    "vlm_temporal_justification": "At second 20, the robot arm begins lowering the yellow block while misaligned with the pedestal. Instead of setting it down on top, the gripper pushes the block against the edge of the pedestal, causing it to tip over and fall onto the table.",
+    "vlm_video_fps": 50.4,
+    "vlm_video_num_frames": 1610,
+    "vlm_video_duration_seconds": 31.924603174603174,
+    "vlm_usage": {
+      "input_tokens": 22983,
+      "output_tokens": 397,
+      "cached_tokens": 2177
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"placement_or_insertion_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 25.0,\n \"failure_window_start_seconds\": 24.5,\n \"failure_window_end_seconds\": 27.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"While attempting to place the yellow block on the wooden block, the robot failed to release the block cleanly and pushed it off the wooden block onto the table.\",\n \"failure_reason_summary\": \"improper release alignment; block knocked off target\",\n \"recovery_action\": \"Re-grasp the yellow block from the tabletop, align it over the wooden block, and place it stably on top.\",\n \"recovery_action_summary\": \"re-grasp yellow block; place on wooden block\",\n \"prevention_action\": \"Align the yellow block directly over the center of the wooden block and fully open the gripper before lifting away to prevent dragging or knocking the block.\",\n \"prevention_action_summary\": \"align yellow block squarely; release cleanly before retracting\",\n \"justification\": \"The robot successfully picked up the yellow block and brought it over the wooden block. However, around 25.0 seconds, during the release phase, the gripper dragged and knocked the yellow block off the wooden block onto the table surface, resulting in a placement failure.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 1129,
+    "failure_sim_step": 1129,
+    "failure_policy_step": 188,
+    "failure_chunk_index": 1,
+    "first_post_failure_policy_step": 189,
+    "failure_type": "placement_or_insertion_failure",
+    "correction_action": "Re-grasp the yellow block from the tabletop, align it over the wooden block, and place it stably on top."
+  }
+}
\ No newline at end of file
diff --git a/blocks/ep013/labels.npz b/blocks/ep013/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..9c2c92fbfaf80f0b4bae85e7e3eb62cff29e4b58
--- /dev/null
+++ b/blocks/ep013/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ccf7a168a8f1c0510311e0782b17a9baf299c9ec104febadde3c3aed084840ea
+size 4142
diff --git a/blocks/ep014/example.md b/blocks/ep014/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..5145ee79995bc7f4c74aa27e677a988d2048b6ca
--- /dev/null
+++ b/blocks/ep014/example.md
@@ -0,0 +1,33 @@
+# t000__ep014__blocks
+
+- **suite / variant:** `` / ``
+- **instruction:** Stack the yellow block on the wooden block
+- **success:** False
+- **control steps:** 443  (8.8s at 50 Hz)
+- **policy inferences:** 75  (replan_steps=6)
+- **video:** `blocks/ep014/rollout.mp4`
+- **features:** `75` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | approaching yellow block |
+| 3.0–6.0 | 151–301 | 25–50 | attempting grasp on yellow block |
+| 6.0–8.8 | 302–442 | 50–73 | lowering gripper over yellow block |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 3.645418326693227 s (00:03.6)
+- **frame / control_step:** 183 (policy_step=30, chunk_index=3)
+- **coarse onset:** 3.5 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot descends onto the yellow block but fails to properly align its fingers around it, causing the gripper to close without securing the block.
+  Summary: misaligned fingers; failed to grasp yellow block
+- **recovery:** Re-align the open gripper directly over the yellow block, lower onto it, close the fingers firmly to secure the block, and lift it before attempting to stack it on the wooden block.
+  Summary: realign gripper; grasp yellow block; lift block
+- **prevention:** Ensure the open end-effector is centered directly over the yellow block before closing the gripper to achieve a firm hold.
+  Summary: center gripper over block; secure grasp
+- **justification:** Around 3.5 seconds into the video, the robot attempts to close its gripper onto the yellow block, but fails to acquire a grasp, lifting back up without the object.
diff --git a/blocks/ep014/labels.json b/blocks/ep014/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..da74b04cef8b4443738cd241915193cd8d977d78
--- /dev/null
+++ b/blocks/ep014/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t000__ep014__blocks",
+  "scene_variant": "",
+  "task_id": 0,
+  "task_file": "",
+  "instruction": "Stack the yellow block on the wooden block",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:01:19+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00796812749004,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "approaching yellow block",
+      "description": "approaching yellow block"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00796812749004,
+      "t_end_sec": 6.01593625498008,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "attempting grasp on yellow block",
+      "description": "attempting grasp on yellow block"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01593625498008,
+      "t_end_sec": 8.824701195219124,
+      "control_step_start": 302,
+      "control_step_end": 442,
+      "policy_step_start": 50,
+      "policy_step_end": 73,
+      "phrase": "lowering gripper over yellow block",
+      "description": "lowering gripper over yellow block"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 3.645418326693227,
+    "vlm_failure_onset_timestamp": "00:03.6",
+    "vlm_failure_onset_frame": 183,
+    "vlm_failure_onset_step": 183,
+    "vlm_coarse_failure_onset_seconds": 3.5,
+    "vlm_failure_window_start_seconds": 3.0,
+    "vlm_failure_window_end_seconds": 4.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot descends onto the yellow block but fails to properly align its fingers around it, causing the gripper to close without securing the block.",
+    "vlm_failure_reason_summary": "misaligned fingers; failed to grasp yellow block",
+    "vlm_recovery_action": "Re-align the open gripper directly over the yellow block, lower onto it, close the fingers firmly to secure the block, and lift it before attempting to stack it on the wooden block.",
+    "vlm_recovery_action_summary": "realign gripper; grasp yellow block; lift block",
+    "vlm_prevention_action": "Ensure the open end-effector is centered directly over the yellow block before closing the gripper to achieve a firm hold.",
+    "vlm_prevention_action_summary": "center gripper over block; secure grasp",
+    "vlm_justification": "Around 3.5 seconds into the video, the robot attempts to close its gripper onto the yellow block, but fails to acquire a grasp, lifting back up without the object.",
+    "vlm_temporal_justification": "At 02:38 (158 seconds into the clip), the robot finishes descending over the yellow block. Its fingers are misaligned with the block's edges, causing the gripper to close at 02:39-02:40 without grasping the block.",
+    "vlm_video_fps": 50.2,
+    "vlm_video_num_frames": 443,
+    "vlm_video_duration_seconds": 8.804780876494023,
+    "vlm_usage": {
+      "input_tokens": 21516,
+      "output_tokens": 383,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 3.5,\n \"failure_window_start_seconds\": 3.0,\n \"failure_window_end_seconds\": 4.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot descends onto the yellow block but fails to properly align its fingers around it, causing the gripper to close without securing the block.\",\n \"failure_reason_summary\": \"misaligned fingers; failed to grasp yellow block\",\n \"recovery_action\": \"Re-align the open gripper directly over the yellow block, lower onto it, close the fingers firmly to secure the block, and lift it before attempting to stack it on the wooden block.\",\n \"recovery_action_summary\": \"realign gripper; grasp yellow block; lift block\",\n \"prevention_action\": \"Ensure the open end-effector is centered directly over the yellow block before closing the gripper to achieve a firm hold.\",\n \"prevention_action_summary\": \"center gripper over block; secure grasp\",\n \"justification\": \"Around 3.5 seconds into the video, the robot attempts to close its gripper onto the yellow block, but fails to acquire a grasp, lifting back up without the object.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 183,
+    "failure_sim_step": 183,
+    "failure_policy_step": 30,
+    "failure_chunk_index": 3,
+    "first_post_failure_policy_step": 31,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-align the open gripper directly over the yellow block, lower onto it, close the fingers firmly to secure the block, and lift it before attempting to stack it on the wooden block."
+  }
+}
\ No newline at end of file
diff --git a/blocks/ep014/labels.npz b/blocks/ep014/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..f489cc997369010e9a5d35f730155a1328c9bf2e
--- /dev/null
+++ b/blocks/ep014/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:275015d724ff2217b90129951b4e6c3786e9b3ee11f722523870809327567d07
+size 3801
diff --git a/blocks/ep015/example.md b/blocks/ep015/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..ce26a259de0d69e6963d085b76534dae68926d83
--- /dev/null
+++ b/blocks/ep015/example.md
@@ -0,0 +1,38 @@
+# t000__ep015__blocks
+
+- **suite / variant:** `` / ``
+- **instruction:** Stack the yellow block on the wooden block
+- **success:** False
+- **control steps:** 1080  (21.3s at 51 Hz)
+- **policy inferences:** 185  (replan_steps=6)
+- **video:** `blocks/ep015/rollout.mp4`
+- **features:** `185` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | approaching yellow block |
+| 3.0–6.0 | 152–303 | 25–50 | grasping yellow block |
+| 6.0–9.0 | 304–455 | 50–75 | carrying yellow block to target |
+| 9.0–12.0 | 456–607 | 76–101 | releasing yellow block off-center |
+| 12.0–15.0 | 608–759 | 101–126 | pushing yellow block aside |
+| 15.0–18.0 | 760–911 | 126–151 | dragging yellow block |
+| 18.0–21.0 | 912–1063 | 152–177 | hovering near yellow block |
+| 21.0–21.3 | 1064–1079 | 177–179 | idle near yellow block |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 9.703557312252965 s (00:09.7)
+- **frame / control_step:** 491 (policy_step=81, chunk_index=5)
+- **coarse onset:** 9.5 s (refined=True)
+- **mode:** placement_or_insertion_failure
+- **reason:** The robot misaligned the yellow block while attempting to stack it, causing the block to hit the edge of the wooden block and slide off onto the tabletop.
+  Summary: misaligned block placement; sliding off target block
+- **recovery:** Re-grasp the yellow block, align it directly above the center of the wooden block, and lower it gently onto the top surface.
+  Summary: re-grasp yellow block; stack cleanly on wooden block
+- **prevention:** Ensure the robot aligns the yellow block squarely over the center of the wooden block before opening the gripper to release it.
+  Summary: center block over target; release gently
+- **justification:** At around 9.5 seconds, the robot lowers the yellow block onto the edge of the wooden block instead of centering it. As the gripper opens, the yellow block slides off the wooden block and falls to the side, failing the goal to stack the yellow block on the wooden block.
diff --git a/blocks/ep015/labels.json b/blocks/ep015/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..6b8fa2440a81e3bc1a3e8255c0c0213b79488197
--- /dev/null
+++ b/blocks/ep015/labels.json
@@ -0,0 +1,153 @@
+{
+  "rollout_id": "t000__ep015__blocks",
+  "scene_variant": "",
+  "task_id": 0,
+  "task_file": "",
+  "instruction": "Stack the yellow block on the wooden block",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:01:54+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0039525691699605,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "approaching yellow block",
+      "description": "approaching yellow block"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0039525691699605,
+      "t_end_sec": 6.007905138339921,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "grasping yellow block",
+      "description": "grasping yellow block"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.007905138339921,
+      "t_end_sec": 9.011857707509881,
+      "control_step_start": 304,
+      "control_step_end": 455,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "carrying yellow block to target",
+      "description": "carrying yellow block to target"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.011857707509881,
+      "t_end_sec": 12.015810276679842,
+      "control_step_start": 456,
+      "control_step_end": 607,
+      "policy_step_start": 76,
+      "policy_step_end": 101,
+      "phrase": "releasing yellow block off-center",
+      "description": "releasing yellow block off-center"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 12.015810276679842,
+      "t_end_sec": 15.019762845849803,
+      "control_step_start": 608,
+      "control_step_end": 759,
+      "policy_step_start": 101,
+      "policy_step_end": 126,
+      "phrase": "pushing yellow block aside",
+      "description": "pushing yellow block aside"
+    },
+    {
+      "segment_index": 5,
+      "t_start_sec": 15.019762845849803,
+      "t_end_sec": 18.023715415019762,
+      "control_step_start": 760,
+      "control_step_end": 911,
+      "policy_step_start": 126,
+      "policy_step_end": 151,
+      "phrase": "dragging yellow block",
+      "description": "dragging yellow block"
+    },
+    {
+      "segment_index": 6,
+      "t_start_sec": 18.023715415019762,
+      "t_end_sec": 21.027667984189723,
+      "control_step_start": 912,
+      "control_step_end": 1063,
+      "policy_step_start": 152,
+      "policy_step_end": 177,
+      "phrase": "hovering near yellow block",
+      "description": "hovering near yellow block"
+    },
+    {
+      "segment_index": 7,
+      "t_start_sec": 21.027667984189723,
+      "t_end_sec": 21.343873517786562,
+      "control_step_start": 1064,
+      "control_step_end": 1079,
+      "policy_step_start": 177,
+      "policy_step_end": 179,
+      "phrase": "idle near yellow block",
+      "description": "idle near yellow block"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "placement_or_insertion_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 9.703557312252965,
+    "vlm_failure_onset_timestamp": "00:09.7",
+    "vlm_failure_onset_frame": 491,
+    "vlm_failure_onset_step": 491,
+    "vlm_coarse_failure_onset_seconds": 9.5,
+    "vlm_failure_window_start_seconds": 8.5,
+    "vlm_failure_window_end_seconds": 11.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot misaligned the yellow block while attempting to stack it, causing the block to hit the edge of the wooden block and slide off onto the tabletop.",
+    "vlm_failure_reason_summary": "misaligned block placement; sliding off target block",
+    "vlm_recovery_action": "Re-grasp the yellow block, align it directly above the center of the wooden block, and lower it gently onto the top surface.",
+    "vlm_recovery_action_summary": "re-grasp yellow block; stack cleanly on wooden block",
+    "vlm_prevention_action": "Ensure the robot aligns the yellow block squarely over the center of the wooden block before opening the gripper to release it.",
+    "vlm_prevention_action_summary": "center block over target; release gently",
+    "vlm_justification": "At around 9.5 seconds, the robot lowers the yellow block onto the edge of the wooden block instead of centering it. As the gripper opens, the yellow block slides off the wooden block and falls to the side, failing the goal to stack the yellow block on the wooden block.",
+    "vlm_temporal_justification": "At 02:42 (162 seconds into the clip), the robot lowers the yellow block directly onto the left edge of the wooden block rather than centered on top of it, causing the yellow block to tilt and slide off onto the tabletop.",
+    "vlm_video_fps": 50.6,
+    "vlm_video_num_frames": 1080,
+    "vlm_video_duration_seconds": 21.324110671936758,
+    "vlm_usage": {
+      "input_tokens": 22412,
+      "output_tokens": 395,
+      "cached_tokens": 1429
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"placement_or_insertion_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 9.5,\n \"failure_window_start_seconds\": 8.5,\n \"failure_window_end_seconds\": 11.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot misaligned the yellow block while attempting to stack it, causing the block to hit the edge of the wooden block and slide off onto the tabletop.\",\n \"failure_reason_summary\": \"misaligned block placement; sliding off target block\",\n \"recovery_action\": \"Re-grasp the yellow block, align it directly above the center of the wooden block, and lower it gently onto the top surface.\",\n \"recovery_action_summary\": \"re-grasp yellow block; stack cleanly on wooden block\",\n \"prevention_action\": \"Ensure the robot aligns the yellow block squarely over the center of the wooden block before opening the gripper to release it.\",\n \"prevention_action_summary\": \"center block over target; release gently\",\n \"justification\": \"At around 9.5 seconds, the robot lowers the yellow block onto the edge of the wooden block instead of centering it. As the gripper opens, the yellow block slides off the wooden block and falls to the side, failing the goal to stack the yellow block on the wooden block.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 491,
+    "failure_sim_step": 491,
+    "failure_policy_step": 81,
+    "failure_chunk_index": 5,
+    "first_post_failure_policy_step": 82,
+    "failure_type": "placement_or_insertion_failure",
+    "correction_action": "Re-grasp the yellow block, align it directly above the center of the wooden block, and lower it gently onto the top surface."
+  }
+}
\ No newline at end of file
diff --git a/blocks/ep015/labels.npz b/blocks/ep015/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..55a0ec059014164dac7654a57aa94638a135037c
--- /dev/null
+++ b/blocks/ep015/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3258a611e8a333a4e83768339ec9f2687971f7adb37e1c151a58ac896b53e77d
+size 3974
diff --git a/blocks/ep016/example.md b/blocks/ep016/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..c87ca2ab8270abeb830902f462ca7c44c3e995ac
--- /dev/null
+++ b/blocks/ep016/example.md
@@ -0,0 +1,37 @@
+# t000__ep016__blocks
+
+- **suite / variant:** `` / ``
+- **instruction:** Stack the yellow block on the wooden block
+- **success:** False
+- **control steps:** 932  (18.5s at 50 Hz)
+- **policy inferences:** 162  (replan_steps=6)
+- **video:** `blocks/ep016/rollout.mp4`
+- **features:** `162` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | approaching yellow block |
+| 3.0–6.0 | 152–303 | 25–50 | closing gripper and missing yellow block |
+| 6.0–9.0 | 304–455 | 50–75 | re-descending and attempting grasp |
+| 9.0–12.0 | 456–607 | 76–101 | pinching yellow block and slipping |
+| 12.0–15.0 | 608–759 | 101–126 | repeated grasp attempts on block |
+| 15.0–18.1 | 760–911 | 126–151 | oscillating over yellow block |
+| 18.1–18.5 | 912–931 | 152–155 | remaining stationary over yellow block |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 3.504950495049505 s (00:03.5)
+- **frame / control_step:** 177 (policy_step=29, chunk_index=3)
+- **coarse onset:** 3.5 s (refined=False)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the yellow block but failed to establish a secure grip, resulting in repeated failed grasp attempts and slipping without picking up the block.
+  Summary: failed grasp on yellow block; unable to lift block
+- **recovery:** Re-open the gripper, center the end-effector directly over the yellow block, lower to surround the block, and close the gripper securely before lifting.
+  Summary: re-open gripper; center over yellow block; grasp securely
+- **prevention:** Align the gripper accurately with the center and orientation of the yellow block before closing the fingers to guarantee a firm grasp on the first try.
+  Summary: align gripper centered over block; grasp firmly
+- **justification:** At approximately 3.5 seconds into the rollout, the robot closes its gripper on the yellow block but fails to secure it. The block slips out, leading to continuous failed grasping attempts throughout the remainder of the video.
diff --git a/blocks/ep016/labels.json b/blocks/ep016/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..d904caa86261af9d26c2dcba149e80ef4d3ad001
--- /dev/null
+++ b/blocks/ep016/labels.json
@@ -0,0 +1,142 @@
+{
+  "rollout_id": "t000__ep016__blocks",
+  "scene_variant": "",
+  "task_id": 0,
+  "task_file": "",
+  "instruction": "Stack the yellow block on the wooden block",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:02:29+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00990099009901,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "approaching yellow block",
+      "description": "approaching yellow block"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00990099009901,
+      "t_end_sec": 6.01980198019802,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "closing gripper and missing yellow block",
+      "description": "closing gripper and missing yellow block"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01980198019802,
+      "t_end_sec": 9.029702970297029,
+      "control_step_start": 304,
+      "control_step_end": 455,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "re-descending and attempting grasp",
+      "description": "re-descending and attempting grasp"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.029702970297029,
+      "t_end_sec": 12.03960396039604,
+      "control_step_start": 456,
+      "control_step_end": 607,
+      "policy_step_start": 76,
+      "policy_step_end": 101,
+      "phrase": "pinching yellow block and slipping",
+      "description": "pinching yellow block and slipping"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 12.03960396039604,
+      "t_end_sec": 15.049504950495049,
+      "control_step_start": 608,
+      "control_step_end": 759,
+      "policy_step_start": 101,
+      "policy_step_end": 126,
+      "phrase": "repeated grasp attempts on block",
+      "description": "repeated grasp attempts on block"
+    },
+    {
+      "segment_index": 5,
+      "t_start_sec": 15.049504950495049,
+      "t_end_sec": 18.059405940594058,
+      "control_step_start": 760,
+      "control_step_end": 911,
+      "policy_step_start": 126,
+      "policy_step_end": 151,
+      "phrase": "oscillating over yellow block",
+      "description": "oscillating over yellow block"
+    },
+    {
+      "segment_index": 6,
+      "t_start_sec": 18.059405940594058,
+      "t_end_sec": 18.455445544554454,
+      "control_step_start": 912,
+      "control_step_end": 931,
+      "policy_step_start": 152,
+      "policy_step_end": 155,
+      "phrase": "remaining stationary over yellow block",
+      "description": "remaining stationary over yellow block"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 3.504950495049505,
+    "vlm_failure_onset_timestamp": "00:03.5",
+    "vlm_failure_onset_frame": 177,
+    "vlm_failure_onset_step": 177,
+    "vlm_coarse_failure_onset_seconds": 3.5,
+    "vlm_failure_window_start_seconds": 3.0,
+    "vlm_failure_window_end_seconds": 4.5,
+    "vlm_temporal_refined": false,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the yellow block but failed to establish a secure grip, resulting in repeated failed grasp attempts and slipping without picking up the block.",
+    "vlm_failure_reason_summary": "failed grasp on yellow block; unable to lift block",
+    "vlm_recovery_action": "Re-open the gripper, center the end-effector directly over the yellow block, lower to surround the block, and close the gripper securely before lifting.",
+    "vlm_recovery_action_summary": "re-open gripper; center over yellow block; grasp securely",
+    "vlm_prevention_action": "Align the gripper accurately with the center and orientation of the yellow block before closing the fingers to guarantee a firm grasp on the first try.",
+    "vlm_prevention_action_summary": "align gripper centered over block; grasp firmly",
+    "vlm_justification": "At approximately 3.5 seconds into the rollout, the robot closes its gripper on the yellow block but fails to secure it. The block slips out, leading to continuous failed grasping attempts throughout the remainder of the video.",
+    "vlm_temporal_justification": "At 02:43, the robot gripper fully closes around the yellow block but fails to establish a secure grip, slipping past the block so that it remains on the table when the arm lifts.",
+    "vlm_video_fps": 50.5,
+    "vlm_video_num_frames": 932,
+    "vlm_video_duration_seconds": 18.435643564356436,
+    "vlm_usage": {
+      "input_tokens": 22215,
+      "output_tokens": 378,
+      "cached_tokens": 1225
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 3.5,\n \"failure_window_start_seconds\": 3.0,\n \"failure_window_end_seconds\": 4.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the yellow block but failed to establish a secure grip, resulting in repeated failed grasp attempts and slipping without picking up the block.\",\n \"failure_reason_summary\": \"failed grasp on yellow block; unable to lift block\",\n \"recovery_action\": \"Re-open the gripper, center the end-effector directly over the yellow block, lower to surround the block, and close the gripper securely before lifting.\",\n \"recovery_action_summary\": \"re-open gripper; center over yellow block; grasp securely\",\n \"prevention_action\": \"Align the gripper accurately with the center and orientation of the yellow block before closing the fingers to guarantee a firm grasp on the first try.\",\n \"prevention_action_summary\": \"align gripper centered over block; grasp firmly\",\n \"justification\": \"At approximately 3.5 seconds into the rollout, the robot closes its gripper on the yellow block but fails to secure it. The block slips out, leading to continuous failed grasping attempts throughout the remainder of the video.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 177,
+    "failure_sim_step": 177,
+    "failure_policy_step": 29,
+    "failure_chunk_index": 3,
+    "first_post_failure_policy_step": 30,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-open the gripper, center the end-effector directly over the yellow block, lower to surround the block, and close the gripper securely before lifting."
+  }
+}
\ No newline at end of file
diff --git a/blocks/ep016/labels.npz b/blocks/ep016/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..a62fbbc29b15756bc414a2c517deecdf70df4202
--- /dev/null
+++ b/blocks/ep016/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:11c5f6d9046a6855f59b30da01d3a058016ae197c72b2301181c1094dd24aafe
+size 4025
diff --git a/blocks/ep017/example.md b/blocks/ep017/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..f1b2e62bcf8525fd87e6d51f83a7d83417a2213b
--- /dev/null
+++ b/blocks/ep017/example.md
@@ -0,0 +1,34 @@
+# t000__ep017__blocks
+
+- **suite / variant:** `` / ``
+- **instruction:** Stack the yellow block on the wooden block
+- **success:** False
+- **control steps:** 475  (9.5s at 50 Hz)
+- **policy inferences:** 82  (replan_steps=6)
+- **video:** `blocks/ep017/rollout.mp4`
+- **features:** `82` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | reaching for yellow block |
+| 3.0–6.0 | 151–301 | 25–50 | grasping and placing yellow block |
+| 6.0–9.0 | 302–452 | 50–75 | hovering above wooden block |
+| 9.0–9.5 | 453–474 | 75–79 | idle near wooden block |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 5.199203187250996 s (00:05.2)
+- **frame / control_step:** 261 (policy_step=43, chunk_index=3)
+- **coarse onset:** 5.2 s (refined=False)
+- **mode:** placement_or_insertion_failure
+- **reason:** The robot successfully picked up the yellow block and moved it over the wooden block, but released it off-center, causing the yellow block to slide off and fall behind the wooden block.
+  Summary: improper block placement; yellow block fell off wooden block
+- **recovery:** Regrasp the fallen yellow block from the table, position it directly over the center of the wooden block, and place it carefully to stack it.
+  Summary: regrasp yellow block; stack onto wooden block
+- **prevention:** Align the yellow block directly over the center of the wooden block before releasing the gripper to ensure a stable stack.
+  Summary: align yellow block centered; place stably on wooden block
+- **justification:** At around 5.2 seconds (frame 00:05 to 00:06), the robot attempts to place the yellow block on the wooden block, but misaligns the placement, causing the yellow block to fall off and fail the goal predicate.
diff --git a/blocks/ep017/labels.json b/blocks/ep017/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..ae8385c275eeb34325c870464147c5e4d0a1253e
--- /dev/null
+++ b/blocks/ep017/labels.json
@@ -0,0 +1,109 @@
+{
+  "rollout_id": "t000__ep017__blocks",
+  "scene_variant": "",
+  "task_id": 0,
+  "task_file": "",
+  "instruction": "Stack the yellow block on the wooden block",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:03:01+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00796812749004,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for yellow block",
+      "description": "reaching for yellow block"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00796812749004,
+      "t_end_sec": 6.01593625498008,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "grasping and placing yellow block",
+      "description": "grasping and placing yellow block"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01593625498008,
+      "t_end_sec": 9.02390438247012,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "hovering above wooden block",
+      "description": "hovering above wooden block"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.02390438247012,
+      "t_end_sec": 9.46215139442231,
+      "control_step_start": 453,
+      "control_step_end": 474,
+      "policy_step_start": 75,
+      "policy_step_end": 79,
+      "phrase": "idle near wooden block",
+      "description": "idle near wooden block"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "placement_or_insertion_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 5.199203187250996,
+    "vlm_failure_onset_timestamp": "00:05.2",
+    "vlm_failure_onset_frame": 261,
+    "vlm_failure_onset_step": 261,
+    "vlm_coarse_failure_onset_seconds": 5.2,
+    "vlm_failure_window_start_seconds": 5.2,
+    "vlm_failure_window_end_seconds": 6.2,
+    "vlm_temporal_refined": false,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot successfully picked up the yellow block and moved it over the wooden block, but released it off-center, causing the yellow block to slide off and fall behind the wooden block.",
+    "vlm_failure_reason_summary": "improper block placement; yellow block fell off wooden block",
+    "vlm_recovery_action": "Regrasp the fallen yellow block from the table, position it directly over the center of the wooden block, and place it carefully to stack it.",
+    "vlm_recovery_action_summary": "regrasp yellow block; stack onto wooden block",
+    "vlm_prevention_action": "Align the yellow block directly over the center of the wooden block before releasing the gripper to ensure a stable stack.",
+    "vlm_prevention_action_summary": "align yellow block centered; place stably on wooden block",
+    "vlm_justification": "At around 5.2 seconds (frame 00:05 to 00:06), the robot attempts to place the yellow block on the wooden block, but misaligns the placement, causing the yellow block to fall off and fail the goal predicate.",
+    "vlm_temporal_justification": "At 03:17, the robot gripper opens to release the yellow block while it is positioned off-center on top of the wooden block, causing the yellow block to slip and fall off behind it.",
+    "vlm_video_fps": 50.2,
+    "vlm_video_num_frames": 475,
+    "vlm_video_duration_seconds": 9.44223107569721,
+    "vlm_usage": {
+      "input_tokens": 21533,
+      "output_tokens": 396,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"placement_or_insertion_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 5.2,\n \"failure_window_start_seconds\": 5.2,\n \"failure_window_end_seconds\": 6.2,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot successfully picked up the yellow block and moved it over the wooden block, but released it off-center, causing the yellow block to slide off and fall behind the wooden block.\",\n \"failure_reason_summary\": \"improper block placement; yellow block fell off wooden block\",\n \"recovery_action\": \"Regrasp the fallen yellow block from the table, position it directly over the center of the wooden block, and place it carefully to stack it.\",\n \"recovery_action_summary\": \"regrasp yellow block; stack onto wooden block\",\n \"prevention_action\": \"Align the yellow block directly over the center of the wooden block before releasing the gripper to ensure a stable stack.\",\n \"prevention_action_summary\": \"align yellow block centered; place stably on wooden block\",\n \"justification\": \"At around 5.2 seconds (frame 00:05 to 00:06), the robot attempts to place the yellow block on the wooden block, but misaligns the placement, causing the yellow block to fall off and fail the goal predicate.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 261,
+    "failure_sim_step": 261,
+    "failure_policy_step": 43,
+    "failure_chunk_index": 3,
+    "first_post_failure_policy_step": 44,
+    "failure_type": "placement_or_insertion_failure",
+    "correction_action": "Regrasp the fallen yellow block from the table, position it directly over the center of the wooden block, and place it carefully to stack it."
+  }
+}
\ No newline at end of file
diff --git a/blocks/ep017/labels.npz b/blocks/ep017/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..766a9528d9ca78adeaa5a6bd53ddbee206cd83a5
--- /dev/null
+++ b/blocks/ep017/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:01b8661df42ff82d0fcd8e8fd7361fc40c592786079781895bb5cee83281a8c4
+size 3870
diff --git a/blocks/ep018/example.md b/blocks/ep018/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..fbabb014d89a16c09822d3023aaf9fc9db0af7a3
--- /dev/null
+++ b/blocks/ep018/example.md
@@ -0,0 +1,34 @@
+# t000__ep018__blocks
+
+- **suite / variant:** `` / ``
+- **instruction:** Stack the yellow block on the wooden block
+- **success:** False
+- **control steps:** 571  (11.3s at 51 Hz)
+- **policy inferences:** 99  (replan_steps=6)
+- **video:** `blocks/ep018/rollout.mp4`
+- **features:** `99` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching toward yellow block |
+| 3.0–6.0 | 152–303 | 25–50 | grasping and lifting yellow block |
+| 6.0–9.0 | 304–455 | 50–75 | moving yellow block to target |
+| 9.0–11.3 | 456–570 | 76–95 | lowering and releasing block |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 8.67588932806324 s (00:08.7)
+- **frame / control_step:** 439 (policy_step=73, chunk_index=1)
+- **coarse onset:** 9.5 s (refined=True)
+- **mode:** placement_or_insertion_failure
+- **reason:** The robot successfully grasped the yellow block and brought it over the wooden block, but released it with poor alignment, causing the block to rest in an unstable tilted position.
+  Summary: misaligned placement; unstable block stacking
+- **recovery:** Re-grasp the yellow block, align it directly over the wooden block's top surface, and place it down cleanly.
+  Summary: re-grasp yellow block; align and stack on wooden block
+- **prevention:** Ensure the yellow block is centered directly above the wooden block before lowering and releasing the gripper.
+  Summary: center yellow block over wooden block; release carefully
+- **justification:** The robot successfully picks up the yellow block and moves it to the target location. However, around 9.5 seconds into the rollout (frames 00:09–00:10), the robot lowers and releases the yellow block without proper alignment, causing it to balance instability / sit tilted on top of the wooden block rather than being properly stacked.
diff --git a/blocks/ep018/labels.json b/blocks/ep018/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..c5d58f06770d9cb3db1f8cc5e9c1c37bf397bf03
--- /dev/null
+++ b/blocks/ep018/labels.json
@@ -0,0 +1,109 @@
+{
+  "rollout_id": "t000__ep018__blocks",
+  "scene_variant": "",
+  "task_id": 0,
+  "task_file": "",
+  "instruction": "Stack the yellow block on the wooden block",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:03:35+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0039525691699605,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching toward yellow block",
+      "description": "reaching toward yellow block"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0039525691699605,
+      "t_end_sec": 6.007905138339921,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "grasping and lifting yellow block",
+      "description": "grasping and lifting yellow block"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.007905138339921,
+      "t_end_sec": 9.011857707509881,
+      "control_step_start": 304,
+      "control_step_end": 455,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "moving yellow block to target",
+      "description": "moving yellow block to target"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.011857707509881,
+      "t_end_sec": 11.284584980237154,
+      "control_step_start": 456,
+      "control_step_end": 570,
+      "policy_step_start": 76,
+      "policy_step_end": 95,
+      "phrase": "lowering and releasing block",
+      "description": "lowering and releasing block"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "placement_or_insertion_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 8.67588932806324,
+    "vlm_failure_onset_timestamp": "00:08.7",
+    "vlm_failure_onset_frame": 439,
+    "vlm_failure_onset_step": 439,
+    "vlm_coarse_failure_onset_seconds": 9.5,
+    "vlm_failure_window_start_seconds": 9.0,
+    "vlm_failure_window_end_seconds": 10.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot successfully grasped the yellow block and brought it over the wooden block, but released it with poor alignment, causing the block to rest in an unstable tilted position.",
+    "vlm_failure_reason_summary": "misaligned placement; unstable block stacking",
+    "vlm_recovery_action": "Re-grasp the yellow block, align it directly over the wooden block's top surface, and place it down cleanly.",
+    "vlm_recovery_action_summary": "re-grasp yellow block; align and stack on wooden block",
+    "vlm_prevention_action": "Ensure the yellow block is centered directly above the wooden block before lowering and releasing the gripper.",
+    "vlm_prevention_action_summary": "center yellow block over wooden block; release carefully",
+    "vlm_justification": "The robot successfully picks up the yellow block and moves it to the target location. However, around 9.5 seconds into the rollout (frames 00:09\u201300:10), the robot lowers and releases the yellow block without proper alignment, causing it to balance instability / sit tilted on top of the wooden block rather than being properly stacked.",
+    "vlm_temporal_justification": "At 01:50 (110 seconds into the clip), the robot begins rotating its wrist away from proper alignment while positioning the yellow block above the wooden block, directly leading to the misaligned placement.",
+    "vlm_video_fps": 50.6,
+    "vlm_video_num_frames": 571,
+    "vlm_video_duration_seconds": 11.264822134387352,
+    "vlm_usage": {
+      "input_tokens": 17814,
+      "output_tokens": 400,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"placement_or_insertion_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 9.5,\n \"failure_window_start_seconds\": 9.0,\n \"failure_window_end_seconds\": 10.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot successfully grasped the yellow block and brought it over the wooden block, but released it with poor alignment, causing the block to rest in an unstable tilted position.\",\n \"failure_reason_summary\": \"misaligned placement; unstable block stacking\",\n \"recovery_action\": \"Re-grasp the yellow block, align it directly over the wooden block's top surface, and place it down cleanly.\",\n \"recovery_action_summary\": \"re-grasp yellow block; align and stack on wooden block\",\n \"prevention_action\": \"Ensure the yellow block is centered directly above the wooden block before lowering and releasing the gripper.\",\n \"prevention_action_summary\": \"center yellow block over wooden block; release carefully\",\n \"justification\": \"The robot successfully picks up the yellow block and moves it to the target location. However, around 9.5 seconds into the rollout (frames 00:09\u201300:10), the robot lowers and releases the yellow block without proper alignment, causing it to balance instability / sit tilted on top of the wooden block rather than being properly stacked.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 439,
+    "failure_sim_step": 439,
+    "failure_policy_step": 73,
+    "failure_chunk_index": 1,
+    "first_post_failure_policy_step": 74,
+    "failure_type": "placement_or_insertion_failure",
+    "correction_action": "Re-grasp the yellow block, align it directly over the wooden block's top surface, and place it down cleanly."
+  }
+}
\ No newline at end of file
diff --git a/blocks/ep018/labels.npz b/blocks/ep018/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..79e3f364f3450c9f5c54b0b20760c72c884f370c
--- /dev/null
+++ b/blocks/ep018/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2f5b36a3495a0bcc0dedf331c32c007ae45f8eaa3e559dfba17d7077ef9d01d4
+size 3866
diff --git a/blocks/ep019/example.md b/blocks/ep019/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..b86a5649e48577780da706ecf6406f08ece00926
--- /dev/null
+++ b/blocks/ep019/example.md
@@ -0,0 +1,25 @@
+# t000__ep019__blocks
+
+- **suite / variant:** `` / ``
+- **instruction:** Stack the yellow block on the wooden block
+- **success:** True
+- **control steps:** 964  (19.0s at 51 Hz)
+- **policy inferences:** 165  (replan_steps=6)
+- **video:** `blocks/ep019/rollout.mp4`
+- **features:** `165` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching for yellow block |
+| 3.0–6.0 | 152–303 | 25–50 | picking up yellow block |
+| 6.0–9.0 | 304–455 | 50–75 | moving yellow block to wooden block |
+| 9.0–12.0 | 456–607 | 76–101 | stacking yellow block on wooden block |
+| 12.0–15.0 | 608–759 | 101–126 | releasing yellow block |
+| 15.0–18.0 | 760–911 | 126–151 | retracting arm |
+| 18.0–19.0 | 912–963 | 152–160 | retracting arm |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/blocks/ep019/labels.json b/blocks/ep019/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..3b91a1daa36a18a21b93551463085239656f8d2a
--- /dev/null
+++ b/blocks/ep019/labels.json
@@ -0,0 +1,112 @@
+{
+  "rollout_id": "t000__ep019__blocks",
+  "scene_variant": "",
+  "task_id": 0,
+  "task_file": "",
+  "instruction": "Stack the yellow block on the wooden block",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:03:46+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9980276134122286,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for yellow block",
+      "description": "reaching for yellow block"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9980276134122286,
+      "t_end_sec": 5.996055226824457,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "picking up yellow block",
+      "description": "picking up yellow block"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.996055226824457,
+      "t_end_sec": 8.994082840236686,
+      "control_step_start": 304,
+      "control_step_end": 455,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "moving yellow block to wooden block",
+      "description": "moving yellow block to wooden block"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.994082840236686,
+      "t_end_sec": 11.992110453648914,
+      "control_step_start": 456,
+      "control_step_end": 607,
+      "policy_step_start": 76,
+      "policy_step_end": 101,
+      "phrase": "stacking yellow block on wooden block",
+      "description": "stacking yellow block on wooden block"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 11.992110453648914,
+      "t_end_sec": 14.990138067061142,
+      "control_step_start": 608,
+      "control_step_end": 759,
+      "policy_step_start": 101,
+      "policy_step_end": 126,
+      "phrase": "releasing yellow block",
+      "description": "releasing yellow block"
+    },
+    {
+      "segment_index": 5,
+      "t_start_sec": 14.990138067061142,
+      "t_end_sec": 17.988165680473372,
+      "control_step_start": 760,
+      "control_step_end": 911,
+      "policy_step_start": 126,
+      "policy_step_end": 151,
+      "phrase": "retracting arm",
+      "description": "retracting arm"
+    },
+    {
+      "segment_index": 6,
+      "t_start_sec": 17.988165680473372,
+      "t_end_sec": 19.013806706114398,
+      "control_step_start": 912,
+      "control_step_end": 963,
+      "policy_step_start": 152,
+      "policy_step_end": 160,
+      "phrase": "retracting arm",
+      "description": "retracting arm"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/blocks/ep019/labels.npz b/blocks/ep019/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..429668e9d17a495a42f7e2a1ac1a49a2193fcf2e
--- /dev/null
+++ b/blocks/ep019/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:97c86a5328f2c389296404348872f088154cc342b906dbc44e96f8f9f6f566eb
+size 3200
diff --git a/blocks/ep020/example.md b/blocks/ep020/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..9954bc0c3c5c049feec8042d81f87378cb4bc494
--- /dev/null
+++ b/blocks/ep020/example.md
@@ -0,0 +1,33 @@
+# t000__ep020__blocks
+
+- **suite / variant:** `` / ``
+- **instruction:** Stack the yellow block on the wooden block
+- **success:** True
+- **control steps:** 2153  (42.7s at 50 Hz)
+- **policy inferences:** 371  (replan_steps=6)
+- **video:** `blocks/ep020/rollout.mp4`
+- **features:** `371` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | reaching for yellow block |
+| 3.0–6.0 | 151–301 | 25–50 | grasping and moving yellow block |
+| 6.0–9.0 | 302–452 | 50–75 | placing yellow block on wooden block |
+| 9.0–12.0 | 453–603 | 75–100 | holding yellow block in place |
+| 12.0–15.0 | 604–754 | 100–125 | holding yellow block in place |
+| 15.0–18.0 | 755–905 | 125–150 | holding yellow block in place |
+| 18.0–21.0 | 906–1056 | 151–176 | holding yellow block in place |
+| 21.0–24.0 | 1057–1207 | 176–201 | holding yellow block in place |
+| 24.0–27.0 | 1208–1358 | 201–226 | holding yellow block in place |
+| 27.0–30.0 | 1359–1509 | 226–251 | holding yellow block in place |
+| 30.0–33.0 | 1510–1660 | 251–276 | holding yellow block in place |
+| 33.0–36.0 | 1661–1811 | 276–301 | holding yellow block in place |
+| 36.0–38.9 | 1812–1962 | 302–327 | holding yellow block in place |
+| 38.9–41.9 | 1963–2113 | 327–352 | releasing block and raising gripper |
+| 41.9–42.7 | 2114–2152 | 352–358 | retracting arm to home position |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/blocks/ep020/labels.json b/blocks/ep020/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..de63f08dcc5966919d19386df1b40364363acda1
--- /dev/null
+++ b/blocks/ep020/labels.json
@@ -0,0 +1,200 @@
+{
+  "rollout_id": "t000__ep020__blocks",
+  "scene_variant": "",
+  "task_id": 0,
+  "task_file": "",
+  "instruction": "Stack the yellow block on the wooden block",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:04:00+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.996031746031746,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for yellow block",
+      "description": "reaching for yellow block"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.996031746031746,
+      "t_end_sec": 5.992063492063492,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "grasping and moving yellow block",
+      "description": "grasping and moving yellow block"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.992063492063492,
+      "t_end_sec": 8.988095238095239,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "placing yellow block on wooden block",
+      "description": "placing yellow block on wooden block"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.988095238095239,
+      "t_end_sec": 11.984126984126984,
+      "control_step_start": 453,
+      "control_step_end": 603,
+      "policy_step_start": 75,
+      "policy_step_end": 100,
+      "phrase": "holding yellow block in place",
+      "description": "holding yellow block in place"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 11.984126984126984,
+      "t_end_sec": 14.980158730158731,
+      "control_step_start": 604,
+      "control_step_end": 754,
+      "policy_step_start": 100,
+      "policy_step_end": 125,
+      "phrase": "holding yellow block in place",
+      "description": "holding yellow block in place"
+    },
+    {
+      "segment_index": 5,
+      "t_start_sec": 14.980158730158731,
+      "t_end_sec": 17.976190476190478,
+      "control_step_start": 755,
+      "control_step_end": 905,
+      "policy_step_start": 125,
+      "policy_step_end": 150,
+      "phrase": "holding yellow block in place",
+      "description": "holding yellow block in place"
+    },
+    {
+      "segment_index": 6,
+      "t_start_sec": 17.976190476190478,
+      "t_end_sec": 20.97222222222222,
+      "control_step_start": 906,
+      "control_step_end": 1056,
+      "policy_step_start": 151,
+      "policy_step_end": 176,
+      "phrase": "holding yellow block in place",
+      "description": "holding yellow block in place"
+    },
+    {
+      "segment_index": 7,
+      "t_start_sec": 20.97222222222222,
+      "t_end_sec": 23.96825396825397,
+      "control_step_start": 1057,
+      "control_step_end": 1207,
+      "policy_step_start": 176,
+      "policy_step_end": 201,
+      "phrase": "holding yellow block in place",
+      "description": "holding yellow block in place"
+    },
+    {
+      "segment_index": 8,
+      "t_start_sec": 23.96825396825397,
+      "t_end_sec": 26.964285714285715,
+      "control_step_start": 1208,
+      "control_step_end": 1358,
+      "policy_step_start": 201,
+      "policy_step_end": 226,
+      "phrase": "holding yellow block in place",
+      "description": "holding yellow block in place"
+    },
+    {
+      "segment_index": 9,
+      "t_start_sec": 26.964285714285715,
+      "t_end_sec": 29.960317460317462,
+      "control_step_start": 1359,
+      "control_step_end": 1509,
+      "policy_step_start": 226,
+      "policy_step_end": 251,
+      "phrase": "holding yellow block in place",
+      "description": "holding yellow block in place"
+    },
+    {
+      "segment_index": 10,
+      "t_start_sec": 29.960317460317462,
+      "t_end_sec": 32.95634920634921,
+      "control_step_start": 1510,
+      "control_step_end": 1660,
+      "policy_step_start": 251,
+      "policy_step_end": 276,
+      "phrase": "holding yellow block in place",
+      "description": "holding yellow block in place"
+    },
+    {
+      "segment_index": 11,
+      "t_start_sec": 32.95634920634921,
+      "t_end_sec": 35.952380952380956,
+      "control_step_start": 1661,
+      "control_step_end": 1811,
+      "policy_step_start": 276,
+      "policy_step_end": 301,
+      "phrase": "holding yellow block in place",
+      "description": "holding yellow block in place"
+    },
+    {
+      "segment_index": 12,
+      "t_start_sec": 35.952380952380956,
+      "t_end_sec": 38.9484126984127,
+      "control_step_start": 1812,
+      "control_step_end": 1962,
+      "policy_step_start": 302,
+      "policy_step_end": 327,
+      "phrase": "holding yellow block in place",
+      "description": "holding yellow block in place"
+    },
+    {
+      "segment_index": 13,
+      "t_start_sec": 38.9484126984127,
+      "t_end_sec": 41.94444444444444,
+      "control_step_start": 1963,
+      "control_step_end": 2113,
+      "policy_step_start": 327,
+      "policy_step_end": 352,
+      "phrase": "releasing block and raising gripper",
+      "description": "releasing block and raising gripper"
+    },
+    {
+      "segment_index": 14,
+      "t_start_sec": 41.94444444444444,
+      "t_end_sec": 42.71825396825397,
+      "control_step_start": 2114,
+      "control_step_end": 2152,
+      "policy_step_start": 352,
+      "policy_step_end": 358,
+      "phrase": "retracting arm to home position",
+      "description": "retracting arm to home position"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/blocks/ep020/labels.npz b/blocks/ep020/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..0d61b81f8ea5c633c2b4d37453aa4022796580da
--- /dev/null
+++ b/blocks/ep020/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4443fce6a0044dd1bc067859d2c76c9d0f3749e3dd659ea292113210b8d790ec
+size 3404
diff --git a/blocks/ep021/example.md b/blocks/ep021/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..2774a99c55f8f6f38eb21a02b5e2496e204cd632
--- /dev/null
+++ b/blocks/ep021/example.md
@@ -0,0 +1,34 @@
+# t000__ep021__blocks
+
+- **suite / variant:** `` / ``
+- **instruction:** Stack the yellow block on the wooden block
+- **success:** False
+- **control steps:** 505  (10.0s at 50 Hz)
+- **policy inferences:** 88  (replan_steps=6)
+- **video:** `blocks/ep021/rollout.mp4`
+- **features:** `88` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | approaching yellow block |
+| 3.0–6.0 | 152–303 | 25–50 | closing gripper on yellow block |
+| 6.0–9.0 | 304–455 | 50–75 | lifting empty gripper |
+| 9.0–10.0 | 456–504 | 76–84 | hovering above workspace |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 5.069306930693069 s (00:05.1)
+- **frame / control_step:** 256 (policy_step=42, chunk_index=4)
+- **coarse onset:** 5.5 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot gripper attempted to grasp the yellow block, but closed loosely/improperly and slipped off, lifting away without holding the block.
+  Summary: failed grasp on yellow block; gripper slipped
+- **recovery:** Lower the gripper back over the yellow block, close firmly around it to secure a grasp, and then stack it onto the wooden block.
+  Summary: re-align gripper; regrasp yellow block; stack on wooden block
+- **prevention:** Center the gripper precisely over the yellow block and ensure the fingers close completely before initiating the lift.
+  Summary: align gripper centered on yellow block; secure grasp before lifting
+- **justification:** None
diff --git a/blocks/ep021/labels.json b/blocks/ep021/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..f086ca8aa7519fa57890b507425572c3369ae09d
--- /dev/null
+++ b/blocks/ep021/labels.json
@@ -0,0 +1,109 @@
+{
+  "rollout_id": "t000__ep021__blocks",
+  "scene_variant": "",
+  "task_id": 0,
+  "task_file": "",
+  "instruction": "Stack the yellow block on the wooden block",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:04:35+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00990099009901,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "approaching yellow block",
+      "description": "approaching yellow block"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00990099009901,
+      "t_end_sec": 6.01980198019802,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "closing gripper on yellow block",
+      "description": "closing gripper on yellow block"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01980198019802,
+      "t_end_sec": 9.029702970297029,
+      "control_step_start": 304,
+      "control_step_end": 455,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "lifting empty gripper",
+      "description": "lifting empty gripper"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.029702970297029,
+      "t_end_sec": 10.0,
+      "control_step_start": 456,
+      "control_step_end": 504,
+      "policy_step_start": 76,
+      "policy_step_end": 84,
+      "phrase": "hovering above workspace",
+      "description": "hovering above workspace"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 5.069306930693069,
+    "vlm_failure_onset_timestamp": "00:05.1",
+    "vlm_failure_onset_frame": 256,
+    "vlm_failure_onset_step": 256,
+    "vlm_coarse_failure_onset_seconds": 5.5,
+    "vlm_failure_window_start_seconds": 5.0,
+    "vlm_failure_window_end_seconds": 6.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot gripper attempted to grasp the yellow block, but closed loosely/improperly and slipped off, lifting away without holding the block.",
+    "vlm_failure_reason_summary": "failed grasp on yellow block; gripper slipped",
+    "vlm_recovery_action": "Lower the gripper back over the yellow block, close firmly around it to secure a grasp, and then stack it onto the wooden block.",
+    "vlm_recovery_action_summary": "re-align gripper; regrasp yellow block; stack on wooden block",
+    "vlm_prevention_action": "Center the gripper precisely over the yellow block and ensure the fingers close completely before initiating the lift.",
+    "vlm_prevention_action_summary": "align gripper centered on yellow block; secure grasp before lifting",
+    "vlm_justification": null,
+    "vlm_temporal_justification": "At timestamp 02:10 (130 seconds into the clip), the robot gripper begins lifting upward after attempting to grasp the yellow block, but the block remains entirely stationary on the table surface, making the grasp failure visibly apparent.",
+    "vlm_video_fps": 50.5,
+    "vlm_video_num_frames": 505,
+    "vlm_video_duration_seconds": 9.98019801980198,
+    "vlm_usage": {
+      "input_tokens": 21709,
+      "output_tokens": 322,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 5.5,\n \"failure_window_start_seconds\": 5.0,\n \"failure_window_end_seconds\": 6.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot gripper attempted to grasp the yellow block, but closed loosely/improperly and slipped off, lifting away without holding the block.\",\n \"failure_reason_summary\": \"failed grasp on yellow block; gripper slipped\",\n \"recovery_action\": \"Lower the gripper back over the yellow block, close firmly around it to secure a grasp, and then stack it onto the wooden block.\",\n \"recovery_action_summary\": \"re-align gripper; regrasp yellow block; stack on wooden block\",\n \"prevention_action\": \"Center the gripper precisely over the yellow block and ensure the fingers close completely before initiating the lift.\",\n \"prevention_action_summary\": \"align gripper centered on yellow block; secure grasp before lifting\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 256,
+    "failure_sim_step": 256,
+    "failure_policy_step": 42,
+    "failure_chunk_index": 4,
+    "first_post_failure_policy_step": 43,
+    "failure_type": "grasp_failure",
+    "correction_action": "Lower the gripper back over the yellow block, close firmly around it to secure a grasp, and then stack it onto the wooden block."
+  }
+}
\ No newline at end of file
diff --git a/blocks/ep021/labels.npz b/blocks/ep021/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..13378c0bd21945f78e31ca505f982e8f21ffff94
--- /dev/null
+++ b/blocks/ep021/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4a8e17368ebb4c56a37e75a484b71bd639e583329291a4b806182af99ce9f8a4
+size 3845
diff --git a/blocks/ep022/example.md b/blocks/ep022/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..61b8b5171843c1b9cfdfca784f23265cedc8150b
--- /dev/null
+++ b/blocks/ep022/example.md
@@ -0,0 +1,34 @@
+# t000__ep022__blocks
+
+- **suite / variant:** `` / ``
+- **instruction:** Stack the yellow block on the wooden block
+- **success:** False
+- **control steps:** 563  (11.1s at 51 Hz)
+- **policy inferences:** 96  (replan_steps=6)
+- **video:** `blocks/ep022/rollout.mp4`
+- **features:** `96` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | lowering arm toward yellow block |
+| 3.0–6.0 | 152–303 | 25–50 | attempting grasp on yellow block |
+| 6.0–9.0 | 304–455 | 50–75 | moving empty gripper toward wooden block |
+| 9.0–11.1 | 456–562 | 76–93 | lowering empty gripper onto wooden block |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 5.0 s (00:05.0)
+- **frame / control_step:** 253 (policy_step=42, chunk_index=1)
+- **coarse onset:** 5.0 s (refined=False)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the yellow block, but failed to secure it in its gripper before lifting up and moving to the wooden block empty-handed.
+  Summary: failed grasp on yellow block; empty gripper movement
+- **recovery:** Lower the end-effector back to the yellow block, realign the gripper fingers around it, establish a firm grasp, and stack it onto the wooden block.
+  Summary: re-align gripper; grasp yellow block; stack on wooden block
+- **prevention:** Align the gripper properly with the center of the yellow block and close the fingers securely before initiating the lift motion.
+  Summary: align gripper over yellow block; ensure firm grasp
+- **justification:** At around 5.0 seconds, the gripper closes around the yellow block but fails to lift it. The robot continues its motion as if it had picked up the object, carrying an empty gripper over to the wooden block.
diff --git a/blocks/ep022/labels.json b/blocks/ep022/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..fe0d32fad49081a6c34859580d1f4103a77e39fc
--- /dev/null
+++ b/blocks/ep022/labels.json
@@ -0,0 +1,109 @@
+{
+  "rollout_id": "t000__ep022__blocks",
+  "scene_variant": "",
+  "task_id": 0,
+  "task_file": "",
+  "instruction": "Stack the yellow block on the wooden block",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:05:06+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0039525691699605,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "lowering arm toward yellow block",
+      "description": "lowering arm toward yellow block"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0039525691699605,
+      "t_end_sec": 6.007905138339921,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "attempting grasp on yellow block",
+      "description": "attempting grasp on yellow block"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.007905138339921,
+      "t_end_sec": 9.011857707509881,
+      "control_step_start": 304,
+      "control_step_end": 455,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "moving empty gripper toward wooden block",
+      "description": "moving empty gripper toward wooden block"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.011857707509881,
+      "t_end_sec": 11.126482213438734,
+      "control_step_start": 456,
+      "control_step_end": 562,
+      "policy_step_start": 76,
+      "policy_step_end": 93,
+      "phrase": "lowering empty gripper onto wooden block",
+      "description": "lowering empty gripper onto wooden block"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 5.0,
+    "vlm_failure_onset_timestamp": "00:05.0",
+    "vlm_failure_onset_frame": 253,
+    "vlm_failure_onset_step": 253,
+    "vlm_coarse_failure_onset_seconds": 5.0,
+    "vlm_failure_window_start_seconds": 4.5,
+    "vlm_failure_window_end_seconds": 6.0,
+    "vlm_temporal_refined": false,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the yellow block, but failed to secure it in its gripper before lifting up and moving to the wooden block empty-handed.",
+    "vlm_failure_reason_summary": "failed grasp on yellow block; empty gripper movement",
+    "vlm_recovery_action": "Lower the end-effector back to the yellow block, realign the gripper fingers around it, establish a firm grasp, and stack it onto the wooden block.",
+    "vlm_recovery_action_summary": "re-align gripper; grasp yellow block; stack on wooden block",
+    "vlm_prevention_action": "Align the gripper properly with the center of the yellow block and close the fingers securely before initiating the lift motion.",
+    "vlm_prevention_action_summary": "align gripper over yellow block; ensure firm grasp",
+    "vlm_justification": "At around 5.0 seconds, the gripper closes around the yellow block but fails to lift it. The robot continues its motion as if it had picked up the object, carrying an empty gripper over to the wooden block.",
+    "vlm_temporal_justification": "The gripper fingers finish closing at 03:12, touching each other completely without securing the yellow block, demonstrating a failed grasp.",
+    "vlm_video_fps": 50.6,
+    "vlm_video_num_frames": 563,
+    "vlm_video_duration_seconds": 11.106719367588932,
+    "vlm_usage": {
+      "input_tokens": 21776,
+      "output_tokens": 363,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 5.0,\n \"failure_window_start_seconds\": 4.5,\n \"failure_window_end_seconds\": 6.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the yellow block, but failed to secure it in its gripper before lifting up and moving to the wooden block empty-handed.\",\n \"failure_reason_summary\": \"failed grasp on yellow block; empty gripper movement\",\n \"recovery_action\": \"Lower the end-effector back to the yellow block, realign the gripper fingers around it, establish a firm grasp, and stack it onto the wooden block.\",\n \"recovery_action_summary\": \"re-align gripper; grasp yellow block; stack on wooden block\",\n \"prevention_action\": \"Align the gripper properly with the center of the yellow block and close the fingers securely before initiating the lift motion.\",\n \"prevention_action_summary\": \"align gripper over yellow block; ensure firm grasp\",\n \"justification\": \"At around 5.0 seconds, the gripper closes around the yellow block but fails to lift it. The robot continues its motion as if it had picked up the object, carrying an empty gripper over to the wooden block.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 253,
+    "failure_sim_step": 253,
+    "failure_policy_step": 42,
+    "failure_chunk_index": 1,
+    "first_post_failure_policy_step": 43,
+    "failure_type": "grasp_failure",
+    "correction_action": "Lower the end-effector back to the yellow block, realign the gripper fingers around it, establish a firm grasp, and stack it onto the wooden block."
+  }
+}
\ No newline at end of file
diff --git a/blocks/ep022/labels.npz b/blocks/ep022/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..a15367b275d56ce19570424a487c80c92d78b95c
--- /dev/null
+++ b/blocks/ep022/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6e68d46d3b70399d03251c0b590648c1e2a523cd912ddd09d51a2df33788e2cd
+size 3883
diff --git a/blocks/ep023/example.md b/blocks/ep023/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..3f979eb659c3de50f2807d4d47446815fee311be
--- /dev/null
+++ b/blocks/ep023/example.md
@@ -0,0 +1,34 @@
+# t000__ep023__blocks
+
+- **suite / variant:** `` / ``
+- **instruction:** Stack the yellow block on the wooden block
+- **success:** False
+- **control steps:** 511  (10.1s at 50 Hz)
+- **policy inferences:** 88  (replan_steps=6)
+- **video:** `blocks/ep023/rollout.mp4`
+- **features:** `88` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | lowering gripper toward yellow block |
+| 3.0–6.0 | 151–301 | 25–50 | grasping yellow block and lifting |
+| 6.0–9.0 | 302–452 | 50–75 | placing yellow block with wrist rotation |
+| 9.0–10.1 | 453–510 | 75–85 | holding position over blocks |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 6.011904761904762 s (00:06.0)
+- **frame / control_step:** 303 (policy_step=50, chunk_index=3)
+- **coarse onset:** 7.8 s (refined=True)
+- **mode:** placement_or_insertion_failure
+- **reason:** The robot successfully picked up the yellow block and positioned it over the wooden block, but rotated its end-effector awkwardly during placement, failing to release the block cleanly and leaving it misaligned.
+  Summary: improper release; misaligned block placement
+- **recovery:** Regrasp the yellow block, align it directly over the center of the wooden block, and release it flat without rotating the wrist.
+  Summary: regrasp yellow block; stack flat on wooden block
+- **prevention:** Maintain a upright, stable end-effector orientation when lowering the yellow block onto the wooden block, and fully open the gripper to release.
+  Summary: maintain vertical orientation; release gripper cleanly
+- **justification:** Around timestamp 7.8s (frames 00:07-00:08), the robot reaches the wooden block but twists its wrist significantly while placing the yellow block instead of opening the gripper smoothly, causing an unstable and failed placement.
diff --git a/blocks/ep023/labels.json b/blocks/ep023/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..5cb565d8c01fa45708207b519c28f56339b85e8e
--- /dev/null
+++ b/blocks/ep023/labels.json
@@ -0,0 +1,109 @@
+{
+  "rollout_id": "t000__ep023__blocks",
+  "scene_variant": "",
+  "task_id": 0,
+  "task_file": "",
+  "instruction": "Stack the yellow block on the wooden block",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:05:40+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.996031746031746,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "lowering gripper toward yellow block",
+      "description": "lowering gripper toward yellow block"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.996031746031746,
+      "t_end_sec": 5.992063492063492,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "grasping yellow block and lifting",
+      "description": "grasping yellow block and lifting"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.992063492063492,
+      "t_end_sec": 8.988095238095239,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "placing yellow block with wrist rotation",
+      "description": "placing yellow block with wrist rotation"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.988095238095239,
+      "t_end_sec": 10.13888888888889,
+      "control_step_start": 453,
+      "control_step_end": 510,
+      "policy_step_start": 75,
+      "policy_step_end": 85,
+      "phrase": "holding position over blocks",
+      "description": "holding position over blocks"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "placement_or_insertion_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 6.011904761904762,
+    "vlm_failure_onset_timestamp": "00:06.0",
+    "vlm_failure_onset_frame": 303,
+    "vlm_failure_onset_step": 303,
+    "vlm_coarse_failure_onset_seconds": 7.8,
+    "vlm_failure_window_start_seconds": 7.3,
+    "vlm_failure_window_end_seconds": 8.3,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot successfully picked up the yellow block and positioned it over the wooden block, but rotated its end-effector awkwardly during placement, failing to release the block cleanly and leaving it misaligned.",
+    "vlm_failure_reason_summary": "improper release; misaligned block placement",
+    "vlm_recovery_action": "Regrasp the yellow block, align it directly over the center of the wooden block, and release it flat without rotating the wrist.",
+    "vlm_recovery_action_summary": "regrasp yellow block; stack flat on wooden block",
+    "vlm_prevention_action": "Maintain a upright, stable end-effector orientation when lowering the yellow block onto the wooden block, and fully open the gripper to release.",
+    "vlm_prevention_action_summary": "maintain vertical orientation; release gripper cleanly",
+    "vlm_justification": "Around timestamp 7.8s (frames 00:07-00:08), the robot reaches the wooden block but twists its wrist significantly while placing the yellow block instead of opening the gripper smoothly, causing an unstable and failed placement.",
+    "vlm_temporal_justification": "At second 61 (01:01 in the video), as the robot lowers the yellow block onto the wooden block, its end-effector/wrist begins an awkward rotation away from the proper vertical alignment, leading directly to the failed placement.",
+    "vlm_video_fps": 50.4,
+    "vlm_video_num_frames": 511,
+    "vlm_video_duration_seconds": 10.119047619047619,
+    "vlm_usage": {
+      "input_tokens": 19457,
+      "output_tokens": 395,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"placement_or_insertion_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 7.8,\n \"failure_window_start_seconds\": 7.3,\n \"failure_window_end_seconds\": 8.3,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot successfully picked up the yellow block and positioned it over the wooden block, but rotated its end-effector awkwardly during placement, failing to release the block cleanly and leaving it misaligned.\",\n \"failure_reason_summary\": \"improper release; misaligned block placement\",\n \"recovery_action\": \"Regrasp the yellow block, align it directly over the center of the wooden block, and release it flat without rotating the wrist.\",\n \"recovery_action_summary\": \"regrasp yellow block; stack flat on wooden block\",\n \"prevention_action\": \"Maintain a upright, stable end-effector orientation when lowering the yellow block onto the wooden block, and fully open the gripper to release.\",\n \"prevention_action_summary\": \"maintain vertical orientation; release gripper cleanly\",\n \"justification\": \"Around timestamp 7.8s (frames 00:07-00:08), the robot reaches the wooden block but twists its wrist significantly while placing the yellow block instead of opening the gripper smoothly, causing an unstable and failed placement.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 303,
+    "failure_sim_step": 303,
+    "failure_policy_step": 50,
+    "failure_chunk_index": 3,
+    "first_post_failure_policy_step": 51,
+    "failure_type": "placement_or_insertion_failure",
+    "correction_action": "Regrasp the yellow block, align it directly over the center of the wooden block, and release it flat without rotating the wrist."
+  }
+}
\ No newline at end of file
diff --git a/blocks/ep023/labels.npz b/blocks/ep023/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..e1c8341a2ac626255c88b56bfd690b3e56a2b461
--- /dev/null
+++ b/blocks/ep023/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8ba3a9e5dabe0fb6b215be818614cd2c4d4797b8922abcd8f54a9ddb10d3d132
+size 3956
diff --git a/blocks/ep024/example.md b/blocks/ep024/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..d5716ff9a243b9091a2dd8edd65de290dd8e9a29
--- /dev/null
+++ b/blocks/ep024/example.md
@@ -0,0 +1,34 @@
+# t000__ep024__blocks
+
+- **suite / variant:** `` / ``
+- **instruction:** Stack the yellow block on the wooden block
+- **success:** False
+- **control steps:** 481  (9.5s at 50 Hz)
+- **policy inferences:** 82  (replan_steps=6)
+- **video:** `blocks/ep024/rollout.mp4`
+- **features:** `82` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | lowering arm toward yellow block |
+| 3.0–6.0 | 152–303 | 25–50 | attempting grasp on yellow block |
+| 6.0–9.0 | 304–455 | 50–75 | lifting empty gripper |
+| 9.0–9.5 | 456–480 | 76–80 | hovering above table |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 6.2772277227722775 s (00:06.3)
+- **frame / control_step:** 317 (policy_step=52, chunk_index=5)
+- **coarse onset:** 5.0 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the yellow block, but failed to secure a firm hold, causing the block to slip out of the gripper as it closed.
+  Summary: failed grasp on yellow block; object slipped from gripper
+- **recovery:** Re-align the end-effector over the yellow block, close the gripper to grasp it securely, and then place it onto the wooden block.
+  Summary: re-align gripper; regrasp yellow block; stack on wooden block
+- **prevention:** Position the gripper directly over the center of the yellow block before closing it to ensure a stable grasp prior to lifting.
+  Summary: align gripper centered on yellow block; secure grasp before lifting
+- **justification:** At 5.0 seconds, the robot closes its gripper on the yellow block, but fails to pinch it securely. By 6.0 seconds, the block slips out and remains on the table while the empty gripper lifts away, rendering the task uncompleted.
diff --git a/blocks/ep024/labels.json b/blocks/ep024/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..c6e034e6622b460a46acc52e16c8700046207f60
--- /dev/null
+++ b/blocks/ep024/labels.json
@@ -0,0 +1,109 @@
+{
+  "rollout_id": "t000__ep024__blocks",
+  "scene_variant": "",
+  "task_id": 0,
+  "task_file": "",
+  "instruction": "Stack the yellow block on the wooden block",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:06:18+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00990099009901,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "lowering arm toward yellow block",
+      "description": "lowering arm toward yellow block"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00990099009901,
+      "t_end_sec": 6.01980198019802,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "attempting grasp on yellow block",
+      "description": "attempting grasp on yellow block"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01980198019802,
+      "t_end_sec": 9.029702970297029,
+      "control_step_start": 304,
+      "control_step_end": 455,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "lifting empty gripper",
+      "description": "lifting empty gripper"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.029702970297029,
+      "t_end_sec": 9.524752475247524,
+      "control_step_start": 456,
+      "control_step_end": 480,
+      "policy_step_start": 76,
+      "policy_step_end": 80,
+      "phrase": "hovering above table",
+      "description": "hovering above table"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 6.2772277227722775,
+    "vlm_failure_onset_timestamp": "00:06.3",
+    "vlm_failure_onset_frame": 317,
+    "vlm_failure_onset_step": 317,
+    "vlm_coarse_failure_onset_seconds": 5.0,
+    "vlm_failure_window_start_seconds": 4.5,
+    "vlm_failure_window_end_seconds": 6.0,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the yellow block, but failed to secure a firm hold, causing the block to slip out of the gripper as it closed.",
+    "vlm_failure_reason_summary": "failed grasp on yellow block; object slipped from gripper",
+    "vlm_recovery_action": "Re-align the end-effector over the yellow block, close the gripper to grasp it securely, and then place it onto the wooden block.",
+    "vlm_recovery_action_summary": "re-align gripper; regrasp yellow block; stack on wooden block",
+    "vlm_prevention_action": "Position the gripper directly over the center of the yellow block before closing it to ensure a stable grasp prior to lifting.",
+    "vlm_prevention_action_summary": "align gripper centered on yellow block; secure grasp before lifting",
+    "vlm_justification": "At 5.0 seconds, the robot closes its gripper on the yellow block, but fails to pinch it securely. By 6.0 seconds, the block slips out and remains on the table while the empty gripper lifts away, rendering the task uncompleted.",
+    "vlm_temporal_justification": "At timestamp 03:37 (217 seconds into the video clip), the gripper fingers close past the edges of the yellow block instead of securing it, causing the block to slip out from the grasp as the gripper closes completely.",
+    "vlm_video_fps": 50.5,
+    "vlm_video_num_frames": 481,
+    "vlm_video_duration_seconds": 9.504950495049505,
+    "vlm_usage": {
+      "input_tokens": 21711,
+      "output_tokens": 392,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 5.0,\n \"failure_window_start_seconds\": 4.5,\n \"failure_window_end_seconds\": 6.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the yellow block, but failed to secure a firm hold, causing the block to slip out of the gripper as it closed.\",\n \"failure_reason_summary\": \"failed grasp on yellow block; object slipped from gripper\",\n \"recovery_action\": \"Re-align the end-effector over the yellow block, close the gripper to grasp it securely, and then place it onto the wooden block.\",\n \"recovery_action_summary\": \"re-align gripper; regrasp yellow block; stack on wooden block\",\n \"prevention_action\": \"Position the gripper directly over the center of the yellow block before closing it to ensure a stable grasp prior to lifting.\",\n \"prevention_action_summary\": \"align gripper centered on yellow block; secure grasp before lifting\",\n \"justification\": \"At 5.0 seconds, the robot closes its gripper on the yellow block, but fails to pinch it securely. By 6.0 seconds, the block slips out and remains on the table while the empty gripper lifts away, rendering the task uncompleted.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 317,
+    "failure_sim_step": 317,
+    "failure_policy_step": 52,
+    "failure_chunk_index": 5,
+    "first_post_failure_policy_step": 53,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-align the end-effector over the yellow block, close the gripper to grasp it securely, and then place it onto the wooden block."
+  }
+}
\ No newline at end of file
diff --git a/blocks/ep024/labels.npz b/blocks/ep024/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..84c29bccd0ffee068b023c7c26591ad6cb6b8f37
--- /dev/null
+++ b/blocks/ep024/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fbdc10eaaa2c219d0d21effde39dfe4dc72101a6c497fa16bff482676cebe0f0
+size 3869
diff --git a/blocks/ep025/example.md b/blocks/ep025/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..5b2339f0d035a049416d76e5f90a597c644f2593
--- /dev/null
+++ b/blocks/ep025/example.md
@@ -0,0 +1,37 @@
+# t000__ep025__blocks
+
+- **suite / variant:** `` / ``
+- **instruction:** Stack the yellow block on the wooden block
+- **success:** False
+- **control steps:** 1029  (20.5s at 50 Hz)
+- **policy inferences:** 178  (replan_steps=6)
+- **video:** `blocks/ep025/rollout.mp4`
+- **features:** `178` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | approaching yellow block |
+| 3.0–6.0 | 151–301 | 25–50 | grasping and moving yellow block |
+| 6.0–9.0 | 302–452 | 50–75 | positioning block on wooden block |
+| 9.0–12.0 | 453–603 | 75–100 | holding yellow block on target |
+| 12.0–15.0 | 604–754 | 100–125 | remaining stationary over block |
+| 15.0–18.0 | 755–905 | 125–150 | idle with gripper closed |
+| 18.0–20.5 | 906–1028 | 151–171 | idle with gripper closed |
+
+## Failure
+
+- **yes**
+- **onset type:** operator_intervention
+- **onset:** 6.242544731610338 s (00:06.2)
+- **frame / control_step:** 314 (policy_step=52, chunk_index=2)
+- **coarse onset:** 7.0 s (refined=True)
+- **mode:** stuck_or_no_progress
+- **reason:** After placing the yellow block on top of the wooden block, the robot failed to open its gripper to release the object and remained stuck in position until time expired.
+  Summary: stuck on target; failure to release gripper
+- **recovery:** Open the gripper to release the yellow block onto the wooden block, then raise the arm away from the stack.
+  Summary: open gripper; retract arm
+- **prevention:** Command the gripper to open immediately once stable contact with the wooden block is made, then retract the arm.
+  Summary: release block on contact; retract arm
+- **justification:** The robot successfully picked up the yellow block and aligned it on top of the wooden block by around 00:06. However, at around 00:07, it stopped making progress and remained frozen with its gripper holding the block until the end of the episode.
diff --git a/blocks/ep025/labels.json b/blocks/ep025/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..06333de8308e93e06630493564caa491f6108c96
--- /dev/null
+++ b/blocks/ep025/labels.json
@@ -0,0 +1,142 @@
+{
+  "rollout_id": "t000__ep025__blocks",
+  "scene_variant": "",
+  "task_id": 0,
+  "task_file": "",
+  "instruction": "Stack the yellow block on the wooden block",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:06:54+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.001988071570577,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "approaching yellow block",
+      "description": "approaching yellow block"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.001988071570577,
+      "t_end_sec": 6.003976143141154,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "grasping and moving yellow block",
+      "description": "grasping and moving yellow block"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.003976143141154,
+      "t_end_sec": 9.00596421471173,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "positioning block on wooden block",
+      "description": "positioning block on wooden block"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.00596421471173,
+      "t_end_sec": 12.007952286282308,
+      "control_step_start": 453,
+      "control_step_end": 603,
+      "policy_step_start": 75,
+      "policy_step_end": 100,
+      "phrase": "holding yellow block on target",
+      "description": "holding yellow block on target"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 12.007952286282308,
+      "t_end_sec": 15.009940357852884,
+      "control_step_start": 604,
+      "control_step_end": 754,
+      "policy_step_start": 100,
+      "policy_step_end": 125,
+      "phrase": "remaining stationary over block",
+      "description": "remaining stationary over block"
+    },
+    {
+      "segment_index": 5,
+      "t_start_sec": 15.009940357852884,
+      "t_end_sec": 18.01192842942346,
+      "control_step_start": 755,
+      "control_step_end": 905,
+      "policy_step_start": 125,
+      "policy_step_end": 150,
+      "phrase": "idle with gripper closed",
+      "description": "idle with gripper closed"
+    },
+    {
+      "segment_index": 6,
+      "t_start_sec": 18.01192842942346,
+      "t_end_sec": 20.457256461232607,
+      "control_step_start": 906,
+      "control_step_end": 1028,
+      "policy_step_start": 151,
+      "policy_step_end": 171,
+      "phrase": "idle with gripper closed",
+      "description": "idle with gripper closed"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "stuck_or_no_progress",
+    "vlm_failure_onset_type": "operator_intervention",
+    "vlm_failure_onset_seconds": 6.242544731610338,
+    "vlm_failure_onset_timestamp": "00:06.2",
+    "vlm_failure_onset_frame": 314,
+    "vlm_failure_onset_step": 314,
+    "vlm_coarse_failure_onset_seconds": 7.0,
+    "vlm_failure_window_start_seconds": 6.5,
+    "vlm_failure_window_end_seconds": 7.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "After placing the yellow block on top of the wooden block, the robot failed to open its gripper to release the object and remained stuck in position until time expired.",
+    "vlm_failure_reason_summary": "stuck on target; failure to release gripper",
+    "vlm_recovery_action": "Open the gripper to release the yellow block onto the wooden block, then raise the arm away from the stack.",
+    "vlm_recovery_action_summary": "open gripper; retract arm",
+    "vlm_prevention_action": "Command the gripper to open immediately once stable contact with the wooden block is made, then retract the arm.",
+    "vlm_prevention_action_summary": "release block on contact; retract arm",
+    "vlm_justification": "The robot successfully picked up the yellow block and aligned it on top of the wooden block by around 00:06. However, at around 00:07, it stopped making progress and remained frozen with its gripper holding the block until the end of the episode.",
+    "vlm_temporal_justification": "At clip second 113 (01:53), the robot finishes placing the yellow block on top of the wooden block. From this point forward, the robotic arm remains completely motionless without opening its gripper to release the block, marking the onset of the failure.",
+    "vlm_video_fps": 50.3,
+    "vlm_video_num_frames": 1029,
+    "vlm_video_duration_seconds": 20.43737574552684,
+    "vlm_usage": {
+      "input_tokens": 22223,
+      "output_tokens": 384,
+      "cached_tokens": 1361
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"stuck_or_no_progress\",\n \"onset_type\": \"operator_intervention\",\n \"failure_onset_seconds\": 7.0,\n \"failure_window_start_seconds\": 6.5,\n \"failure_window_end_seconds\": 7.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"After placing the yellow block on top of the wooden block, the robot failed to open its gripper to release the object and remained stuck in position until time expired.\",\n \"failure_reason_summary\": \"stuck on target; failure to release gripper\",\n \"recovery_action\": \"Open the gripper to release the yellow block onto the wooden block, then raise the arm away from the stack.\",\n \"recovery_action_summary\": \"open gripper; retract arm\",\n \"prevention_action\": \"Command the gripper to open immediately once stable contact with the wooden block is made, then retract the arm.\",\n \"prevention_action_summary\": \"release block on contact; retract arm\",\n \"justification\": \"The robot successfully picked up the yellow block and aligned it on top of the wooden block by around 00:06. However, at around 00:07, it stopped making progress and remained frozen with its gripper holding the block until the end of the episode.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 314,
+    "failure_sim_step": 314,
+    "failure_policy_step": 52,
+    "failure_chunk_index": 2,
+    "first_post_failure_policy_step": 53,
+    "failure_type": "stuck_or_no_progress",
+    "correction_action": "Open the gripper to release the yellow block onto the wooden block, then raise the arm away from the stack."
+  }
+}
\ No newline at end of file
diff --git a/blocks/ep025/labels.npz b/blocks/ep025/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..1259308941d175a3447ccdfbad4bf8adaf12de8e
--- /dev/null
+++ b/blocks/ep025/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7cdf4d4125f497f4ac70468dea01e66aee92b5e264361670b0d4d83c4a5875ef
+size 3910
diff --git a/blocks/ep026/example.md b/blocks/ep026/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..4ce43091cc0d5d398d9d8ccbe53f1fb37d93935b
--- /dev/null
+++ b/blocks/ep026/example.md
@@ -0,0 +1,34 @@
+# t000__ep026__blocks
+
+- **suite / variant:** `` / ``
+- **instruction:** Stack the yellow block on the wooden block
+- **success:** False
+- **control steps:** 577  (11.4s at 51 Hz)
+- **policy inferences:** 99  (replan_steps=6)
+- **video:** `blocks/ep026/rollout.mp4`
+- **features:** `99` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching for and grasping yellow block |
+| 3.0–6.0 | 152–303 | 25–50 | lifting and carrying yellow block |
+| 6.0–9.0 | 304–455 | 50–75 | releasing yellow block beside wooden block |
+| 9.0–11.4 | 456–576 | 76–96 | hovering above wooden block |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 7.411067193675889 s (00:07.4)
+- **frame / control_step:** 375 (policy_step=62, chunk_index=3)
+- **coarse onset:** 5.8 s (refined=True)
+- **mode:** placement_or_insertion_failure
+- **reason:** The robot brought the yellow block over the wooden block but misaligned the placement, causing the yellow block to slide off the side and fall onto the table.
+  Summary: misaligned block placement; object fell off target
+- **recovery:** Re-grasp the yellow block from the table surface, align it directly over the center of the wooden block, and stack it cleanly.
+  Summary: re-grasp yellow block; align and stack on wooden block
+- **prevention:** Align the yellow block directly over the center of the wooden block before opening the gripper to ensure a stable stacked placement.
+  Summary: align block centrally; release gently on top
+- **justification:** At around 5.8 seconds, the robot released the yellow block over the edge of the wooden block without proper centering, causing the yellow block to fall onto the table next to the target block instead of remaining stacked on top.
diff --git a/blocks/ep026/labels.json b/blocks/ep026/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..ba01e9439b30a77d94e69f3d9fa986a5223e9cd0
--- /dev/null
+++ b/blocks/ep026/labels.json
@@ -0,0 +1,109 @@
+{
+  "rollout_id": "t000__ep026__blocks",
+  "scene_variant": "",
+  "task_id": 0,
+  "task_file": "",
+  "instruction": "Stack the yellow block on the wooden block",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:07:29+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0039525691699605,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for and grasping yellow block",
+      "description": "reaching for and grasping yellow block"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0039525691699605,
+      "t_end_sec": 6.007905138339921,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "lifting and carrying yellow block",
+      "description": "lifting and carrying yellow block"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.007905138339921,
+      "t_end_sec": 9.011857707509881,
+      "control_step_start": 304,
+      "control_step_end": 455,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "releasing yellow block beside wooden block",
+      "description": "releasing yellow block beside wooden block"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.011857707509881,
+      "t_end_sec": 11.403162055335969,
+      "control_step_start": 456,
+      "control_step_end": 576,
+      "policy_step_start": 76,
+      "policy_step_end": 96,
+      "phrase": "hovering above wooden block",
+      "description": "hovering above wooden block"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "placement_or_insertion_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 7.411067193675889,
+    "vlm_failure_onset_timestamp": "00:07.4",
+    "vlm_failure_onset_frame": 375,
+    "vlm_failure_onset_step": 375,
+    "vlm_coarse_failure_onset_seconds": 5.8,
+    "vlm_failure_window_start_seconds": 5.3,
+    "vlm_failure_window_end_seconds": 6.3,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot brought the yellow block over the wooden block but misaligned the placement, causing the yellow block to slide off the side and fall onto the table.",
+    "vlm_failure_reason_summary": "misaligned block placement; object fell off target",
+    "vlm_recovery_action": "Re-grasp the yellow block from the table surface, align it directly over the center of the wooden block, and stack it cleanly.",
+    "vlm_recovery_action_summary": "re-grasp yellow block; align and stack on wooden block",
+    "vlm_prevention_action": "Align the yellow block directly over the center of the wooden block before opening the gripper to ensure a stable stacked placement.",
+    "vlm_prevention_action_summary": "align block centrally; release gently on top",
+    "vlm_justification": "At around 5.8 seconds, the robot released the yellow block over the edge of the wooden block without proper centering, causing the yellow block to fall onto the table next to the target block instead of remaining stacked on top.",
+    "vlm_temporal_justification": "At 03:54 (clip second 234), the robot lowers the yellow block onto the far edge of the wooden block, causing it to rest tilted and misaligned, making the placement failure visibly obvious before the gripper releases.",
+    "vlm_video_fps": 50.6,
+    "vlm_video_num_frames": 577,
+    "vlm_video_duration_seconds": 11.383399209486166,
+    "vlm_usage": {
+      "input_tokens": 21781,
+      "output_tokens": 383,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"placement_or_insertion_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 5.8,\n \"failure_window_start_seconds\": 5.3,\n \"failure_window_end_seconds\": 6.3,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot brought the yellow block over the wooden block but misaligned the placement, causing the yellow block to slide off the side and fall onto the table.\",\n \"failure_reason_summary\": \"misaligned block placement; object fell off target\",\n \"recovery_action\": \"Re-grasp the yellow block from the table surface, align it directly over the center of the wooden block, and stack it cleanly.\",\n \"recovery_action_summary\": \"re-grasp yellow block; align and stack on wooden block\",\n \"prevention_action\": \"Align the yellow block directly over the center of the wooden block before opening the gripper to ensure a stable stacked placement.\",\n \"prevention_action_summary\": \"align block centrally; release gently on top\",\n \"justification\": \"At around 5.8 seconds, the robot released the yellow block over the edge of the wooden block without proper centering, causing the yellow block to fall onto the table next to the target block instead of remaining stacked on top.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 375,
+    "failure_sim_step": 375,
+    "failure_policy_step": 62,
+    "failure_chunk_index": 3,
+    "first_post_failure_policy_step": 63,
+    "failure_type": "placement_or_insertion_failure",
+    "correction_action": "Re-grasp the yellow block from the table surface, align it directly over the center of the wooden block, and stack it cleanly."
+  }
+}
\ No newline at end of file
diff --git a/blocks/ep026/labels.npz b/blocks/ep026/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..1fb28a3286fea2af2f167e5d344f4ad95fa37443
--- /dev/null
+++ b/blocks/ep026/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f58b48066de0cbc955f5f4106eb1e66af33fbb6265b05b3eec965a66e13ac12b
+size 3876
diff --git a/blocks/ep027/example.md b/blocks/ep027/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..2ab8abbdf535c99aef0727b4c09f183e192352df
--- /dev/null
+++ b/blocks/ep027/example.md
@@ -0,0 +1,35 @@
+# t000__ep027__blocks
+
+- **suite / variant:** `` / ``
+- **instruction:** Stack the yellow block on the wooden block
+- **success:** False
+- **control steps:** 645  (12.8s at 50 Hz)
+- **policy inferences:** 112  (replan_steps=6)
+- **video:** `blocks/ep027/rollout.mp4`
+- **features:** `112` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | lowering arm toward yellow block |
+| 3.0–6.0 | 151–301 | 25–50 | grasping and lifting yellow block |
+| 6.0–9.0 | 302–452 | 50–75 | placing block on wooden block |
+| 9.0–12.0 | 453–603 | 75–100 | retracting gripper and hovering |
+| 12.0–12.8 | 604–644 | 100–107 | lowering toward table |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 5.059760956175299 s (00:05.1)
+- **frame / control_step:** 254 (policy_step=42, chunk_index=2)
+- **coarse onset:** 8.0 s (refined=True)
+- **mode:** placement_or_insertion_failure
+- **reason:** The robot attempted to stack the yellow block on the wooden block, but misaligned the block upon release, causing it to slide off onto the table surface.
+  Summary: misaligned block placement; block sliding off target
+- **recovery:** Re-grasp the fallen yellow block from the table surface, align it directly over the center of the wooden block, and release it gently.
+  Summary: re-grasp yellow block; place on wooden block
+- **prevention:** Ensure the yellow block is centered and flat on top of the wooden block before releasing the gripper to achieve a stable stack.
+  Summary: center yellow block; release stably
+- **justification:** None
diff --git a/blocks/ep027/labels.json b/blocks/ep027/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..623abaeb541ec1bcc4db2682589176dd4ca385d0
--- /dev/null
+++ b/blocks/ep027/labels.json
@@ -0,0 +1,120 @@
+{
+  "rollout_id": "t000__ep027__blocks",
+  "scene_variant": "",
+  "task_id": 0,
+  "task_file": "",
+  "instruction": "Stack the yellow block on the wooden block",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:08:01+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00796812749004,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "lowering arm toward yellow block",
+      "description": "lowering arm toward yellow block"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00796812749004,
+      "t_end_sec": 6.01593625498008,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "grasping and lifting yellow block",
+      "description": "grasping and lifting yellow block"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01593625498008,
+      "t_end_sec": 9.02390438247012,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "placing block on wooden block",
+      "description": "placing block on wooden block"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.02390438247012,
+      "t_end_sec": 12.03187250996016,
+      "control_step_start": 453,
+      "control_step_end": 603,
+      "policy_step_start": 75,
+      "policy_step_end": 100,
+      "phrase": "retracting gripper and hovering",
+      "description": "retracting gripper and hovering"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 12.03187250996016,
+      "t_end_sec": 12.848605577689243,
+      "control_step_start": 604,
+      "control_step_end": 644,
+      "policy_step_start": 100,
+      "policy_step_end": 107,
+      "phrase": "lowering toward table",
+      "description": "lowering toward table"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "placement_or_insertion_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 5.059760956175299,
+    "vlm_failure_onset_timestamp": "00:05.1",
+    "vlm_failure_onset_frame": 254,
+    "vlm_failure_onset_step": 254,
+    "vlm_coarse_failure_onset_seconds": 8.0,
+    "vlm_failure_window_start_seconds": 7.5,
+    "vlm_failure_window_end_seconds": 8.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to stack the yellow block on the wooden block, but misaligned the block upon release, causing it to slide off onto the table surface.",
+    "vlm_failure_reason_summary": "misaligned block placement; block sliding off target",
+    "vlm_recovery_action": "Re-grasp the fallen yellow block from the table surface, align it directly over the center of the wooden block, and release it gently.",
+    "vlm_recovery_action_summary": "re-grasp yellow block; place on wooden block",
+    "vlm_prevention_action": "Ensure the yellow block is centered and flat on top of the wooden block before releasing the gripper to achieve a stable stack.",
+    "vlm_prevention_action_summary": "center yellow block; release stably",
+    "vlm_justification": null,
+    "vlm_temporal_justification": "At 03:01, the robot gripper opens to release the yellow block onto the wooden block. However, the block is misaligned over the right edge of the wooden block, causing it to tilt and fall onto the table surface immediately upon release.",
+    "vlm_video_fps": 50.2,
+    "vlm_video_num_frames": 645,
+    "vlm_video_duration_seconds": 12.828685258964143,
+    "vlm_usage": {
+      "input_tokens": 21781,
+      "output_tokens": 331,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"placement_or_insertion_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 8.0,\n \"failure_window_start_seconds\": 7.5,\n \"failure_window_end_seconds\": 8.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to stack the yellow block on the wooden block, but misaligned the block upon release, causing it to slide off onto the table surface.\",\n \"failure_reason_summary\": \"misaligned block placement; block sliding off target\",\n \"recovery_action\": \"Re-grasp the fallen yellow block from the table surface, align it directly over the center of the wooden block, and release it gently.\",\n \"recovery_action_summary\": \"re-grasp yellow block; place on wooden block\",\n \"prevention_action\": \"Ensure the yellow block is centered and flat on top of the wooden block before releasing the gripper to achieve a stable stack.\",\n \"prevention_action_summary\": \"center yellow block; release stably\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 254,
+    "failure_sim_step": 254,
+    "failure_policy_step": 42,
+    "failure_chunk_index": 2,
+    "first_post_failure_policy_step": 43,
+    "failure_type": "placement_or_insertion_failure",
+    "correction_action": "Re-grasp the fallen yellow block from the table surface, align it directly over the center of the wooden block, and release it gently."
+  }
+}
\ No newline at end of file
diff --git a/blocks/ep027/labels.npz b/blocks/ep027/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..de793820e7171b9953460fc495a6d2108894cc55
--- /dev/null
+++ b/blocks/ep027/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b4ed3e1121b6dbdfdeee41a256699f1e0f73dc7e881c0da6244d20314633eb00
+size 3891
diff --git a/blocks/ep028/example.md b/blocks/ep028/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..708e010348923dafac2bb683d16f2d95ff4ffcbd
--- /dev/null
+++ b/blocks/ep028/example.md
@@ -0,0 +1,37 @@
+# t000__ep028__blocks
+
+- **suite / variant:** `` / ``
+- **instruction:** Stack the yellow block on the wooden block
+- **success:** False
+- **control steps:** 934  (18.5s at 50 Hz)
+- **policy inferences:** 160  (replan_steps=6)
+- **video:** `blocks/ep028/rollout.mp4`
+- **features:** `160` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | approaching yellow block |
+| 3.0–6.0 | 151–301 | 25–50 | grasping yellow block |
+| 6.0–9.0 | 302–452 | 50–75 | moving yellow block toward wooden block |
+| 9.0–12.0 | 453–603 | 75–100 | lowering yellow block onto wooden block |
+| 12.0–15.0 | 604–754 | 100–125 | positioning yellow block on wooden block |
+| 15.0–18.0 | 755–905 | 125–150 | releasing yellow block; block slipping off |
+| 18.0–18.5 | 906–933 | 151–155 | retreating gripper |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 16.567460317460316 s (00:16.6)
+- **frame / control_step:** 835 (policy_step=139, chunk_index=1)
+- **coarse onset:** 15.5 s (refined=True)
+- **mode:** placement_or_insertion_failure
+- **reason:** The robot released the yellow block onto the wooden block without proper alignment, causing the yellow block to slide off and fall onto the table as the gripper opened.
+  Summary: unstable placement; block sliding off
+- **recovery:** Re-grasp the yellow block from the table surface, align it directly over the center of the wooden block, and place it gently.
+  Summary: re-grasp yellow block; stack stably on wooden block
+- **prevention:** Ensure the yellow block is centered and leveled over the wooden block prior to opening the gripper.
+  Summary: center yellow block; release gently on target
+- **justification:** The robot successfully picked up the yellow block and brought it to the wooden block, but upon releasing at ~15.5s, the placement was misaligned and unstable, causing the yellow block to fall off onto the table by 17.0s.
diff --git a/blocks/ep028/labels.json b/blocks/ep028/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..b219ff049701476baa294122046b975a868ef4a8
--- /dev/null
+++ b/blocks/ep028/labels.json
@@ -0,0 +1,142 @@
+{
+  "rollout_id": "t000__ep028__blocks",
+  "scene_variant": "",
+  "task_id": 0,
+  "task_file": "",
+  "instruction": "Stack the yellow block on the wooden block",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:08:35+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.996031746031746,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "approaching yellow block",
+      "description": "approaching yellow block"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.996031746031746,
+      "t_end_sec": 5.992063492063492,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "grasping yellow block",
+      "description": "grasping yellow block"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.992063492063492,
+      "t_end_sec": 8.988095238095239,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "moving yellow block toward wooden block",
+      "description": "moving yellow block toward wooden block"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.988095238095239,
+      "t_end_sec": 11.984126984126984,
+      "control_step_start": 453,
+      "control_step_end": 603,
+      "policy_step_start": 75,
+      "policy_step_end": 100,
+      "phrase": "lowering yellow block onto wooden block",
+      "description": "lowering yellow block onto wooden block"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 11.984126984126984,
+      "t_end_sec": 14.980158730158731,
+      "control_step_start": 604,
+      "control_step_end": 754,
+      "policy_step_start": 100,
+      "policy_step_end": 125,
+      "phrase": "positioning yellow block on wooden block",
+      "description": "positioning yellow block on wooden block"
+    },
+    {
+      "segment_index": 5,
+      "t_start_sec": 14.980158730158731,
+      "t_end_sec": 17.976190476190478,
+      "control_step_start": 755,
+      "control_step_end": 905,
+      "policy_step_start": 125,
+      "policy_step_end": 150,
+      "phrase": "releasing yellow block; block slipping off",
+      "description": "releasing yellow block; block slipping off"
+    },
+    {
+      "segment_index": 6,
+      "t_start_sec": 17.976190476190478,
+      "t_end_sec": 18.53174603174603,
+      "control_step_start": 906,
+      "control_step_end": 933,
+      "policy_step_start": 151,
+      "policy_step_end": 155,
+      "phrase": "retreating gripper",
+      "description": "retreating gripper"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "placement_or_insertion_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 16.567460317460316,
+    "vlm_failure_onset_timestamp": "00:16.6",
+    "vlm_failure_onset_frame": 835,
+    "vlm_failure_onset_step": 835,
+    "vlm_coarse_failure_onset_seconds": 15.5,
+    "vlm_failure_window_start_seconds": 14.5,
+    "vlm_failure_window_end_seconds": 17.0,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot released the yellow block onto the wooden block without proper alignment, causing the yellow block to slide off and fall onto the table as the gripper opened.",
+    "vlm_failure_reason_summary": "unstable placement; block sliding off",
+    "vlm_recovery_action": "Re-grasp the yellow block from the table surface, align it directly over the center of the wooden block, and place it gently.",
+    "vlm_recovery_action_summary": "re-grasp yellow block; stack stably on wooden block",
+    "vlm_prevention_action": "Ensure the yellow block is centered and leveled over the wooden block prior to opening the gripper.",
+    "vlm_prevention_action_summary": "center yellow block; release gently on target",
+    "vlm_justification": "The robot successfully picked up the yellow block and brought it to the wooden block, but upon releasing at ~15.5s, the placement was misaligned and unstable, causing the yellow block to fall off onto the table by 17.0s.",
+    "vlm_temporal_justification": "At 03:25 (second 205 of the clip), the robot opens its gripper to release the yellow block while it is poorly aligned with the wooden block below, causing the yellow block to slide off and fail the stacking task.",
+    "vlm_video_fps": 50.4,
+    "vlm_video_num_frames": 934,
+    "vlm_video_duration_seconds": 18.511904761904763,
+    "vlm_usage": {
+      "input_tokens": 22161,
+      "output_tokens": 386,
+      "cached_tokens": 1293
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"placement_or_insertion_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 15.5,\n \"failure_window_start_seconds\": 14.5,\n \"failure_window_end_seconds\": 17.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot released the yellow block onto the wooden block without proper alignment, causing the yellow block to slide off and fall onto the table as the gripper opened.\",\n \"failure_reason_summary\": \"unstable placement; block sliding off\",\n \"recovery_action\": \"Re-grasp the yellow block from the table surface, align it directly over the center of the wooden block, and place it gently.\",\n \"recovery_action_summary\": \"re-grasp yellow block; stack stably on wooden block\",\n \"prevention_action\": \"Ensure the yellow block is centered and leveled over the wooden block prior to opening the gripper.\",\n \"prevention_action_summary\": \"center yellow block; release gently on target\",\n \"justification\": \"The robot successfully picked up the yellow block and brought it to the wooden block, but upon releasing at ~15.5s, the placement was misaligned and unstable, causing the yellow block to fall off onto the table by 17.0s.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 835,
+    "failure_sim_step": 835,
+    "failure_policy_step": 139,
+    "failure_chunk_index": 1,
+    "first_post_failure_policy_step": 140,
+    "failure_type": "placement_or_insertion_failure",
+    "correction_action": "Re-grasp the yellow block from the table surface, align it directly over the center of the wooden block, and place it gently."
+  }
+}
\ No newline at end of file
diff --git a/blocks/ep028/labels.npz b/blocks/ep028/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..227e2502335a491c2cd628264bb8740ddc3401e0
--- /dev/null
+++ b/blocks/ep028/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:de27c21183c816b685eabecc8b50b8be74bb7b00722d4f784d5cd5fa8bb80071
+size 3919
diff --git a/blocks/ep029/example.md b/blocks/ep029/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..c7e6bd9fb336023d6508a8714c54cb644a1acbf2
--- /dev/null
+++ b/blocks/ep029/example.md
@@ -0,0 +1,34 @@
+# t000__ep029__blocks
+
+- **suite / variant:** `` / ``
+- **instruction:** Stack the yellow block on the wooden block
+- **success:** False
+- **control steps:** 461  (9.1s at 51 Hz)
+- **policy inferences:** 79  (replan_steps=6)
+- **video:** `blocks/ep029/rollout.mp4`
+- **features:** `79` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | opening gripper; hovering at home |
+| 3.0–6.0 | 152–303 | 25–50 | idle at home position |
+| 6.0–9.0 | 304–455 | 50–75 | idle at home position |
+| 9.0–9.1 | 456–460 | 76–76 | idle at home position |
+
+## Failure
+
+- **yes**
+- **onset type:** operator_intervention
+- **onset:** 0.05917159763313609 s (00:00.1)
+- **frame / control_step:** 3 (policy_step=0, chunk_index=3)
+- **coarse onset:** 2.5 s (refined=True)
+- **mode:** stuck_or_no_progress
+- **reason:** The robot arm remained stationary in its initial home position after opening its gripper, making no progress toward reaching or grasping the yellow block.
+  Summary: robot stationary at home position; no progress toward blocks
+- **recovery:** Guide the robot arm down to the yellow block, grasp it, and place it on top of the wooden block.
+  Summary: guide arm to yellow block; grasp and stack on wooden block
+- **prevention:** Ensure the policy generates active trajectory commands toward the yellow block immediately following task initialization.
+  Summary: generate motion commands toward yellow block; execute grasp
+- **justification:** The robot opens its end-effector shortly after starting, but then stays completely frozen at its home pose throughout the rest of the rollout without approaching the objects.
diff --git a/blocks/ep029/labels.json b/blocks/ep029/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..5c1a41cd0b703496ea415824cd313b282925b9ee
--- /dev/null
+++ b/blocks/ep029/labels.json
@@ -0,0 +1,109 @@
+{
+  "rollout_id": "t000__ep029__blocks",
+  "scene_variant": "",
+  "task_id": 0,
+  "task_file": "",
+  "instruction": "Stack the yellow block on the wooden block",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:09:05+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9980276134122286,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "opening gripper; hovering at home",
+      "description": "opening gripper; hovering at home"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9980276134122286,
+      "t_end_sec": 5.996055226824457,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "idle at home position",
+      "description": "idle at home position"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.996055226824457,
+      "t_end_sec": 8.994082840236686,
+      "control_step_start": 304,
+      "control_step_end": 455,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "idle at home position",
+      "description": "idle at home position"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.994082840236686,
+      "t_end_sec": 9.092702169625246,
+      "control_step_start": 456,
+      "control_step_end": 460,
+      "policy_step_start": 76,
+      "policy_step_end": 76,
+      "phrase": "idle at home position",
+      "description": "idle at home position"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "stuck_or_no_progress",
+    "vlm_failure_onset_type": "operator_intervention",
+    "vlm_failure_onset_seconds": 0.05917159763313609,
+    "vlm_failure_onset_timestamp": "00:00.1",
+    "vlm_failure_onset_frame": 3,
+    "vlm_failure_onset_step": 3,
+    "vlm_coarse_failure_onset_seconds": 2.5,
+    "vlm_failure_window_start_seconds": 2.0,
+    "vlm_failure_window_end_seconds": 3.0,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot arm remained stationary in its initial home position after opening its gripper, making no progress toward reaching or grasping the yellow block.",
+    "vlm_failure_reason_summary": "robot stationary at home position; no progress toward blocks",
+    "vlm_recovery_action": "Guide the robot arm down to the yellow block, grasp it, and place it on top of the wooden block.",
+    "vlm_recovery_action_summary": "guide arm to yellow block; grasp and stack on wooden block",
+    "vlm_prevention_action": "Ensure the policy generates active trajectory commands toward the yellow block immediately following task initialization.",
+    "vlm_prevention_action_summary": "generate motion commands toward yellow block; execute grasp",
+    "vlm_justification": "The robot opens its end-effector shortly after starting, but then stays completely frozen at its home pose throughout the rest of the rollout without approaching the objects.",
+    "vlm_temporal_justification": "The robot arm opens its gripper during the first two frames. By frame 3 (00:03), the gripper is fully open and the arm becomes completely stationary in its home position, making no further movement toward the target objects.",
+    "vlm_video_fps": 50.7,
+    "vlm_video_num_frames": 461,
+    "vlm_video_duration_seconds": 9.072978303747535,
+    "vlm_usage": {
+      "input_tokens": 20066,
+      "output_tokens": 349,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"stuck_or_no_progress\",\n \"onset_type\": \"operator_intervention\",\n \"failure_onset_seconds\": 2.5,\n \"failure_window_start_seconds\": 2.0,\n \"failure_window_end_seconds\": 3.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot arm remained stationary in its initial home position after opening its gripper, making no progress toward reaching or grasping the yellow block.\",\n \"failure_reason_summary\": \"robot stationary at home position; no progress toward blocks\",\n \"recovery_action\": \"Guide the robot arm down to the yellow block, grasp it, and place it on top of the wooden block.\",\n \"recovery_action_summary\": \"guide arm to yellow block; grasp and stack on wooden block\",\n \"prevention_action\": \"Ensure the policy generates active trajectory commands toward the yellow block immediately following task initialization.\",\n \"prevention_action_summary\": \"generate motion commands toward yellow block; execute grasp\",\n \"justification\": \"The robot opens its end-effector shortly after starting, but then stays completely frozen at its home pose throughout the rest of the rollout without approaching the objects.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 3,
+    "failure_sim_step": 3,
+    "failure_policy_step": 0,
+    "failure_chunk_index": 3,
+    "first_post_failure_policy_step": 1,
+    "failure_type": "stuck_or_no_progress",
+    "correction_action": "Guide the robot arm down to the yellow block, grasp it, and place it on top of the wooden block."
+  }
+}
\ No newline at end of file
diff --git a/blocks/ep029/labels.npz b/blocks/ep029/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..4073b413315d3a10db0e3488aa018a9650856bff
--- /dev/null
+++ b/blocks/ep029/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ab207c8c021b745f78d612094a1a5c2eec7279ef7d8887e68ddba1c032c4250e
+size 3825
diff --git a/blocks/ep030/example.md b/blocks/ep030/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..b5ac5400e5424eb0a3c4760f0e6a118b67575b54
--- /dev/null
+++ b/blocks/ep030/example.md
@@ -0,0 +1,33 @@
+# t000__ep030__blocks
+
+- **suite / variant:** `` / ``
+- **instruction:** Stack the yellow block on the wooden block
+- **success:** False
+- **control steps:** 394  (7.8s at 51 Hz)
+- **policy inferences:** 68  (replan_steps=6)
+- **video:** `blocks/ep030/rollout.mp4`
+- **features:** `68` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | lowering end-effector toward yellow block |
+| 3.0–6.0 | 152–303 | 25–50 | closing gripper without holding block |
+| 6.0–7.8 | 304–393 | 50–65 | moving empty gripper over wooden block |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 4.220907297830374 s (00:04.2)
+- **frame / control_step:** 214 (policy_step=35, chunk_index=4)
+- **coarse onset:** 4.0 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot closed its gripper above the yellow block without properly enveloping or grasping it, then moved away toward the wooden block with empty fingers.
+  Summary: failed grasp on yellow block; empty gripper lift
+- **recovery:** Move the end-effector back to the yellow block, open the gripper, position it fully around the block, grasp it securely, and then carry it to the wooden block.
+  Summary: re-align gripper; grasp yellow block; stack on wooden block
+- **prevention:** Lower the end-effector further to align the gripper fingers around the yellow block before closing them to guarantee a firm hold.
+  Summary: lower gripper fully; secure block before lifting
+- **justification:** At around 4.0 seconds (frame 00:04), the gripper closes in mid-air above the yellow block, failing to grasp it. The robot then proceeds to move toward the target without the block.
diff --git a/blocks/ep030/labels.json b/blocks/ep030/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..ebd0b339ef7189e3b68b0cefe2c9f582c7bbf2bb
--- /dev/null
+++ b/blocks/ep030/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t000__ep030__blocks",
+  "scene_variant": "",
+  "task_id": 0,
+  "task_file": "",
+  "instruction": "Stack the yellow block on the wooden block",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:09:35+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9980276134122286,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "lowering end-effector toward yellow block",
+      "description": "lowering end-effector toward yellow block"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9980276134122286,
+      "t_end_sec": 5.996055226824457,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "closing gripper without holding block",
+      "description": "closing gripper without holding block"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.996055226824457,
+      "t_end_sec": 7.77120315581854,
+      "control_step_start": 304,
+      "control_step_end": 393,
+      "policy_step_start": 50,
+      "policy_step_end": 65,
+      "phrase": "moving empty gripper over wooden block",
+      "description": "moving empty gripper over wooden block"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 4.220907297830374,
+    "vlm_failure_onset_timestamp": "00:04.2",
+    "vlm_failure_onset_frame": 214,
+    "vlm_failure_onset_step": 214,
+    "vlm_coarse_failure_onset_seconds": 4.0,
+    "vlm_failure_window_start_seconds": 3.5,
+    "vlm_failure_window_end_seconds": 4.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot closed its gripper above the yellow block without properly enveloping or grasping it, then moved away toward the wooden block with empty fingers.",
+    "vlm_failure_reason_summary": "failed grasp on yellow block; empty gripper lift",
+    "vlm_recovery_action": "Move the end-effector back to the yellow block, open the gripper, position it fully around the block, grasp it securely, and then carry it to the wooden block.",
+    "vlm_recovery_action_summary": "re-align gripper; grasp yellow block; stack on wooden block",
+    "vlm_prevention_action": "Lower the end-effector further to align the gripper fingers around the yellow block before closing them to guarantee a firm hold.",
+    "vlm_prevention_action_summary": "lower gripper fully; secure block before lifting",
+    "vlm_justification": "At around 4.0 seconds (frame 00:04), the gripper closes in mid-air above the yellow block, failing to grasp it. The robot then proceeds to move toward the target without the block.",
+    "vlm_temporal_justification": "At timestamp 02:43 (frame 163 of the clip), the robot gripper begins closing its fingers while still positioned above the yellow block without lowering far enough to envelop or grasp it.",
+    "vlm_video_fps": 50.7,
+    "vlm_video_num_frames": 394,
+    "vlm_video_duration_seconds": 7.751479289940828,
+    "vlm_usage": {
+      "input_tokens": 21579,
+      "output_tokens": 377,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 4.0,\n \"failure_window_start_seconds\": 3.5,\n \"failure_window_end_seconds\": 4.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot closed its gripper above the yellow block without properly enveloping or grasping it, then moved away toward the wooden block with empty fingers.\",\n \"failure_reason_summary\": \"failed grasp on yellow block; empty gripper lift\",\n \"recovery_action\": \"Move the end-effector back to the yellow block, open the gripper, position it fully around the block, grasp it securely, and then carry it to the wooden block.\",\n \"recovery_action_summary\": \"re-align gripper; grasp yellow block; stack on wooden block\",\n \"prevention_action\": \"Lower the end-effector further to align the gripper fingers around the yellow block before closing them to guarantee a firm hold.\",\n \"prevention_action_summary\": \"lower gripper fully; secure block before lifting\",\n \"justification\": \"At around 4.0 seconds (frame 00:04), the gripper closes in mid-air above the yellow block, failing to grasp it. The robot then proceeds to move toward the target without the block.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 214,
+    "failure_sim_step": 214,
+    "failure_policy_step": 35,
+    "failure_chunk_index": 4,
+    "first_post_failure_policy_step": 36,
+    "failure_type": "grasp_failure",
+    "correction_action": "Move the end-effector back to the yellow block, open the gripper, position it fully around the block, grasp it securely, and then carry it to the wooden block."
+  }
+}
\ No newline at end of file
diff --git a/blocks/ep030/labels.npz b/blocks/ep030/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..ed99052c493b790420595c22a61d3c0abc1c3e30
--- /dev/null
+++ b/blocks/ep030/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:69139d5c1040f98672656d6c503fc25fdbaf100f66b500805b6abf23406a8f5f
+size 3860
diff --git a/blocks/ep031/example.md b/blocks/ep031/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..3d49c7f96ee6fb06a6e19ce0f7eb2f91d101c6c7
--- /dev/null
+++ b/blocks/ep031/example.md
@@ -0,0 +1,33 @@
+# t000__ep031__blocks
+
+- **suite / variant:** `` / ``
+- **instruction:** Stack the yellow block on the wooden block
+- **success:** False
+- **control steps:** 364  (7.3s at 50 Hz)
+- **policy inferences:** 63  (replan_steps=6)
+- **video:** `blocks/ep031/rollout.mp4`
+- **features:** `63` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–149 | 0–24 | lowering arm toward yellow block |
+| 3.0–6.0 | 150–299 | 25–49 | closing gripper; lifting empty |
+| 6.0–7.3 | 300–363 | 50–60 | hovering empty gripper above block |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 4.06 s (00:04.1)
+- **frame / control_step:** 203 (policy_step=33, chunk_index=5)
+- **coarse onset:** 4.8 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot aligned with the yellow block and attempted to grasp it, but the gripper slipped during closure, failing to hold the block as the arm lifted.
+  Summary: grasp failed; yellow block not acquired
+- **recovery:** Lower the arm back down, reposition the gripper over the center of the yellow block, securely close the fingers around it, and resume stacking on the wooden block.
+  Summary: reposition gripper; regrasp yellow block; stack on wooden block
+- **prevention:** Ensure precise alignment and grasp depth over the yellow block before closing the gripper so that it is firmly held during lifting.
+  Summary: center gripper over block; secure grasp before lifting
+- **justification:** Around 4.8 seconds into the video, the robot closes its gripper on the yellow block, but fails to establish a stable grasp. The fingers close completely while the yellow block remains on the table, and the arm lifts up empty.
diff --git a/blocks/ep031/labels.json b/blocks/ep031/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..9c57a2ba22e3106931e2845ad07d273f1af419c7
--- /dev/null
+++ b/blocks/ep031/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t000__ep031__blocks",
+  "scene_variant": "",
+  "task_id": 0,
+  "task_file": "",
+  "instruction": "Stack the yellow block on the wooden block",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:10:08+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0,
+      "control_step_start": 0,
+      "control_step_end": 149,
+      "policy_step_start": 0,
+      "policy_step_end": 24,
+      "phrase": "lowering arm toward yellow block",
+      "description": "lowering arm toward yellow block"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0,
+      "t_end_sec": 6.0,
+      "control_step_start": 150,
+      "control_step_end": 299,
+      "policy_step_start": 25,
+      "policy_step_end": 49,
+      "phrase": "closing gripper; lifting empty",
+      "description": "closing gripper; lifting empty"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.0,
+      "t_end_sec": 7.28,
+      "control_step_start": 300,
+      "control_step_end": 363,
+      "policy_step_start": 50,
+      "policy_step_end": 60,
+      "phrase": "hovering empty gripper above block",
+      "description": "hovering empty gripper above block"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 4.06,
+    "vlm_failure_onset_timestamp": "00:04.1",
+    "vlm_failure_onset_frame": 203,
+    "vlm_failure_onset_step": 203,
+    "vlm_coarse_failure_onset_seconds": 4.8,
+    "vlm_failure_window_start_seconds": 4.0,
+    "vlm_failure_window_end_seconds": 5.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot aligned with the yellow block and attempted to grasp it, but the gripper slipped during closure, failing to hold the block as the arm lifted.",
+    "vlm_failure_reason_summary": "grasp failed; yellow block not acquired",
+    "vlm_recovery_action": "Lower the arm back down, reposition the gripper over the center of the yellow block, securely close the fingers around it, and resume stacking on the wooden block.",
+    "vlm_recovery_action_summary": "reposition gripper; regrasp yellow block; stack on wooden block",
+    "vlm_prevention_action": "Ensure precise alignment and grasp depth over the yellow block before closing the gripper so that it is firmly held during lifting.",
+    "vlm_prevention_action_summary": "center gripper over block; secure grasp before lifting",
+    "vlm_justification": "Around 4.8 seconds into the video, the robot closes its gripper on the yellow block, but fails to establish a stable grasp. The fingers close completely while the yellow block remains on the table, and the arm lifts up empty.",
+    "vlm_temporal_justification": "At 01:53 (113 seconds into the clip), as the gripper fingers close around the yellow block, they begin slipping off its edges instead of establishing a secure hold, leading to a failed grasp as the fingers close empty.",
+    "vlm_video_fps": 50.0,
+    "vlm_video_num_frames": 364,
+    "vlm_video_duration_seconds": 7.26,
+    "vlm_usage": {
+      "input_tokens": 19566,
+      "output_tokens": 398,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n  \"failure_mode\": \"grasp_failure\",\n  \"onset_type\": \"obvious_mistake\",\n  \"failure_onset_seconds\": 4.8,\n  \"failure_window_start_seconds\": 4.0,\n  \"failure_window_end_seconds\": 5.5,\n  \"confidence\": \"high\",\n  \"failure_reason\": \"The robot aligned with the yellow block and attempted to grasp it, but the gripper slipped during closure, failing to hold the block as the arm lifted.\",\n  \"failure_reason_summary\": \"grasp failed; yellow block not acquired\",\n  \"recovery_action\": \"Lower the arm back down, reposition the gripper over the center of the yellow block, securely close the fingers around it, and resume stacking on the wooden block.\",\n  \"recovery_action_summary\": \"reposition gripper; regrasp yellow block; stack on wooden block\",\n  \"prevention_action\": \"Ensure precise alignment and grasp depth over the yellow block before closing the gripper so that it is firmly held during lifting.\",\n  \"prevention_action_summary\": \"center gripper over block; secure grasp before lifting\",\n  \"justification\": \"Around 4.8 seconds into the video, the robot closes its gripper on the yellow block, but fails to establish a stable grasp. The fingers close completely while the yellow block remains on the table, and the arm lifts up empty.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 203,
+    "failure_sim_step": 203,
+    "failure_policy_step": 33,
+    "failure_chunk_index": 5,
+    "first_post_failure_policy_step": 34,
+    "failure_type": "grasp_failure",
+    "correction_action": "Lower the arm back down, reposition the gripper over the center of the yellow block, securely close the fingers around it, and resume stacking on the wooden block."
+  }
+}
\ No newline at end of file
diff --git a/blocks/ep031/labels.npz b/blocks/ep031/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..edc56eb92fcb1c74d32031bb094c1078680c1fe1
--- /dev/null
+++ b/blocks/ep031/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7804885e5c0e0c1ff7c5b2714a4c066b452df165ce06d09e8bbe995d7ed4dcfd
+size 3862
diff --git a/blocks/ep032/example.md b/blocks/ep032/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..540f7af36b442e01c5b148da0e60f1ba67930968
--- /dev/null
+++ b/blocks/ep032/example.md
@@ -0,0 +1,26 @@
+# t000__ep032__blocks
+
+- **suite / variant:** `` / ``
+- **instruction:** Stack the yellow block on the wooden block
+- **success:** True
+- **control steps:** 1200  (23.8s at 50 Hz)
+- **policy inferences:** 206  (replan_steps=6)
+- **video:** `blocks/ep032/rollout.mp4`
+- **features:** `206` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | reaching and grasping yellow block |
+| 3.0–6.0 | 151–301 | 25–50 | carrying yellow block to wooden block |
+| 6.0–9.0 | 302–452 | 50–75 | lowering yellow block onto wooden block |
+| 9.0–12.0 | 453–603 | 75–100 | placing yellow block on wooden block |
+| 12.0–15.0 | 604–754 | 100–125 | correcting placement of yellow block |
+| 15.0–18.0 | 755–905 | 125–150 | repositioning yellow block on wooden block |
+| 18.0–21.0 | 906–1056 | 151–176 | balancing yellow block on wooden block |
+| 21.0–23.8 | 1057–1199 | 176–199 | holding stacked yellow block |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/blocks/ep032/labels.json b/blocks/ep032/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..2725e3103d82237c7c6d643e1b9397f1a88b94a4
--- /dev/null
+++ b/blocks/ep032/labels.json
@@ -0,0 +1,123 @@
+{
+  "rollout_id": "t000__ep032__blocks",
+  "scene_variant": "",
+  "task_id": 0,
+  "task_file": "",
+  "instruction": "Stack the yellow block on the wooden block",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:10:18+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.996031746031746,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching and grasping yellow block",
+      "description": "reaching and grasping yellow block"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.996031746031746,
+      "t_end_sec": 5.992063492063492,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "carrying yellow block to wooden block",
+      "description": "carrying yellow block to wooden block"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.992063492063492,
+      "t_end_sec": 8.988095238095239,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "lowering yellow block onto wooden block",
+      "description": "lowering yellow block onto wooden block"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.988095238095239,
+      "t_end_sec": 11.984126984126984,
+      "control_step_start": 453,
+      "control_step_end": 603,
+      "policy_step_start": 75,
+      "policy_step_end": 100,
+      "phrase": "placing yellow block on wooden block",
+      "description": "placing yellow block on wooden block"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 11.984126984126984,
+      "t_end_sec": 14.980158730158731,
+      "control_step_start": 604,
+      "control_step_end": 754,
+      "policy_step_start": 100,
+      "policy_step_end": 125,
+      "phrase": "correcting placement of yellow block",
+      "description": "correcting placement of yellow block"
+    },
+    {
+      "segment_index": 5,
+      "t_start_sec": 14.980158730158731,
+      "t_end_sec": 17.976190476190478,
+      "control_step_start": 755,
+      "control_step_end": 905,
+      "policy_step_start": 125,
+      "policy_step_end": 150,
+      "phrase": "repositioning yellow block on wooden block",
+      "description": "repositioning yellow block on wooden block"
+    },
+    {
+      "segment_index": 6,
+      "t_start_sec": 17.976190476190478,
+      "t_end_sec": 20.97222222222222,
+      "control_step_start": 906,
+      "control_step_end": 1056,
+      "policy_step_start": 151,
+      "policy_step_end": 176,
+      "phrase": "balancing yellow block on wooden block",
+      "description": "balancing yellow block on wooden block"
+    },
+    {
+      "segment_index": 7,
+      "t_start_sec": 20.97222222222222,
+      "t_end_sec": 23.80952380952381,
+      "control_step_start": 1057,
+      "control_step_end": 1199,
+      "policy_step_start": 176,
+      "policy_step_end": 199,
+      "phrase": "holding stacked yellow block",
+      "description": "holding stacked yellow block"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/blocks/ep032/labels.npz b/blocks/ep032/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..2688409bbf236a4eede4cc559185f65547453644
--- /dev/null
+++ b/blocks/ep032/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1f80074a5d6c079d5681a5dd45405ebf08f940725a2b852fa8943e24d352c55c
+size 3268
diff --git a/blocks/ep033/example.md b/blocks/ep033/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..a9fdffe12ff3a5f6fbe88f2a5d8ce9e4d5b80eb8
--- /dev/null
+++ b/blocks/ep033/example.md
@@ -0,0 +1,37 @@
+# t000__ep033__blocks
+
+- **suite / variant:** `` / ``
+- **instruction:** Stack the yellow block on the wooden block
+- **success:** False
+- **control steps:** 981  (19.6s at 50 Hz)
+- **policy inferences:** 169  (replan_steps=6)
+- **video:** `blocks/ep033/rollout.mp4`
+- **features:** `169` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–149 | 0–24 | reaching and grasping yellow block |
+| 3.0–6.0 | 150–299 | 25–49 | carrying yellow block toward target |
+| 6.0–9.0 | 300–449 | 50–74 | lowering yellow block over target |
+| 9.0–12.0 | 450–599 | 75–99 | releasing yellow block onto table |
+| 12.0–15.0 | 600–749 | 100–124 | hovering gripper above table |
+| 15.0–18.0 | 750–899 | 125–149 | remaining idle near blocks |
+| 18.0–19.6 | 900–980 | 150–163 | idle near wooden block |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 7.485029940119761 s (00:07.5)
+- **frame / control_step:** 375 (policy_step=62, chunk_index=3)
+- **coarse onset:** 9.0 s (refined=True)
+- **mode:** placement_or_insertion_failure
+- **reason:** The robot carries the yellow block to the wooden block but misaligns the placement, releasing the yellow block beside the wooden block onto the table surface rather than stacking it directly on top.
+  Summary: misaligned placement; yellow block released beside wooden block
+- **recovery:** Re-grasp the yellow block from the table, lift it above the wooden block, align it properly with the top surface, and place it down securely.
+  Summary: re-grasp yellow block; center and stack on wooden block
+- **prevention:** Position the yellow block directly over the center of the wooden block at a sufficient height before opening the gripper to release it.
+  Summary: center yellow block over target; release carefully onto surface
+- **justification:** At around 9.0 seconds, as the robot attempts to place the yellow block on top of the wooden block, it lowers the gripper off-center, causing the yellow block to slip down and land on the table beside the wooden block.
diff --git a/blocks/ep033/labels.json b/blocks/ep033/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..6d44916b7c77d4e76e3de74f1fa04d85348aa41a
--- /dev/null
+++ b/blocks/ep033/labels.json
@@ -0,0 +1,142 @@
+{
+  "rollout_id": "t000__ep033__blocks",
+  "scene_variant": "",
+  "task_id": 0,
+  "task_file": "",
+  "instruction": "Stack the yellow block on the wooden block",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:10:54+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9940119760479043,
+      "control_step_start": 0,
+      "control_step_end": 149,
+      "policy_step_start": 0,
+      "policy_step_end": 24,
+      "phrase": "reaching and grasping yellow block",
+      "description": "reaching and grasping yellow block"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9940119760479043,
+      "t_end_sec": 5.9880239520958085,
+      "control_step_start": 150,
+      "control_step_end": 299,
+      "policy_step_start": 25,
+      "policy_step_end": 49,
+      "phrase": "carrying yellow block toward target",
+      "description": "carrying yellow block toward target"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.9880239520958085,
+      "t_end_sec": 8.982035928143713,
+      "control_step_start": 300,
+      "control_step_end": 449,
+      "policy_step_start": 50,
+      "policy_step_end": 74,
+      "phrase": "lowering yellow block over target",
+      "description": "lowering yellow block over target"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.982035928143713,
+      "t_end_sec": 11.976047904191617,
+      "control_step_start": 450,
+      "control_step_end": 599,
+      "policy_step_start": 75,
+      "policy_step_end": 99,
+      "phrase": "releasing yellow block onto table",
+      "description": "releasing yellow block onto table"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 11.976047904191617,
+      "t_end_sec": 14.970059880239521,
+      "control_step_start": 600,
+      "control_step_end": 749,
+      "policy_step_start": 100,
+      "policy_step_end": 124,
+      "phrase": "hovering gripper above table",
+      "description": "hovering gripper above table"
+    },
+    {
+      "segment_index": 5,
+      "t_start_sec": 14.970059880239521,
+      "t_end_sec": 17.964071856287426,
+      "control_step_start": 750,
+      "control_step_end": 899,
+      "policy_step_start": 125,
+      "policy_step_end": 149,
+      "phrase": "remaining idle near blocks",
+      "description": "remaining idle near blocks"
+    },
+    {
+      "segment_index": 6,
+      "t_start_sec": 17.964071856287426,
+      "t_end_sec": 19.580838323353294,
+      "control_step_start": 900,
+      "control_step_end": 980,
+      "policy_step_start": 150,
+      "policy_step_end": 163,
+      "phrase": "idle near wooden block",
+      "description": "idle near wooden block"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "placement_or_insertion_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 7.485029940119761,
+    "vlm_failure_onset_timestamp": "00:07.5",
+    "vlm_failure_onset_frame": 375,
+    "vlm_failure_onset_step": 375,
+    "vlm_coarse_failure_onset_seconds": 9.0,
+    "vlm_failure_window_start_seconds": 8.0,
+    "vlm_failure_window_end_seconds": 11.0,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot carries the yellow block to the wooden block but misaligns the placement, releasing the yellow block beside the wooden block onto the table surface rather than stacking it directly on top.",
+    "vlm_failure_reason_summary": "misaligned placement; yellow block released beside wooden block",
+    "vlm_recovery_action": "Re-grasp the yellow block from the table, lift it above the wooden block, align it properly with the top surface, and place it down securely.",
+    "vlm_recovery_action_summary": "re-grasp yellow block; center and stack on wooden block",
+    "vlm_prevention_action": "Position the yellow block directly over the center of the wooden block at a sufficient height before opening the gripper to release it.",
+    "vlm_prevention_action_summary": "center yellow block over target; release carefully onto surface",
+    "vlm_justification": "At around 9.0 seconds, as the robot attempts to place the yellow block on top of the wooden block, it lowers the gripper off-center, causing the yellow block to slip down and land on the table beside the wooden block.",
+    "vlm_temporal_justification": "At timestamp 01:14 (74 seconds into the clip), the robot begins rotating and tilting the yellow block away from proper vertical alignment above the wooden block, initiating the misaligned placement.",
+    "vlm_video_fps": 50.1,
+    "vlm_video_num_frames": 981,
+    "vlm_video_duration_seconds": 19.560878243512974,
+    "vlm_usage": {
+      "input_tokens": 22103,
+      "output_tokens": 391,
+      "cached_tokens": 1361
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"placement_or_insertion_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 9.0,\n \"failure_window_start_seconds\": 8.0,\n \"failure_window_end_seconds\": 11.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot carries the yellow block to the wooden block but misaligns the placement, releasing the yellow block beside the wooden block onto the table surface rather than stacking it directly on top.\",\n \"failure_reason_summary\": \"misaligned placement; yellow block released beside wooden block\",\n \"recovery_action\": \"Re-grasp the yellow block from the table, lift it above the wooden block, align it properly with the top surface, and place it down securely.\",\n \"recovery_action_summary\": \"re-grasp yellow block; center and stack on wooden block\",\n \"prevention_action\": \"Position the yellow block directly over the center of the wooden block at a sufficient height before opening the gripper to release it.\",\n \"prevention_action_summary\": \"center yellow block over target; release carefully onto surface\",\n \"justification\": \"At around 9.0 seconds, as the robot attempts to place the yellow block on top of the wooden block, it lowers the gripper off-center, causing the yellow block to slip down and land on the table beside the wooden block.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 375,
+    "failure_sim_step": 375,
+    "failure_policy_step": 62,
+    "failure_chunk_index": 3,
+    "first_post_failure_policy_step": 63,
+    "failure_type": "placement_or_insertion_failure",
+    "correction_action": "Re-grasp the yellow block from the table, lift it above the wooden block, align it properly with the top surface, and place it down securely."
+  }
+}
\ No newline at end of file
diff --git a/blocks/ep033/labels.npz b/blocks/ep033/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..5950e9bd7e6d260d0efaa507cf3a34b4003687d7
--- /dev/null
+++ b/blocks/ep033/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bfd059e45c236e8da3fbb4f7d2e4883745f96dba8cc6a39b7861e031e19cb543
+size 4065
diff --git a/blocks/ep034/example.md b/blocks/ep034/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..bbfcad5075d6e3ada6ac22db12e8e5f34ff28d73
--- /dev/null
+++ b/blocks/ep034/example.md
@@ -0,0 +1,37 @@
+# t000__ep034__blocks
+
+- **suite / variant:** `` / ``
+- **instruction:** Stack the yellow block on the wooden block
+- **success:** False
+- **control steps:** 930  (18.5s at 50 Hz)
+- **policy inferences:** 160  (replan_steps=6)
+- **video:** `blocks/ep034/rollout.mp4`
+- **features:** `160` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | reaching for and grasping yellow block |
+| 3.0–6.0 | 151–301 | 25–50 | lifting yellow block; misplacing on table |
+| 6.0–9.0 | 302–452 | 50–75 | releasing block; hovering low near table |
+| 9.0–12.0 | 453–603 | 75–100 | hovering near misplaced block; idle |
+| 12.0–15.0 | 604–754 | 100–125 | holding position near table; idle |
+| 15.0–18.0 | 755–905 | 125–150 | remaining stationary near block |
+| 18.0–18.5 | 906–929 | 151–154 | remaining stationary at rollout end |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 5.09920634920635 s (00:05.1)
+- **frame / control_step:** 257 (policy_step=42, chunk_index=5)
+- **coarse onset:** 6.0 s (refined=True)
+- **mode:** placement_or_insertion_failure
+- **reason:** The robot successfully picked up the yellow block and brought it toward the wooden block, but misplaced it off the edge of the wooden block onto the table instead of stacking it directly on top.
+  Summary: misaligned placement; yellow block placed on table beside wooden block
+- **recovery:** Re-grasp the yellow block from the table surface, lift it above the wooden block, align it over the top center of the wooden block, and release.
+  Summary: re-grasp yellow block; center and stack on wooden block
+- **prevention:** Maintain sufficient clearance height above the wooden block during approach and align the yellow block centered over the top surface before releasing.
+  Summary: maintain clearance; center yellow block over wooden block before release
+- **justification:** Around 6.0 seconds (frame 00:06), the robot lowers the yellow block, but misses the top surface of the wooden block and places it onto the table adjacent to it. The yellow block remains on the table for the rest of the rollout, failing the stacking goal.
diff --git a/blocks/ep034/labels.json b/blocks/ep034/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..f3bfc2289ed599aae926e0925a114ff36c7d87fc
--- /dev/null
+++ b/blocks/ep034/labels.json
@@ -0,0 +1,142 @@
+{
+  "rollout_id": "t000__ep034__blocks",
+  "scene_variant": "",
+  "task_id": 0,
+  "task_file": "",
+  "instruction": "Stack the yellow block on the wooden block",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:11:30+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.996031746031746,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for and grasping yellow block",
+      "description": "reaching for and grasping yellow block"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.996031746031746,
+      "t_end_sec": 5.992063492063492,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "lifting yellow block; misplacing on table",
+      "description": "lifting yellow block; misplacing on table"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.992063492063492,
+      "t_end_sec": 8.988095238095239,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "releasing block; hovering low near table",
+      "description": "releasing block; hovering low near table"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.988095238095239,
+      "t_end_sec": 11.984126984126984,
+      "control_step_start": 453,
+      "control_step_end": 603,
+      "policy_step_start": 75,
+      "policy_step_end": 100,
+      "phrase": "hovering near misplaced block; idle",
+      "description": "hovering near misplaced block; idle"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 11.984126984126984,
+      "t_end_sec": 14.980158730158731,
+      "control_step_start": 604,
+      "control_step_end": 754,
+      "policy_step_start": 100,
+      "policy_step_end": 125,
+      "phrase": "holding position near table; idle",
+      "description": "holding position near table; idle"
+    },
+    {
+      "segment_index": 5,
+      "t_start_sec": 14.980158730158731,
+      "t_end_sec": 17.976190476190478,
+      "control_step_start": 755,
+      "control_step_end": 905,
+      "policy_step_start": 125,
+      "policy_step_end": 150,
+      "phrase": "remaining stationary near block",
+      "description": "remaining stationary near block"
+    },
+    {
+      "segment_index": 6,
+      "t_start_sec": 17.976190476190478,
+      "t_end_sec": 18.452380952380953,
+      "control_step_start": 906,
+      "control_step_end": 929,
+      "policy_step_start": 151,
+      "policy_step_end": 154,
+      "phrase": "remaining stationary at rollout end",
+      "description": "remaining stationary at rollout end"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "placement_or_insertion_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 5.09920634920635,
+    "vlm_failure_onset_timestamp": "00:05.1",
+    "vlm_failure_onset_frame": 257,
+    "vlm_failure_onset_step": 257,
+    "vlm_coarse_failure_onset_seconds": 6.0,
+    "vlm_failure_window_start_seconds": 5.5,
+    "vlm_failure_window_end_seconds": 7.0,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot successfully picked up the yellow block and brought it toward the wooden block, but misplaced it off the edge of the wooden block onto the table instead of stacking it directly on top.",
+    "vlm_failure_reason_summary": "misaligned placement; yellow block placed on table beside wooden block",
+    "vlm_recovery_action": "Re-grasp the yellow block from the table surface, lift it above the wooden block, align it over the top center of the wooden block, and release.",
+    "vlm_recovery_action_summary": "re-grasp yellow block; center and stack on wooden block",
+    "vlm_prevention_action": "Maintain sufficient clearance height above the wooden block during approach and align the yellow block centered over the top surface before releasing.",
+    "vlm_prevention_action_summary": "maintain clearance; center yellow block over wooden block before release",
+    "vlm_justification": "Around 6.0 seconds (frame 00:06), the robot lowers the yellow block, but misses the top surface of the wooden block and places it onto the table adjacent to it. The yellow block remains on the table for the rest of the rollout, failing the stacking goal.",
+    "vlm_temporal_justification": "At timestamp 01:46 (106 seconds into the clip), the robot lowers and releases the yellow block directly onto the table surface next to the wooden block instead of positioning and placing it on top of the wooden block.",
+    "vlm_video_fps": 50.4,
+    "vlm_video_num_frames": 930,
+    "vlm_video_duration_seconds": 18.432539682539684,
+    "vlm_usage": {
+      "input_tokens": 22103,
+      "output_tokens": 412,
+      "cached_tokens": 1225
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"placement_or_insertion_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 6.0,\n \"failure_window_start_seconds\": 5.5,\n \"failure_window_end_seconds\": 7.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot successfully picked up the yellow block and brought it toward the wooden block, but misplaced it off the edge of the wooden block onto the table instead of stacking it directly on top.\",\n \"failure_reason_summary\": \"misaligned placement; yellow block placed on table beside wooden block\",\n \"recovery_action\": \"Re-grasp the yellow block from the table surface, lift it above the wooden block, align it over the top center of the wooden block, and release.\",\n \"recovery_action_summary\": \"re-grasp yellow block; center and stack on wooden block\",\n \"prevention_action\": \"Maintain sufficient clearance height above the wooden block during approach and align the yellow block centered over the top surface before releasing.\",\n \"prevention_action_summary\": \"maintain clearance; center yellow block over wooden block before release\",\n \"justification\": \"Around 6.0 seconds (frame 00:06), the robot lowers the yellow block, but misses the top surface of the wooden block and places it onto the table adjacent to it. The yellow block remains on the table for the rest of the rollout, failing the stacking goal.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 257,
+    "failure_sim_step": 257,
+    "failure_policy_step": 42,
+    "failure_chunk_index": 5,
+    "first_post_failure_policy_step": 43,
+    "failure_type": "placement_or_insertion_failure",
+    "correction_action": "Re-grasp the yellow block from the table surface, lift it above the wooden block, align it over the top center of the wooden block, and release."
+  }
+}
\ No newline at end of file
diff --git a/blocks/ep034/labels.npz b/blocks/ep034/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..d6895745f6806c43ee45107e7fe38f79f20278fd
--- /dev/null
+++ b/blocks/ep034/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ddf5ca2f20faacc379617c69f26032c6be45236b71cbf24a2403a5eb29f8593e
+size 4111
diff --git a/blocks/ep035/example.md b/blocks/ep035/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..0e503516a9b0c88434401fa10223777e3d7451fe
--- /dev/null
+++ b/blocks/ep035/example.md
@@ -0,0 +1,38 @@
+# t000__ep035__blocks
+
+- **suite / variant:** `` / ``
+- **instruction:** Stack the yellow block on the wooden block
+- **success:** False
+- **control steps:** 1165  (23.1s at 50 Hz)
+- **policy inferences:** 200  (replan_steps=6)
+- **video:** `blocks/ep035/rollout.mp4`
+- **features:** `200` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | approaching yellow block |
+| 3.0–6.0 | 151–301 | 25–50 | grasping yellow block |
+| 6.0–9.0 | 302–452 | 50–75 | lifting yellow block |
+| 9.0–12.0 | 453–603 | 75–100 | placing yellow block on wooden block |
+| 12.0–15.0 | 604–754 | 100–125 | holding block; remaining stationary |
+| 15.0–18.0 | 755–905 | 125–150 | idle; holding block |
+| 18.0–21.0 | 906–1056 | 151–176 | idle / stationary position |
+| 21.0–23.1 | 1057–1164 | 176–194 | idle / no progress |
+
+## Failure
+
+- **yes**
+- **onset type:** operator_intervention
+- **onset:** 12.063492063492063 s (00:12.1)
+- **frame / control_step:** 608 (policy_step=101, chunk_index=2)
+- **coarse onset:** 13.0 s (refined=True)
+- **mode:** stuck_or_no_progress
+- **reason:** The robot moved the yellow block over the wooden block and lowered it, but then froze without opening its gripper to release the object.
+  Summary: robot freezes; gripper fails to open
+- **recovery:** Open the gripper to release the yellow block onto the wooden block, then retract the end-effector.
+  Summary: open gripper; release yellow block; retract arm
+- **prevention:** Command the gripper to open and the arm to retract immediately after reaching the target placement pose above the wooden block.
+  Summary: open gripper after placement; retract arm
+- **justification:** By timestamp 00:13, the robot has positioned the yellow block directly on top of the wooden block, but it freezes in place and keeps its gripper clamped shut for the remainder of the episode without completing the release.
diff --git a/blocks/ep035/labels.json b/blocks/ep035/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..fcd8128f95aa0c080ec56bddcabb61b028e8a6b9
--- /dev/null
+++ b/blocks/ep035/labels.json
@@ -0,0 +1,153 @@
+{
+  "rollout_id": "t000__ep035__blocks",
+  "scene_variant": "",
+  "task_id": 0,
+  "task_file": "",
+  "instruction": "Stack the yellow block on the wooden block",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:12:09+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.996031746031746,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "approaching yellow block",
+      "description": "approaching yellow block"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.996031746031746,
+      "t_end_sec": 5.992063492063492,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "grasping yellow block",
+      "description": "grasping yellow block"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.992063492063492,
+      "t_end_sec": 8.988095238095239,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "lifting yellow block",
+      "description": "lifting yellow block"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.988095238095239,
+      "t_end_sec": 11.984126984126984,
+      "control_step_start": 453,
+      "control_step_end": 603,
+      "policy_step_start": 75,
+      "policy_step_end": 100,
+      "phrase": "placing yellow block on wooden block",
+      "description": "placing yellow block on wooden block"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 11.984126984126984,
+      "t_end_sec": 14.980158730158731,
+      "control_step_start": 604,
+      "control_step_end": 754,
+      "policy_step_start": 100,
+      "policy_step_end": 125,
+      "phrase": "holding block; remaining stationary",
+      "description": "holding block; remaining stationary"
+    },
+    {
+      "segment_index": 5,
+      "t_start_sec": 14.980158730158731,
+      "t_end_sec": 17.976190476190478,
+      "control_step_start": 755,
+      "control_step_end": 905,
+      "policy_step_start": 125,
+      "policy_step_end": 150,
+      "phrase": "idle; holding block",
+      "description": "idle; holding block"
+    },
+    {
+      "segment_index": 6,
+      "t_start_sec": 17.976190476190478,
+      "t_end_sec": 20.97222222222222,
+      "control_step_start": 906,
+      "control_step_end": 1056,
+      "policy_step_start": 151,
+      "policy_step_end": 176,
+      "phrase": "idle / stationary position",
+      "description": "idle / stationary position"
+    },
+    {
+      "segment_index": 7,
+      "t_start_sec": 20.97222222222222,
+      "t_end_sec": 23.115079365079367,
+      "control_step_start": 1057,
+      "control_step_end": 1164,
+      "policy_step_start": 176,
+      "policy_step_end": 194,
+      "phrase": "idle / no progress",
+      "description": "idle / no progress"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "stuck_or_no_progress",
+    "vlm_failure_onset_type": "operator_intervention",
+    "vlm_failure_onset_seconds": 12.063492063492063,
+    "vlm_failure_onset_timestamp": "00:12.1",
+    "vlm_failure_onset_frame": 608,
+    "vlm_failure_onset_step": 608,
+    "vlm_coarse_failure_onset_seconds": 13.0,
+    "vlm_failure_window_start_seconds": 12.0,
+    "vlm_failure_window_end_seconds": 14.0,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot moved the yellow block over the wooden block and lowered it, but then froze without opening its gripper to release the object.",
+    "vlm_failure_reason_summary": "robot freezes; gripper fails to open",
+    "vlm_recovery_action": "Open the gripper to release the yellow block onto the wooden block, then retract the end-effector.",
+    "vlm_recovery_action_summary": "open gripper; release yellow block; retract arm",
+    "vlm_prevention_action": "Command the gripper to open and the arm to retract immediately after reaching the target placement pose above the wooden block.",
+    "vlm_prevention_action_summary": "open gripper after placement; retract arm",
+    "vlm_justification": "By timestamp 00:13, the robot has positioned the yellow block directly on top of the wooden block, but it freezes in place and keeps its gripper clamped shut for the remainder of the episode without completing the release.",
+    "vlm_temporal_justification": "By second 104 of the clip, the robot arm completes its descent, placing the yellow block directly onto the wooden block, and comes to a complete stop. From this point forward, it remains frozen without opening its gripper or making further progress.",
+    "vlm_video_fps": 50.4,
+    "vlm_video_num_frames": 1165,
+    "vlm_video_duration_seconds": 23.095238095238095,
+    "vlm_usage": {
+      "input_tokens": 22407,
+      "output_tokens": 367,
+      "cached_tokens": 1565
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"stuck_or_no_progress\",\n \"onset_type\": \"operator_intervention\",\n \"failure_onset_seconds\": 13.0,\n \"failure_window_start_seconds\": 12.0,\n \"failure_window_end_seconds\": 14.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot moved the yellow block over the wooden block and lowered it, but then froze without opening its gripper to release the object.\",\n \"failure_reason_summary\": \"robot freezes; gripper fails to open\",\n \"recovery_action\": \"Open the gripper to release the yellow block onto the wooden block, then retract the end-effector.\",\n \"recovery_action_summary\": \"open gripper; release yellow block; retract arm\",\n \"prevention_action\": \"Command the gripper to open and the arm to retract immediately after reaching the target placement pose above the wooden block.\",\n \"prevention_action_summary\": \"open gripper after placement; retract arm\",\n \"justification\": \"By timestamp 00:13, the robot has positioned the yellow block directly on top of the wooden block, but it freezes in place and keeps its gripper clamped shut for the remainder of the episode without completing the release.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 608,
+    "failure_sim_step": 608,
+    "failure_policy_step": 101,
+    "failure_chunk_index": 2,
+    "first_post_failure_policy_step": 102,
+    "failure_type": "stuck_or_no_progress",
+    "correction_action": "Open the gripper to release the yellow block onto the wooden block, then retract the end-effector."
+  }
+}
\ No newline at end of file
diff --git a/blocks/ep035/labels.npz b/blocks/ep035/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..c348992b16107aa7896e99c40c3e97a0a08f6ec3
--- /dev/null
+++ b/blocks/ep035/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a2a0fa02cef67cd659b40bdda861d2692c009c812cf4acc7ded403d99ff1850c
+size 3899
diff --git a/blocks/ep036/example.md b/blocks/ep036/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..8726f5e0d543f4e06ac34847b078649dea5d4bb1
--- /dev/null
+++ b/blocks/ep036/example.md
@@ -0,0 +1,35 @@
+# t000__ep036__blocks
+
+- **suite / variant:** `` / ``
+- **instruction:** Stack the yellow block on the wooden block
+- **success:** False
+- **control steps:** 618  (12.3s at 50 Hz)
+- **policy inferences:** 108  (replan_steps=6)
+- **video:** `blocks/ep036/rollout.mp4`
+- **features:** `108` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | reaching for yellow block |
+| 3.0–6.0 | 151–301 | 25–50 | closing gripper and lifting empty |
+| 6.0–9.0 | 302–452 | 50–75 | moving toward wooden block |
+| 9.0–12.0 | 453–603 | 75–100 | grasping wooden block |
+| 12.0–12.3 | 604–617 | 100–102 | holding wooden block |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 4.820717131474104 s (00:04.8)
+- **frame / control_step:** 242 (policy_step=40, chunk_index=2)
+- **coarse onset:** 4.5 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the yellow block but failed to secure its hold, lifting up empty before moving away to grab the wooden block instead.
+  Summary: failed grasp on yellow block; empty end-effector lift
+- **recovery:** Re-align the end-effector over the yellow block, close the gripper securely around it, and lift it to stack on top of the wooden block.
+  Summary: re-grasp yellow block; stack onto wooden block
+- **prevention:** Ensure precise alignment and sufficient gripper contact around the yellow block before executing the lift to prevent slipping.
+  Summary: align gripper centered; secure grasp before lifting
+- **justification:** None
diff --git a/blocks/ep036/labels.json b/blocks/ep036/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..42d4e0bbc418bc1413664b2ab743cd0001bb1d8b
--- /dev/null
+++ b/blocks/ep036/labels.json
@@ -0,0 +1,120 @@
+{
+  "rollout_id": "t000__ep036__blocks",
+  "scene_variant": "",
+  "task_id": 0,
+  "task_file": "",
+  "instruction": "Stack the yellow block on the wooden block",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:12:39+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00796812749004,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for yellow block",
+      "description": "reaching for yellow block"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00796812749004,
+      "t_end_sec": 6.01593625498008,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "closing gripper and lifting empty",
+      "description": "closing gripper and lifting empty"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01593625498008,
+      "t_end_sec": 9.02390438247012,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "moving toward wooden block",
+      "description": "moving toward wooden block"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.02390438247012,
+      "t_end_sec": 12.03187250996016,
+      "control_step_start": 453,
+      "control_step_end": 603,
+      "policy_step_start": 75,
+      "policy_step_end": 100,
+      "phrase": "grasping wooden block",
+      "description": "grasping wooden block"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 12.03187250996016,
+      "t_end_sec": 12.310756972111554,
+      "control_step_start": 604,
+      "control_step_end": 617,
+      "policy_step_start": 100,
+      "policy_step_end": 102,
+      "phrase": "holding wooden block",
+      "description": "holding wooden block"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 4.820717131474104,
+    "vlm_failure_onset_timestamp": "00:04.8",
+    "vlm_failure_onset_frame": 242,
+    "vlm_failure_onset_step": 242,
+    "vlm_coarse_failure_onset_seconds": 4.5,
+    "vlm_failure_window_start_seconds": 4.0,
+    "vlm_failure_window_end_seconds": 5.0,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the yellow block but failed to secure its hold, lifting up empty before moving away to grab the wooden block instead.",
+    "vlm_failure_reason_summary": "failed grasp on yellow block; empty end-effector lift",
+    "vlm_recovery_action": "Re-align the end-effector over the yellow block, close the gripper securely around it, and lift it to stack on top of the wooden block.",
+    "vlm_recovery_action_summary": "re-grasp yellow block; stack onto wooden block",
+    "vlm_prevention_action": "Ensure precise alignment and sufficient gripper contact around the yellow block before executing the lift to prevent slipping.",
+    "vlm_prevention_action_summary": "align gripper centered; secure grasp before lifting",
+    "vlm_justification": null,
+    "vlm_temporal_justification": "The robot's gripper closes fully at 02:47 without securing the yellow block, leaving it sitting untouched on the table as the arm prepares to lift up empty.",
+    "vlm_video_fps": 50.2,
+    "vlm_video_num_frames": 618,
+    "vlm_video_duration_seconds": 12.290836653386453,
+    "vlm_usage": {
+      "input_tokens": 21708,
+      "output_tokens": 310,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 4.5,\n \"failure_window_start_seconds\": 4.0,\n \"failure_window_end_seconds\": 5.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the yellow block but failed to secure its hold, lifting up empty before moving away to grab the wooden block instead.\",\n \"failure_reason_summary\": \"failed grasp on yellow block; empty end-effector lift\",\n \"recovery_action\": \"Re-align the end-effector over the yellow block, close the gripper securely around it, and lift it to stack on top of the wooden block.\",\n \"recovery_action_summary\": \"re-grasp yellow block; stack onto wooden block\",\n \"prevention_action\": \"Ensure precise alignment and sufficient gripper contact around the yellow block before executing the lift to prevent slipping.\",\n \"prevention_action_summary\": \"align gripper centered; secure grasp before lifting\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 242,
+    "failure_sim_step": 242,
+    "failure_policy_step": 40,
+    "failure_chunk_index": 2,
+    "first_post_failure_policy_step": 41,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-align the end-effector over the yellow block, close the gripper securely around it, and lift it to stack on top of the wooden block."
+  }
+}
\ No newline at end of file
diff --git a/blocks/ep036/labels.npz b/blocks/ep036/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..ed89401fd8de09103b2ec739e8ee1da7573751ef
--- /dev/null
+++ b/blocks/ep036/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:401fe0317ee7fa53c4ca976d652726219c78be015a333bd62587b8a6d6ac3c2e
+size 3880
diff --git a/blocks/ep037/example.md b/blocks/ep037/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..5c3eaff44ec11359cf64d3e653627200b7ecca20
--- /dev/null
+++ b/blocks/ep037/example.md
@@ -0,0 +1,24 @@
+# t000__ep037__blocks
+
+- **suite / variant:** `` / ``
+- **instruction:** Stack the yellow block on the wooden block
+- **success:** True
+- **control steps:** 871  (17.3s at 50 Hz)
+- **policy inferences:** 149  (replan_steps=6)
+- **video:** `blocks/ep037/rollout.mp4`
+- **features:** `149` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | reaching for yellow block |
+| 3.0–6.0 | 151–301 | 25–50 | grasping and lifting yellow block |
+| 6.0–9.0 | 302–452 | 50–75 | moving yellow block over wooden block |
+| 9.0–12.0 | 453–603 | 75–100 | lowering yellow block onto wooden block |
+| 12.0–15.0 | 604–754 | 100–125 | releasing block and lifting gripper |
+| 15.0–17.3 | 755–870 | 125–145 | retracting arm |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/blocks/ep037/labels.json b/blocks/ep037/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..e8779da3ba03670144515a3e125d060ff00f0ef5
--- /dev/null
+++ b/blocks/ep037/labels.json
@@ -0,0 +1,101 @@
+{
+  "rollout_id": "t000__ep037__blocks",
+  "scene_variant": "",
+  "task_id": 0,
+  "task_file": "",
+  "instruction": "Stack the yellow block on the wooden block",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:12:48+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.996031746031746,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for yellow block",
+      "description": "reaching for yellow block"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.996031746031746,
+      "t_end_sec": 5.992063492063492,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "grasping and lifting yellow block",
+      "description": "grasping and lifting yellow block"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.992063492063492,
+      "t_end_sec": 8.988095238095239,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "moving yellow block over wooden block",
+      "description": "moving yellow block over wooden block"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.988095238095239,
+      "t_end_sec": 11.984126984126984,
+      "control_step_start": 453,
+      "control_step_end": 603,
+      "policy_step_start": 75,
+      "policy_step_end": 100,
+      "phrase": "lowering yellow block onto wooden block",
+      "description": "lowering yellow block onto wooden block"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 11.984126984126984,
+      "t_end_sec": 14.980158730158731,
+      "control_step_start": 604,
+      "control_step_end": 754,
+      "policy_step_start": 100,
+      "policy_step_end": 125,
+      "phrase": "releasing block and lifting gripper",
+      "description": "releasing block and lifting gripper"
+    },
+    {
+      "segment_index": 5,
+      "t_start_sec": 14.980158730158731,
+      "t_end_sec": 17.28174603174603,
+      "control_step_start": 755,
+      "control_step_end": 870,
+      "policy_step_start": 125,
+      "policy_step_end": 145,
+      "phrase": "retracting arm",
+      "description": "retracting arm"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/blocks/ep037/labels.npz b/blocks/ep037/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..80f7c58f2e5df2cc055d61d4b8cdcf204006151d
--- /dev/null
+++ b/blocks/ep037/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6e4664c5ff0de97acf3b5dbda41981876c9dd661694013fccacf8f482b612a19
+size 3200
diff --git a/blocks/ep038/example.md b/blocks/ep038/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..35274cf26de99ad770f39b7c8019016f5a4fefcf
--- /dev/null
+++ b/blocks/ep038/example.md
@@ -0,0 +1,42 @@
+# t000__ep038__blocks
+
+- **suite / variant:** `` / ``
+- **instruction:** Stack the yellow block on the wooden block
+- **success:** False
+- **control steps:** 1749  (34.7s at 50 Hz)
+- **policy inferences:** 301  (replan_steps=6)
+- **video:** `blocks/ep038/rollout.mp4`
+- **features:** `301` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | reaching for yellow block |
+| 3.0–6.0 | 151–301 | 25–50 | grasping and lifting yellow block |
+| 6.0–9.0 | 302–452 | 50–75 | carrying yellow block to target |
+| 9.0–12.0 | 453–603 | 75–100 | dropping yellow block beside wooden block |
+| 12.0–15.0 | 604–754 | 100–125 | hovering gripper near blocks |
+| 15.0–18.0 | 755–905 | 125–150 | holding position above blocks |
+| 18.0–21.0 | 906–1056 | 151–176 | oscillating slight gripper movements |
+| 21.0–24.0 | 1057–1207 | 176–201 | remaining stationary above wooden block |
+| 24.0–27.0 | 1208–1358 | 201–226 | idle hovering near blocks |
+| 27.0–30.0 | 1359–1509 | 226–251 | holding position over wooden block |
+| 30.0–33.0 | 1510–1660 | 251–276 | remaining idle above blocks |
+| 33.0–34.7 | 1661–1748 | 276–291 | stationary near wooden block |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 8.174603174603174 s (00:08.2)
+- **frame / control_step:** 412 (policy_step=68, chunk_index=4)
+- **coarse onset:** 10.5 s (refined=True)
+- **mode:** placement_or_insertion_failure
+- **reason:** The robot brought the yellow block over the wooden block but misaligned the placement, causing the yellow block to slip off the top of the wooden block rather than resting stably on top.
+  Summary: misaligned placement; yellow block slipping off target
+- **recovery:** Re-grasp the yellow block, lift it above the wooden block, carefully align it over the top surface, and release it.
+  Summary: re-grasp yellow block; align and stack on wooden block
+- **prevention:** Ensure the yellow block is positioned directly over the center of the wooden block before releasing the gripper.
+  Summary: center yellow block over wooden block; release carefully
+- **justification:** At around 10.5 seconds, the robot attempts to place the yellow block onto the wooden block, but because of poor alignment, the block slides off to the side rather than remaining stacked on top.
diff --git a/blocks/ep038/labels.json b/blocks/ep038/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..eb0e907faaa3593aca7bafb758e63b3c21f2a7be
--- /dev/null
+++ b/blocks/ep038/labels.json
@@ -0,0 +1,197 @@
+{
+  "rollout_id": "t000__ep038__blocks",
+  "scene_variant": "",
+  "task_id": 0,
+  "task_file": "",
+  "instruction": "Stack the yellow block on the wooden block",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:13:24+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.996031746031746,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for yellow block",
+      "description": "reaching for yellow block"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.996031746031746,
+      "t_end_sec": 5.992063492063492,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "grasping and lifting yellow block",
+      "description": "grasping and lifting yellow block"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.992063492063492,
+      "t_end_sec": 8.988095238095239,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "carrying yellow block to target",
+      "description": "carrying yellow block to target"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.988095238095239,
+      "t_end_sec": 11.984126984126984,
+      "control_step_start": 453,
+      "control_step_end": 603,
+      "policy_step_start": 75,
+      "policy_step_end": 100,
+      "phrase": "dropping yellow block beside wooden block",
+      "description": "dropping yellow block beside wooden block"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 11.984126984126984,
+      "t_end_sec": 14.980158730158731,
+      "control_step_start": 604,
+      "control_step_end": 754,
+      "policy_step_start": 100,
+      "policy_step_end": 125,
+      "phrase": "hovering gripper near blocks",
+      "description": "hovering gripper near blocks"
+    },
+    {
+      "segment_index": 5,
+      "t_start_sec": 14.980158730158731,
+      "t_end_sec": 17.976190476190478,
+      "control_step_start": 755,
+      "control_step_end": 905,
+      "policy_step_start": 125,
+      "policy_step_end": 150,
+      "phrase": "holding position above blocks",
+      "description": "holding position above blocks"
+    },
+    {
+      "segment_index": 6,
+      "t_start_sec": 17.976190476190478,
+      "t_end_sec": 20.97222222222222,
+      "control_step_start": 906,
+      "control_step_end": 1056,
+      "policy_step_start": 151,
+      "policy_step_end": 176,
+      "phrase": "oscillating slight gripper movements",
+      "description": "oscillating slight gripper movements"
+    },
+    {
+      "segment_index": 7,
+      "t_start_sec": 20.97222222222222,
+      "t_end_sec": 23.96825396825397,
+      "control_step_start": 1057,
+      "control_step_end": 1207,
+      "policy_step_start": 176,
+      "policy_step_end": 201,
+      "phrase": "remaining stationary above wooden block",
+      "description": "remaining stationary above wooden block"
+    },
+    {
+      "segment_index": 8,
+      "t_start_sec": 23.96825396825397,
+      "t_end_sec": 26.964285714285715,
+      "control_step_start": 1208,
+      "control_step_end": 1358,
+      "policy_step_start": 201,
+      "policy_step_end": 226,
+      "phrase": "idle hovering near blocks",
+      "description": "idle hovering near blocks"
+    },
+    {
+      "segment_index": 9,
+      "t_start_sec": 26.964285714285715,
+      "t_end_sec": 29.960317460317462,
+      "control_step_start": 1359,
+      "control_step_end": 1509,
+      "policy_step_start": 226,
+      "policy_step_end": 251,
+      "phrase": "holding position over wooden block",
+      "description": "holding position over wooden block"
+    },
+    {
+      "segment_index": 10,
+      "t_start_sec": 29.960317460317462,
+      "t_end_sec": 32.95634920634921,
+      "control_step_start": 1510,
+      "control_step_end": 1660,
+      "policy_step_start": 251,
+      "policy_step_end": 276,
+      "phrase": "remaining idle above blocks",
+      "description": "remaining idle above blocks"
+    },
+    {
+      "segment_index": 11,
+      "t_start_sec": 32.95634920634921,
+      "t_end_sec": 34.702380952380956,
+      "control_step_start": 1661,
+      "control_step_end": 1748,
+      "policy_step_start": 276,
+      "policy_step_end": 291,
+      "phrase": "stationary near wooden block",
+      "description": "stationary near wooden block"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "placement_or_insertion_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 8.174603174603174,
+    "vlm_failure_onset_timestamp": "00:08.2",
+    "vlm_failure_onset_frame": 412,
+    "vlm_failure_onset_step": 412,
+    "vlm_coarse_failure_onset_seconds": 10.5,
+    "vlm_failure_window_start_seconds": 10.0,
+    "vlm_failure_window_end_seconds": 12.0,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot brought the yellow block over the wooden block but misaligned the placement, causing the yellow block to slip off the top of the wooden block rather than resting stably on top.",
+    "vlm_failure_reason_summary": "misaligned placement; yellow block slipping off target",
+    "vlm_recovery_action": "Re-grasp the yellow block, lift it above the wooden block, carefully align it over the top surface, and release it.",
+    "vlm_recovery_action_summary": "re-grasp yellow block; align and stack on wooden block",
+    "vlm_prevention_action": "Ensure the yellow block is positioned directly over the center of the wooden block before releasing the gripper.",
+    "vlm_prevention_action_summary": "center yellow block over wooden block; release carefully",
+    "vlm_justification": "At around 10.5 seconds, the robot attempts to place the yellow block onto the wooden block, but because of poor alignment, the block slides off to the side rather than remaining stacked on top.",
+    "vlm_temporal_justification": "At second 34 (00:34), the yellow block makes contact with the edge of the wooden block in a clearly misaligned position, causing it to tilt and slide off rather than rest stably on top.",
+    "vlm_video_fps": 50.4,
+    "vlm_video_num_frames": 1749,
+    "vlm_video_duration_seconds": 34.682539682539684,
+    "vlm_usage": {
+      "input_tokens": 23174,
+      "output_tokens": 377,
+      "cached_tokens": 2381
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"placement_or_insertion_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 10.5,\n \"failure_window_start_seconds\": 10.0,\n \"failure_window_end_seconds\": 12.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot brought the yellow block over the wooden block but misaligned the placement, causing the yellow block to slip off the top of the wooden block rather than resting stably on top.\",\n \"failure_reason_summary\": \"misaligned placement; yellow block slipping off target\",\n \"recovery_action\": \"Re-grasp the yellow block, lift it above the wooden block, carefully align it over the top surface, and release it.\",\n \"recovery_action_summary\": \"re-grasp yellow block; align and stack on wooden block\",\n \"prevention_action\": \"Ensure the yellow block is positioned directly over the center of the wooden block before releasing the gripper.\",\n \"prevention_action_summary\": \"center yellow block over wooden block; release carefully\",\n \"justification\": \"At around 10.5 seconds, the robot attempts to place the yellow block onto the wooden block, but because of poor alignment, the block slides off to the side rather than remaining stacked on top.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 412,
+    "failure_sim_step": 412,
+    "failure_policy_step": 68,
+    "failure_chunk_index": 4,
+    "first_post_failure_policy_step": 69,
+    "failure_type": "placement_or_insertion_failure",
+    "correction_action": "Re-grasp the yellow block, lift it above the wooden block, carefully align it over the top surface, and release it."
+  }
+}
\ No newline at end of file
diff --git a/blocks/ep038/labels.npz b/blocks/ep038/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..0cd5a3a3fcdf2d013402999f6c7b7b7f93465b43
--- /dev/null
+++ b/blocks/ep038/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8803b2381bf0cddac8f4b7e29d22500a7637fe1bdffe055551a8ceee958e21d5
+size 4173
diff --git a/blocks/ep039/example.md b/blocks/ep039/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..208a9736750439f15aec520b4b92cf33fbdab6cb
--- /dev/null
+++ b/blocks/ep039/example.md
@@ -0,0 +1,35 @@
+# t000__ep039__blocks
+
+- **suite / variant:** `` / ``
+- **instruction:** Stack the yellow block on the wooden block
+- **success:** False
+- **control steps:** 635  (12.5s at 51 Hz)
+- **policy inferences:** 109  (replan_steps=6)
+- **video:** `blocks/ep039/rollout.mp4`
+- **features:** `109` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching for yellow block |
+| 3.0–6.0 | 152–303 | 25–50 | grasping yellow block |
+| 6.0–9.0 | 304–455 | 50–75 | moving block toward wooden block |
+| 9.0–12.0 | 456–607 | 76–101 | releasing block beside wooden block |
+| 12.0–12.5 | 608–634 | 101–105 | retracting end-effector |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 10.296442687747035 s (00:10.3)
+- **frame / control_step:** 521 (policy_step=86, chunk_index=5)
+- **coarse onset:** 8.5 s (refined=True)
+- **mode:** placement_or_insertion_failure
+- **reason:** The robot moved the yellow block toward the wooden block but released it off-center, causing the yellow block to slide off and land on the table surface next to the wooden block.
+  Summary: misaligned placement; yellow block dropped beside target
+- **recovery:** Re-grasp the yellow block from the table, position it directly above the center of the wooden block, and release it so it rests securely on top.
+  Summary: re-grasp yellow block; stack onto wooden block
+- **prevention:** Ensure the robot aligns the yellow block directly over the top face of the wooden block before releasing the gripper fingers.
+  Summary: center yellow block over target; release carefully
+- **justification:** The robot successfully picks up the yellow block and carries it over to the wooden block. At approximately 8.5 seconds, it opens the gripper while the yellow block is not centered over the wooden block, resulting in the block falling to the side onto the table.
diff --git a/blocks/ep039/labels.json b/blocks/ep039/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..e2c4e014c033a50703295a2b5554839ec29493e6
--- /dev/null
+++ b/blocks/ep039/labels.json
@@ -0,0 +1,120 @@
+{
+  "rollout_id": "t000__ep039__blocks",
+  "scene_variant": "",
+  "task_id": 0,
+  "task_file": "",
+  "instruction": "Stack the yellow block on the wooden block",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:13:57+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0039525691699605,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for yellow block",
+      "description": "reaching for yellow block"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0039525691699605,
+      "t_end_sec": 6.007905138339921,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "grasping yellow block",
+      "description": "grasping yellow block"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.007905138339921,
+      "t_end_sec": 9.011857707509881,
+      "control_step_start": 304,
+      "control_step_end": 455,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "moving block toward wooden block",
+      "description": "moving block toward wooden block"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.011857707509881,
+      "t_end_sec": 12.015810276679842,
+      "control_step_start": 456,
+      "control_step_end": 607,
+      "policy_step_start": 76,
+      "policy_step_end": 101,
+      "phrase": "releasing block beside wooden block",
+      "description": "releasing block beside wooden block"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 12.015810276679842,
+      "t_end_sec": 12.549407114624506,
+      "control_step_start": 608,
+      "control_step_end": 634,
+      "policy_step_start": 101,
+      "policy_step_end": 105,
+      "phrase": "retracting end-effector",
+      "description": "retracting end-effector"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "placement_or_insertion_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 10.296442687747035,
+    "vlm_failure_onset_timestamp": "00:10.3",
+    "vlm_failure_onset_frame": 521,
+    "vlm_failure_onset_step": 521,
+    "vlm_coarse_failure_onset_seconds": 8.5,
+    "vlm_failure_window_start_seconds": 8.0,
+    "vlm_failure_window_end_seconds": 10.0,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot moved the yellow block toward the wooden block but released it off-center, causing the yellow block to slide off and land on the table surface next to the wooden block.",
+    "vlm_failure_reason_summary": "misaligned placement; yellow block dropped beside target",
+    "vlm_recovery_action": "Re-grasp the yellow block from the table, position it directly above the center of the wooden block, and release it so it rests securely on top.",
+    "vlm_recovery_action_summary": "re-grasp yellow block; stack onto wooden block",
+    "vlm_prevention_action": "Ensure the robot aligns the yellow block directly over the top face of the wooden block before releasing the gripper fingers.",
+    "vlm_prevention_action_summary": "center yellow block over target; release carefully",
+    "vlm_justification": "The robot successfully picks up the yellow block and carries it over to the wooden block. At approximately 8.5 seconds, it opens the gripper while the yellow block is not centered over the wooden block, resulting in the block falling to the side onto the table.",
+    "vlm_temporal_justification": "At 04:03 (243 seconds into the clip), the robot opens its gripper to release the yellow block while positioned off-center over the edge of the wooden block, causing the yellow block to slide off onto the table surface.",
+    "vlm_video_fps": 50.6,
+    "vlm_video_num_frames": 635,
+    "vlm_video_duration_seconds": 12.529644268774703,
+    "vlm_usage": {
+      "input_tokens": 21912,
+      "output_tokens": 398,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"placement_or_insertion_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 8.5,\n \"failure_window_start_seconds\": 8.0,\n \"failure_window_end_seconds\": 10.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot moved the yellow block toward the wooden block but released it off-center, causing the yellow block to slide off and land on the table surface next to the wooden block.\",\n \"failure_reason_summary\": \"misaligned placement; yellow block dropped beside target\",\n \"recovery_action\": \"Re-grasp the yellow block from the table, position it directly above the center of the wooden block, and release it so it rests securely on top.\",\n \"recovery_action_summary\": \"re-grasp yellow block; stack onto wooden block\",\n \"prevention_action\": \"Ensure the robot aligns the yellow block directly over the top face of the wooden block before releasing the gripper fingers.\",\n \"prevention_action_summary\": \"center yellow block over target; release carefully\",\n \"justification\": \"The robot successfully picks up the yellow block and carries it over to the wooden block. At approximately 8.5 seconds, it opens the gripper while the yellow block is not centered over the wooden block, resulting in the block falling to the side onto the table.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 521,
+    "failure_sim_step": 521,
+    "failure_policy_step": 86,
+    "failure_chunk_index": 5,
+    "first_post_failure_policy_step": 87,
+    "failure_type": "placement_or_insertion_failure",
+    "correction_action": "Re-grasp the yellow block from the table, position it directly above the center of the wooden block, and release it so it rests securely on top."
+  }
+}
\ No newline at end of file
diff --git a/blocks/ep039/labels.npz b/blocks/ep039/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..ee88420ed0e2568b9f24b0fada39150ad5ee37b9
--- /dev/null
+++ b/blocks/ep039/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1a83fd4724e3666cd80eeb5b731cae2d5965e095181ee3d07b35bf4b86710bd9
+size 3918
diff --git a/blocks/ep040/example.md b/blocks/ep040/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..27d765453748717966549d042bf06602c234a5f6
--- /dev/null
+++ b/blocks/ep040/example.md
@@ -0,0 +1,33 @@
+# t000__ep040__blocks
+
+- **suite / variant:** `` / ``
+- **instruction:** Stack the yellow block on the wooden block
+- **success:** False
+- **control steps:** 426  (8.4s at 51 Hz)
+- **policy inferences:** 74  (replan_steps=6)
+- **video:** `blocks/ep040/rollout.mp4`
+- **features:** `74` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | lowering gripper toward yellow block |
+| 3.0–6.0 | 152–303 | 25–50 | closing gripper and lifting empty |
+| 6.0–8.4 | 304–425 | 50–70 | moving empty gripper over wooden block |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 4.714003944773175 s (00:04.7)
+- **frame / control_step:** 239 (policy_step=39, chunk_index=5)
+- **coarse onset:** 4.2 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the yellow block, but failed to secure its grip, raising its empty gripper into the air while leaving the block on the table.
+  Summary: failed grasp on yellow block; lifting empty gripper
+- **recovery:** Move the arm back over the yellow block, lower the gripper around the block, close the fingers securely, and then pick it up.
+  Summary: re-align gripper over yellow block; grasp and lift yellow block
+- **prevention:** Ensure the gripper is properly aligned and centered around the yellow block before closing the fingers to achieve a firm grasp before lifting.
+  Summary: center gripper on yellow block; secure firm grip before lifting
+- **justification:** At 00:04 (4.0s), the robot closes its gripper on the yellow block, but fails to grip it securely. By 00:05 (5.0s), the arm lifts up leaving the yellow block behind on the table, continuing to execute the stacking motion with an empty gripper.
diff --git a/blocks/ep040/labels.json b/blocks/ep040/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..b4a535cd3912e90b2bdcb80d82a20ef0815c60b2
--- /dev/null
+++ b/blocks/ep040/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t000__ep040__blocks",
+  "scene_variant": "",
+  "task_id": 0,
+  "task_file": "",
+  "instruction": "Stack the yellow block on the wooden block",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:14:31+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9980276134122286,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "lowering gripper toward yellow block",
+      "description": "lowering gripper toward yellow block"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9980276134122286,
+      "t_end_sec": 5.996055226824457,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "closing gripper and lifting empty",
+      "description": "closing gripper and lifting empty"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.996055226824457,
+      "t_end_sec": 8.402366863905325,
+      "control_step_start": 304,
+      "control_step_end": 425,
+      "policy_step_start": 50,
+      "policy_step_end": 70,
+      "phrase": "moving empty gripper over wooden block",
+      "description": "moving empty gripper over wooden block"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 4.714003944773175,
+    "vlm_failure_onset_timestamp": "00:04.7",
+    "vlm_failure_onset_frame": 239,
+    "vlm_failure_onset_step": 239,
+    "vlm_coarse_failure_onset_seconds": 4.2,
+    "vlm_failure_window_start_seconds": 3.5,
+    "vlm_failure_window_end_seconds": 5.0,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the yellow block, but failed to secure its grip, raising its empty gripper into the air while leaving the block on the table.",
+    "vlm_failure_reason_summary": "failed grasp on yellow block; lifting empty gripper",
+    "vlm_recovery_action": "Move the arm back over the yellow block, lower the gripper around the block, close the fingers securely, and then pick it up.",
+    "vlm_recovery_action_summary": "re-align gripper over yellow block; grasp and lift yellow block",
+    "vlm_prevention_action": "Ensure the gripper is properly aligned and centered around the yellow block before closing the fingers to achieve a firm grasp before lifting.",
+    "vlm_prevention_action_summary": "center gripper on yellow block; secure firm grip before lifting",
+    "vlm_justification": "At 00:04 (4.0s), the robot closes its gripper on the yellow block, but fails to grip it securely. By 00:05 (5.0s), the arm lifts up leaving the yellow block behind on the table, continuing to execute the stacking motion with an empty gripper.",
+    "vlm_temporal_justification": "At timestamp 02:58 (178 seconds into the magnified clip), the robot arm begins lifting upward after attempting to clamp down on the yellow block. However, the block stays flat on the table, making the grasp failure visibly obvious as the empty gripper rises.",
+    "vlm_video_fps": 50.7,
+    "vlm_video_num_frames": 426,
+    "vlm_video_duration_seconds": 8.382642998027613,
+    "vlm_usage": {
+      "input_tokens": 21583,
+      "output_tokens": 411,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 4.2,\n \"failure_window_start_seconds\": 3.5,\n \"failure_window_end_seconds\": 5.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the yellow block, but failed to secure its grip, raising its empty gripper into the air while leaving the block on the table.\",\n \"failure_reason_summary\": \"failed grasp on yellow block; lifting empty gripper\",\n \"recovery_action\": \"Move the arm back over the yellow block, lower the gripper around the block, close the fingers securely, and then pick it up.\",\n \"recovery_action_summary\": \"re-align gripper over yellow block; grasp and lift yellow block\",\n \"prevention_action\": \"Ensure the gripper is properly aligned and centered around the yellow block before closing the fingers to achieve a firm grasp before lifting.\",\n \"prevention_action_summary\": \"center gripper on yellow block; secure firm grip before lifting\",\n \"justification\": \"At 00:04 (4.0s), the robot closes its gripper on the yellow block, but fails to grip it securely. By 00:05 (5.0s), the arm lifts up leaving the yellow block behind on the table, continuing to execute the stacking motion with an empty gripper.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 239,
+    "failure_sim_step": 239,
+    "failure_policy_step": 39,
+    "failure_chunk_index": 5,
+    "first_post_failure_policy_step": 40,
+    "failure_type": "grasp_failure",
+    "correction_action": "Move the arm back over the yellow block, lower the gripper around the block, close the fingers securely, and then pick it up."
+  }
+}
\ No newline at end of file
diff --git a/blocks/ep040/labels.npz b/blocks/ep040/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..74fcca8fbcb3906aaa7b56bd4519b7d96cf7dc76
--- /dev/null
+++ b/blocks/ep040/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:36f18ec99156147501d0bf3a8323feaaf7dbda1db3485c3da680919f075bc7dc
+size 3826
diff --git a/blocks/ep041/example.md b/blocks/ep041/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..944a9ea2b788f28c3d67bb4ea8635d7b8829a2be
--- /dev/null
+++ b/blocks/ep041/example.md
@@ -0,0 +1,36 @@
+# t000__ep041__blocks
+
+- **suite / variant:** `` / ``
+- **instruction:** Stack the yellow block on the wooden block
+- **success:** False
+- **control steps:** 877  (17.5s at 50 Hz)
+- **policy inferences:** 151  (replan_steps=6)
+- **video:** `blocks/ep041/rollout.mp4`
+- **features:** `151` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | reaching for yellow block |
+| 3.0–6.0 | 151–301 | 25–50 | grasping and lifting yellow block |
+| 6.0–9.0 | 302–452 | 50–75 | carrying yellow block toward target |
+| 9.0–12.0 | 453–603 | 75–100 | aligning yellow block above wooden block |
+| 12.0–15.0 | 604–754 | 100–125 | releasing yellow block; block falling off |
+| 15.0–17.5 | 755–876 | 125–146 | retracting gripper away from table |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 14.9800796812749 s (00:15.0)
+- **frame / control_step:** 752 (policy_step=125, chunk_index=2)
+- **coarse onset:** 14.5 s (refined=True)
+- **mode:** placement_or_insertion_failure
+- **reason:** The robot positioned the yellow block over the wooden block and released its grip, but improper alignment caused the yellow block to slide off and fall onto the table.
+  Summary: improper placement alignment; block falling off wooden block
+- **recovery:** Re-grasp the yellow block from the table, move it directly over the center of the wooden block, and lower it carefully into place.
+  Summary: re-grasp yellow block; align and stack on wooden block
+- **prevention:** Ensure the yellow block is centered precisely over the wooden block and lowered close to its surface before opening the gripper.
+  Summary: center yellow block; release gently on wooden block
+- **justification:** Around 14.5 seconds into the video, the robot opens its gripper to release the yellow block onto the wooden block. Because the alignment was slightly off-center and unstable, the yellow block slips off the edge and lands next to the wooden block.
diff --git a/blocks/ep041/labels.json b/blocks/ep041/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..1de4fd92aac824606f5cea9856e79fe6ef710d9c
--- /dev/null
+++ b/blocks/ep041/labels.json
@@ -0,0 +1,131 @@
+{
+  "rollout_id": "t000__ep041__blocks",
+  "scene_variant": "",
+  "task_id": 0,
+  "task_file": "",
+  "instruction": "Stack the yellow block on the wooden block",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:15:04+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00796812749004,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for yellow block",
+      "description": "reaching for yellow block"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00796812749004,
+      "t_end_sec": 6.01593625498008,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "grasping and lifting yellow block",
+      "description": "grasping and lifting yellow block"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01593625498008,
+      "t_end_sec": 9.02390438247012,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "carrying yellow block toward target",
+      "description": "carrying yellow block toward target"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.02390438247012,
+      "t_end_sec": 12.03187250996016,
+      "control_step_start": 453,
+      "control_step_end": 603,
+      "policy_step_start": 75,
+      "policy_step_end": 100,
+      "phrase": "aligning yellow block above wooden block",
+      "description": "aligning yellow block above wooden block"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 12.03187250996016,
+      "t_end_sec": 15.0398406374502,
+      "control_step_start": 604,
+      "control_step_end": 754,
+      "policy_step_start": 100,
+      "policy_step_end": 125,
+      "phrase": "releasing yellow block; block falling off",
+      "description": "releasing yellow block; block falling off"
+    },
+    {
+      "segment_index": 5,
+      "t_start_sec": 15.0398406374502,
+      "t_end_sec": 17.47011952191235,
+      "control_step_start": 755,
+      "control_step_end": 876,
+      "policy_step_start": 125,
+      "policy_step_end": 146,
+      "phrase": "retracting gripper away from table",
+      "description": "retracting gripper away from table"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "placement_or_insertion_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 14.9800796812749,
+    "vlm_failure_onset_timestamp": "00:15.0",
+    "vlm_failure_onset_frame": 752,
+    "vlm_failure_onset_step": 752,
+    "vlm_coarse_failure_onset_seconds": 14.5,
+    "vlm_failure_window_start_seconds": 14.0,
+    "vlm_failure_window_end_seconds": 15.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot positioned the yellow block over the wooden block and released its grip, but improper alignment caused the yellow block to slide off and fall onto the table.",
+    "vlm_failure_reason_summary": "improper placement alignment; block falling off wooden block",
+    "vlm_recovery_action": "Re-grasp the yellow block from the table, move it directly over the center of the wooden block, and lower it carefully into place.",
+    "vlm_recovery_action_summary": "re-grasp yellow block; align and stack on wooden block",
+    "vlm_prevention_action": "Ensure the yellow block is centered precisely over the wooden block and lowered close to its surface before opening the gripper.",
+    "vlm_prevention_action_summary": "center yellow block; release gently on wooden block",
+    "vlm_justification": "Around 14.5 seconds into the video, the robot opens its gripper to release the yellow block onto the wooden block. Because the alignment was slightly off-center and unstable, the yellow block slips off the edge and lands next to the wooden block.",
+    "vlm_temporal_justification": "At 02:55 (second 175), the robot opens its gripper while misaligned with the wooden block, causing the yellow block to slide off and fall directly onto the table.",
+    "vlm_video_fps": 50.2,
+    "vlm_video_num_frames": 877,
+    "vlm_video_duration_seconds": 17.45019920318725,
+    "vlm_usage": {
+      "input_tokens": 21846,
+      "output_tokens": 385,
+      "cached_tokens": 1157
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"placement_or_insertion_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 14.5,\n \"failure_window_start_seconds\": 14.0,\n \"failure_window_end_seconds\": 15.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot positioned the yellow block over the wooden block and released its grip, but improper alignment caused the yellow block to slide off and fall onto the table.\",\n \"failure_reason_summary\": \"improper placement alignment; block falling off wooden block\",\n \"recovery_action\": \"Re-grasp the yellow block from the table, move it directly over the center of the wooden block, and lower it carefully into place.\",\n \"recovery_action_summary\": \"re-grasp yellow block; align and stack on wooden block\",\n \"prevention_action\": \"Ensure the yellow block is centered precisely over the wooden block and lowered close to its surface before opening the gripper.\",\n \"prevention_action_summary\": \"center yellow block; release gently on wooden block\",\n \"justification\": \"Around 14.5 seconds into the video, the robot opens its gripper to release the yellow block onto the wooden block. Because the alignment was slightly off-center and unstable, the yellow block slips off the edge and lands next to the wooden block.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 752,
+    "failure_sim_step": 752,
+    "failure_policy_step": 125,
+    "failure_chunk_index": 2,
+    "first_post_failure_policy_step": 126,
+    "failure_type": "placement_or_insertion_failure",
+    "correction_action": "Re-grasp the yellow block from the table, move it directly over the center of the wooden block, and lower it carefully into place."
+  }
+}
\ No newline at end of file
diff --git a/blocks/ep041/labels.npz b/blocks/ep041/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..e1c58ef61e3f13039eb533ca4fb9e582a32ef5f8
--- /dev/null
+++ b/blocks/ep041/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c178d43d77f49922641a5d7955f6666f7bf8c19d80e00d478c8154b37c635b82
+size 3991
diff --git a/blocks/ep042/example.md b/blocks/ep042/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..9588d0bfdf67980e2f0bc219b2dd7b0c51462514
--- /dev/null
+++ b/blocks/ep042/example.md
@@ -0,0 +1,39 @@
+# t000__ep042__blocks
+
+- **suite / variant:** `` / ``
+- **instruction:** Stack the yellow block on the wooden block
+- **success:** False
+- **control steps:** 1285  (25.5s at 50 Hz)
+- **policy inferences:** 220  (replan_steps=6)
+- **video:** `blocks/ep042/rollout.mp4`
+- **features:** `220` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | reaching for yellow block |
+| 3.0–6.0 | 151–301 | 25–50 | grasping yellow block |
+| 6.0–9.0 | 302–452 | 50–75 | lifting yellow block |
+| 9.0–12.0 | 453–603 | 75–100 | moving block over wooden block |
+| 12.0–15.0 | 604–754 | 100–125 | releasing block off target |
+| 15.0–18.0 | 755–905 | 125–150 | hovering above displaced block |
+| 18.0–21.0 | 906–1056 | 151–176 | hovering above table |
+| 21.0–24.0 | 1057–1207 | 176–201 | remaining stationary above block |
+| 24.0–25.5 | 1208–1284 | 201–214 | remaining stationary |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 13.499005964214712 s (00:13.5)
+- **frame / control_step:** 679 (policy_step=113, chunk_index=1)
+- **coarse onset:** 13.5 s (refined=False)
+- **mode:** placement_or_insertion_failure
+- **reason:** The robot released the yellow block too close to the front edge of the wooden block, causing the yellow block to slip off and tilt against the front side of the wooden block.
+  Summary: improper placement alignment; block slipping off target
+- **recovery:** Re-grasp the tilted yellow block, lift it up, align it centered over the wooden block's top face, and release it.
+  Summary: re-grasp yellow block; center over target; release carefully
+- **prevention:** Align the yellow block directly over the center of the wooden block before opening the gripper fingers to ensure a stable stack.
+  Summary: center block on target surface; release when stable
+- **justification:** At approximately 13.5 seconds, the robot opens its gripper while the yellow block is off-center near the edge of the wooden block. As a result, the block slides down and rests precariously tilted against the front of the wooden block, leaving the goal unfulfilled.
diff --git a/blocks/ep042/labels.json b/blocks/ep042/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..cc8af5d4cb13980b5b25e9784dce0fd3ac912138
--- /dev/null
+++ b/blocks/ep042/labels.json
@@ -0,0 +1,164 @@
+{
+  "rollout_id": "t000__ep042__blocks",
+  "scene_variant": "",
+  "task_id": 0,
+  "task_file": "",
+  "instruction": "Stack the yellow block on the wooden block",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:15:37+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.001988071570577,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for yellow block",
+      "description": "reaching for yellow block"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.001988071570577,
+      "t_end_sec": 6.003976143141154,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "grasping yellow block",
+      "description": "grasping yellow block"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.003976143141154,
+      "t_end_sec": 9.00596421471173,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "lifting yellow block",
+      "description": "lifting yellow block"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.00596421471173,
+      "t_end_sec": 12.007952286282308,
+      "control_step_start": 453,
+      "control_step_end": 603,
+      "policy_step_start": 75,
+      "policy_step_end": 100,
+      "phrase": "moving block over wooden block",
+      "description": "moving block over wooden block"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 12.007952286282308,
+      "t_end_sec": 15.009940357852884,
+      "control_step_start": 604,
+      "control_step_end": 754,
+      "policy_step_start": 100,
+      "policy_step_end": 125,
+      "phrase": "releasing block off target",
+      "description": "releasing block off target"
+    },
+    {
+      "segment_index": 5,
+      "t_start_sec": 15.009940357852884,
+      "t_end_sec": 18.01192842942346,
+      "control_step_start": 755,
+      "control_step_end": 905,
+      "policy_step_start": 125,
+      "policy_step_end": 150,
+      "phrase": "hovering above displaced block",
+      "description": "hovering above displaced block"
+    },
+    {
+      "segment_index": 6,
+      "t_start_sec": 18.01192842942346,
+      "t_end_sec": 21.013916500994036,
+      "control_step_start": 906,
+      "control_step_end": 1056,
+      "policy_step_start": 151,
+      "policy_step_end": 176,
+      "phrase": "hovering above table",
+      "description": "hovering above table"
+    },
+    {
+      "segment_index": 7,
+      "t_start_sec": 21.013916500994036,
+      "t_end_sec": 24.015904572564615,
+      "control_step_start": 1057,
+      "control_step_end": 1207,
+      "policy_step_start": 176,
+      "policy_step_end": 201,
+      "phrase": "remaining stationary above block",
+      "description": "remaining stationary above block"
+    },
+    {
+      "segment_index": 8,
+      "t_start_sec": 24.015904572564615,
+      "t_end_sec": 25.54671968190855,
+      "control_step_start": 1208,
+      "control_step_end": 1284,
+      "policy_step_start": 201,
+      "policy_step_end": 214,
+      "phrase": "remaining stationary",
+      "description": "remaining stationary"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "placement_or_insertion_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 13.499005964214712,
+    "vlm_failure_onset_timestamp": "00:13.5",
+    "vlm_failure_onset_frame": 679,
+    "vlm_failure_onset_step": 679,
+    "vlm_coarse_failure_onset_seconds": 13.5,
+    "vlm_failure_window_start_seconds": 12.5,
+    "vlm_failure_window_end_seconds": 14.5,
+    "vlm_temporal_refined": false,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot released the yellow block too close to the front edge of the wooden block, causing the yellow block to slip off and tilt against the front side of the wooden block.",
+    "vlm_failure_reason_summary": "improper placement alignment; block slipping off target",
+    "vlm_recovery_action": "Re-grasp the tilted yellow block, lift it up, align it centered over the wooden block's top face, and release it.",
+    "vlm_recovery_action_summary": "re-grasp yellow block; center over target; release carefully",
+    "vlm_prevention_action": "Align the yellow block directly over the center of the wooden block before opening the gripper fingers to ensure a stable stack.",
+    "vlm_prevention_action_summary": "center block on target surface; release when stable",
+    "vlm_justification": "At approximately 13.5 seconds, the robot opens its gripper while the yellow block is off-center near the edge of the wooden block. As a result, the block slides down and rests precariously tilted against the front of the wooden block, leaving the goal unfulfilled.",
+    "vlm_temporal_justification": "At 04:10, as the gripper releases the yellow block, the block visibly begins tilting and slipping off the front edge of the wooden block due to improper placement alignment.",
+    "vlm_video_fps": 50.3,
+    "vlm_video_num_frames": 1285,
+    "vlm_video_duration_seconds": 25.526838966202785,
+    "vlm_usage": {
+      "input_tokens": 22606,
+      "output_tokens": 392,
+      "cached_tokens": 1769
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"placement_or_insertion_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 13.5,\n \"failure_window_start_seconds\": 12.5,\n \"failure_window_end_seconds\": 14.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot released the yellow block too close to the front edge of the wooden block, causing the yellow block to slip off and tilt against the front side of the wooden block.\",\n \"failure_reason_summary\": \"improper placement alignment; block slipping off target\",\n \"recovery_action\": \"Re-grasp the tilted yellow block, lift it up, align it centered over the wooden block's top face, and release it.\",\n \"recovery_action_summary\": \"re-grasp yellow block; center over target; release carefully\",\n \"prevention_action\": \"Align the yellow block directly over the center of the wooden block before opening the gripper fingers to ensure a stable stack.\",\n \"prevention_action_summary\": \"center block on target surface; release when stable\",\n \"justification\": \"At approximately 13.5 seconds, the robot opens its gripper while the yellow block is off-center near the edge of the wooden block. As a result, the block slides down and rests precariously tilted against the front of the wooden block, leaving the goal unfulfilled.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 679,
+    "failure_sim_step": 679,
+    "failure_policy_step": 113,
+    "failure_chunk_index": 1,
+    "first_post_failure_policy_step": 114,
+    "failure_type": "placement_or_insertion_failure",
+    "correction_action": "Re-grasp the tilted yellow block, lift it up, align it centered over the wooden block's top face, and release it."
+  }
+}
\ No newline at end of file
diff --git a/blocks/ep042/labels.npz b/blocks/ep042/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..bc491e7fd78bf0db1c85be2ad4858c370820a06a
--- /dev/null
+++ b/blocks/ep042/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ac8a243e5dcc0166283cabebc3c20b85df125cba45ac63637e08dff9311002aa
+size 4024
diff --git a/blocks/ep043/example.md b/blocks/ep043/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..26538a305999c5c5e0b2e0928e432405e9c78975
--- /dev/null
+++ b/blocks/ep043/example.md
@@ -0,0 +1,35 @@
+# t000__ep043__blocks
+
+- **suite / variant:** `` / ``
+- **instruction:** Stack the yellow block on the wooden block
+- **success:** False
+- **control steps:** 679  (13.4s at 51 Hz)
+- **policy inferences:** 118  (replan_steps=6)
+- **video:** `blocks/ep043/rollout.mp4`
+- **features:** `118` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching toward yellow block |
+| 3.0–6.0 | 152–303 | 25–50 | closing gripper and lifting empty |
+| 6.0–9.0 | 304–455 | 50–75 | moving empty gripper toward target |
+| 9.0–12.0 | 456–607 | 76–101 | hovering over wooden block |
+| 12.0–13.4 | 608–678 | 101–113 | idle above wooden block |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 4.130434782608695 s (00:04.1)
+- **frame / control_step:** 209 (policy_step=34, chunk_index=5)
+- **coarse onset:** 3.5 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the yellow block, but the gripper closed without securing the object, leaving the yellow block on the table while the arm lifted.
+  Summary: failed grasp on yellow block; arm lifted empty
+- **recovery:** Descend back to the yellow block, realign the gripper, securely close on the block, and lift it to stack on the wooden block.
+  Summary: re-grasp yellow block; stack on wooden block
+- **prevention:** Ensure the end-effector is precisely aligned over the yellow block and maintain sufficient grasp contact before attempting to lift.
+  Summary: align gripper centered; secure grasp before lifting
+- **justification:** At around 3.5 seconds, the gripper attempts to close on the yellow block but fails to acquire it. The robot then lifts its empty gripper into the air and proceeds without the yellow block.
diff --git a/blocks/ep043/labels.json b/blocks/ep043/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..7b0ae80da6abb8c593c1b2989012588aa42176d4
--- /dev/null
+++ b/blocks/ep043/labels.json
@@ -0,0 +1,120 @@
+{
+  "rollout_id": "t000__ep043__blocks",
+  "scene_variant": "",
+  "task_id": 0,
+  "task_file": "",
+  "instruction": "Stack the yellow block on the wooden block",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:16:13+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0039525691699605,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching toward yellow block",
+      "description": "reaching toward yellow block"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0039525691699605,
+      "t_end_sec": 6.007905138339921,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "closing gripper and lifting empty",
+      "description": "closing gripper and lifting empty"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.007905138339921,
+      "t_end_sec": 9.011857707509881,
+      "control_step_start": 304,
+      "control_step_end": 455,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "moving empty gripper toward target",
+      "description": "moving empty gripper toward target"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.011857707509881,
+      "t_end_sec": 12.015810276679842,
+      "control_step_start": 456,
+      "control_step_end": 607,
+      "policy_step_start": 76,
+      "policy_step_end": 101,
+      "phrase": "hovering over wooden block",
+      "description": "hovering over wooden block"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 12.015810276679842,
+      "t_end_sec": 13.41897233201581,
+      "control_step_start": 608,
+      "control_step_end": 678,
+      "policy_step_start": 101,
+      "policy_step_end": 113,
+      "phrase": "idle above wooden block",
+      "description": "idle above wooden block"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 4.130434782608695,
+    "vlm_failure_onset_timestamp": "00:04.1",
+    "vlm_failure_onset_frame": 209,
+    "vlm_failure_onset_step": 209,
+    "vlm_coarse_failure_onset_seconds": 3.5,
+    "vlm_failure_window_start_seconds": 3.0,
+    "vlm_failure_window_end_seconds": 4.0,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the yellow block, but the gripper closed without securing the object, leaving the yellow block on the table while the arm lifted.",
+    "vlm_failure_reason_summary": "failed grasp on yellow block; arm lifted empty",
+    "vlm_recovery_action": "Descend back to the yellow block, realign the gripper, securely close on the block, and lift it to stack on the wooden block.",
+    "vlm_recovery_action_summary": "re-grasp yellow block; stack on wooden block",
+    "vlm_prevention_action": "Ensure the end-effector is precisely aligned over the yellow block and maintain sufficient grasp contact before attempting to lift.",
+    "vlm_prevention_action_summary": "align gripper centered; secure grasp before lifting",
+    "vlm_justification": "At around 3.5 seconds, the gripper attempts to close on the yellow block but fails to acquire it. The robot then lifts its empty gripper into the air and proceeds without the yellow block.",
+    "vlm_temporal_justification": "At clip second 184 (03:04), the gripper fingers finish closing completely against each other above the yellow block, making it visibly clear that the grasp failed before the arm lifts.",
+    "vlm_video_fps": 50.6,
+    "vlm_video_num_frames": 679,
+    "vlm_video_duration_seconds": 13.399209486166008,
+    "vlm_usage": {
+      "input_tokens": 21899,
+      "output_tokens": 362,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 3.5,\n \"failure_window_start_seconds\": 3.0,\n \"failure_window_end_seconds\": 4.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the yellow block, but the gripper closed without securing the object, leaving the yellow block on the table while the arm lifted.\",\n \"failure_reason_summary\": \"failed grasp on yellow block; arm lifted empty\",\n \"recovery_action\": \"Descend back to the yellow block, realign the gripper, securely close on the block, and lift it to stack on the wooden block.\",\n \"recovery_action_summary\": \"re-grasp yellow block; stack on wooden block\",\n \"prevention_action\": \"Ensure the end-effector is precisely aligned over the yellow block and maintain sufficient grasp contact before attempting to lift.\",\n \"prevention_action_summary\": \"align gripper centered; secure grasp before lifting\",\n \"justification\": \"At around 3.5 seconds, the gripper attempts to close on the yellow block but fails to acquire it. The robot then lifts its empty gripper into the air and proceeds without the yellow block.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 209,
+    "failure_sim_step": 209,
+    "failure_policy_step": 34,
+    "failure_chunk_index": 5,
+    "first_post_failure_policy_step": 35,
+    "failure_type": "grasp_failure",
+    "correction_action": "Descend back to the yellow block, realign the gripper, securely close on the block, and lift it to stack on the wooden block."
+  }
+}
\ No newline at end of file
diff --git a/blocks/ep043/labels.npz b/blocks/ep043/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..2dabe1234b3a1dec8fc47e673ed06c4fa21999df
--- /dev/null
+++ b/blocks/ep043/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b066712ce4a3203f4c56342aaed43bfc38f95234dee1c7e7911c8111a12bbfb0
+size 3884
diff --git a/blocks/ep044/example.md b/blocks/ep044/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..dbabf483596f7385827a391270467e91d193fdfa
--- /dev/null
+++ b/blocks/ep044/example.md
@@ -0,0 +1,33 @@
+# t000__ep044__blocks
+
+- **suite / variant:** `` / ``
+- **instruction:** Stack the yellow block on the wooden block
+- **success:** False
+- **control steps:** 439  (8.6s at 51 Hz)
+- **policy inferences:** 76  (replan_steps=6)
+- **video:** `blocks/ep044/rollout.mp4`
+- **features:** `76` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | approaching and aligning with yellow block |
+| 3.0–6.0 | 152–303 | 25–50 | grasping yellow block; block drops |
+| 6.0–8.6 | 304–438 | 50–73 | moving empty gripper over wooden block |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 4.921259842519685 s (00:04.9)
+- **frame / control_step:** 250 (policy_step=41, chunk_index=4)
+- **coarse onset:** 4.8 s (refined=True)
+- **mode:** object_displacement
+- **reason:** The robot successfully grasps the yellow block but loses its grip and drops the block while moving toward the target location.
+  Summary: dropping grasped object; losing grip during transfer
+- **recovery:** Lower the gripper back over the dropped yellow block, establish a secure grasp, lift it, and place it on top of the wooden block.
+  Summary: re-grasp yellow block; stack on wooden block
+- **prevention:** Maintain a firm and centered grip on the yellow block while lifting and transporting it to prevent it from slipping out.
+  Summary: maintain secure grip; transfer block smoothly
+- **justification:** At around 4.8 seconds, the yellow block slips out of the robot's gripper and drops back onto the table. The robot then continues moving its empty gripper toward the wooden block, failing the stacking task.
diff --git a/blocks/ep044/labels.json b/blocks/ep044/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..92e8e2ccff1e8f9aaa436045529588b4e8583766
--- /dev/null
+++ b/blocks/ep044/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t000__ep044__blocks",
+  "scene_variant": "",
+  "task_id": 0,
+  "task_file": "",
+  "instruction": "Stack the yellow block on the wooden block",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:16:46+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9921259842519685,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "approaching and aligning with yellow block",
+      "description": "approaching and aligning with yellow block"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9921259842519685,
+      "t_end_sec": 5.984251968503937,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "grasping yellow block; block drops",
+      "description": "grasping yellow block; block drops"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.984251968503937,
+      "t_end_sec": 8.641732283464567,
+      "control_step_start": 304,
+      "control_step_end": 438,
+      "policy_step_start": 50,
+      "policy_step_end": 73,
+      "phrase": "moving empty gripper over wooden block",
+      "description": "moving empty gripper over wooden block"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "object_displacement",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 4.921259842519685,
+    "vlm_failure_onset_timestamp": "00:04.9",
+    "vlm_failure_onset_frame": 250,
+    "vlm_failure_onset_step": 250,
+    "vlm_coarse_failure_onset_seconds": 4.8,
+    "vlm_failure_window_start_seconds": 4.5,
+    "vlm_failure_window_end_seconds": 5.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot successfully grasps the yellow block but loses its grip and drops the block while moving toward the target location.",
+    "vlm_failure_reason_summary": "dropping grasped object; losing grip during transfer",
+    "vlm_recovery_action": "Lower the gripper back over the dropped yellow block, establish a secure grasp, lift it, and place it on top of the wooden block.",
+    "vlm_recovery_action_summary": "re-grasp yellow block; stack on wooden block",
+    "vlm_prevention_action": "Maintain a firm and centered grip on the yellow block while lifting and transporting it to prevent it from slipping out.",
+    "vlm_prevention_action_summary": "maintain secure grip; transfer block smoothly",
+    "vlm_justification": "At around 4.8 seconds, the yellow block slips out of the robot's gripper and drops back onto the table. The robot then continues moving its empty gripper toward the wooden block, failing the stacking task.",
+    "vlm_temporal_justification": "At clip second 158 (02:38), as the robot arm begins moving laterally towards the wooden block, the yellow block loses stability in the gripper and starts slipping downward, dropping onto the table surface by second 160.",
+    "vlm_video_fps": 50.8,
+    "vlm_video_num_frames": 439,
+    "vlm_video_duration_seconds": 8.622047244094489,
+    "vlm_usage": {
+      "input_tokens": 21638,
+      "output_tokens": 366,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"object_displacement\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 4.8,\n \"failure_window_start_seconds\": 4.5,\n \"failure_window_end_seconds\": 5.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot successfully grasps the yellow block but loses its grip and drops the block while moving toward the target location.\",\n \"failure_reason_summary\": \"dropping grasped object; losing grip during transfer\",\n \"recovery_action\": \"Lower the gripper back over the dropped yellow block, establish a secure grasp, lift it, and place it on top of the wooden block.\",\n \"recovery_action_summary\": \"re-grasp yellow block; stack on wooden block\",\n \"prevention_action\": \"Maintain a firm and centered grip on the yellow block while lifting and transporting it to prevent it from slipping out.\",\n \"prevention_action_summary\": \"maintain secure grip; transfer block smoothly\",\n \"justification\": \"At around 4.8 seconds, the yellow block slips out of the robot's gripper and drops back onto the table. The robot then continues moving its empty gripper toward the wooden block, failing the stacking task.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 250,
+    "failure_sim_step": 250,
+    "failure_policy_step": 41,
+    "failure_chunk_index": 4,
+    "first_post_failure_policy_step": 42,
+    "failure_type": "object_displacement",
+    "correction_action": "Lower the gripper back over the dropped yellow block, establish a secure grasp, lift it, and place it on top of the wooden block."
+  }
+}
\ No newline at end of file
diff --git a/blocks/ep044/labels.npz b/blocks/ep044/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..f9813d0935bc47c6df9b8512f9bc6a5e092b2f0e
--- /dev/null
+++ b/blocks/ep044/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0e9796501f2489b531952b0aa448e2bdcd0cebc87fe6e78953e3841260ac9602
+size 3806
diff --git a/blocks/ep045/example.md b/blocks/ep045/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..aae643b37f57aa5cc49e004aed6277b700240b9d
--- /dev/null
+++ b/blocks/ep045/example.md
@@ -0,0 +1,35 @@
+# t000__ep045__blocks
+
+- **suite / variant:** `` / ``
+- **instruction:** Stack the yellow block on the wooden block
+- **success:** False
+- **control steps:** 695  (13.8s at 50 Hz)
+- **policy inferences:** 119  (replan_steps=6)
+- **video:** `blocks/ep045/rollout.mp4`
+- **features:** `119` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | approaching and grasping yellow block |
+| 3.0–6.0 | 151–301 | 25–50 | lifting and transporting yellow block |
+| 6.0–9.0 | 302–452 | 50–75 | placing block and releasing gripper |
+| 9.0–12.0 | 453–603 | 75–100 | idle over stacked blocks |
+| 12.0–13.8 | 604–694 | 100–115 | remaining stationary over blocks |
+
+## Failure
+
+- **yes**
+- **onset type:** operator_intervention
+- **onset:** 5.0396825396825395 s (00:05.0)
+- **frame / control_step:** 254 (policy_step=42, chunk_index=2)
+- **coarse onset:** 8.0 s (refined=True)
+- **mode:** stuck_or_no_progress
+- **reason:** After releasing the yellow block onto the wooden block, the robot arm stops moving and remains hovering directly over the stacked blocks without retracting upwards.
+  Summary: robot stuck over stacked blocks; failure to retract gripper
+- **recovery:** Lift the end-effector vertically away from the stacked blocks to clear contact and allow the task goal to register.
+  Summary: raise arm vertically; clear workspace
+- **prevention:** Ensure the controller commands an upward retreat trajectory immediately after opening the gripper to complete the stacking routine.
+  Summary: lift end-effector after release; clear stacked blocks
+- **justification:** The robot successfully grasps the yellow block and places it onto the wooden block around 7.5 to 8.0 seconds into the rollout. However, after opening the gripper at 8.0 seconds, the arm freezes and remains stationary on top of the blocks for the remainder of the episode.
diff --git a/blocks/ep045/labels.json b/blocks/ep045/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..6f06df89562a1dc4a7fa264e2934ea74c75124ab
--- /dev/null
+++ b/blocks/ep045/labels.json
@@ -0,0 +1,120 @@
+{
+  "rollout_id": "t000__ep045__blocks",
+  "scene_variant": "",
+  "task_id": 0,
+  "task_file": "",
+  "instruction": "Stack the yellow block on the wooden block",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:17:19+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.996031746031746,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "approaching and grasping yellow block",
+      "description": "approaching and grasping yellow block"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.996031746031746,
+      "t_end_sec": 5.992063492063492,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "lifting and transporting yellow block",
+      "description": "lifting and transporting yellow block"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.992063492063492,
+      "t_end_sec": 8.988095238095239,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "placing block and releasing gripper",
+      "description": "placing block and releasing gripper"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.988095238095239,
+      "t_end_sec": 11.984126984126984,
+      "control_step_start": 453,
+      "control_step_end": 603,
+      "policy_step_start": 75,
+      "policy_step_end": 100,
+      "phrase": "idle over stacked blocks",
+      "description": "idle over stacked blocks"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 11.984126984126984,
+      "t_end_sec": 13.78968253968254,
+      "control_step_start": 604,
+      "control_step_end": 694,
+      "policy_step_start": 100,
+      "policy_step_end": 115,
+      "phrase": "remaining stationary over blocks",
+      "description": "remaining stationary over blocks"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "stuck_or_no_progress",
+    "vlm_failure_onset_type": "operator_intervention",
+    "vlm_failure_onset_seconds": 5.0396825396825395,
+    "vlm_failure_onset_timestamp": "00:05.0",
+    "vlm_failure_onset_frame": 254,
+    "vlm_failure_onset_step": 254,
+    "vlm_coarse_failure_onset_seconds": 8.0,
+    "vlm_failure_window_start_seconds": 8.0,
+    "vlm_failure_window_end_seconds": 13.77,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "After releasing the yellow block onto the wooden block, the robot arm stops moving and remains hovering directly over the stacked blocks without retracting upwards.",
+    "vlm_failure_reason_summary": "robot stuck over stacked blocks; failure to retract gripper",
+    "vlm_recovery_action": "Lift the end-effector vertically away from the stacked blocks to clear contact and allow the task goal to register.",
+    "vlm_recovery_action_summary": "raise arm vertically; clear workspace",
+    "vlm_prevention_action": "Ensure the controller commands an upward retreat trajectory immediately after opening the gripper to complete the stacking routine.",
+    "vlm_prevention_action_summary": "lift end-effector after release; clear stacked blocks",
+    "vlm_justification": "The robot successfully grasps the yellow block and places it onto the wooden block around 7.5 to 8.0 seconds into the rollout. However, after opening the gripper at 8.0 seconds, the arm freezes and remains stationary on top of the blocks for the remainder of the episode.",
+    "vlm_temporal_justification": "By second 2 (corresponding to frame 254 of the original video), the robot has placed the yellow block on the wooden block and completely stops moving, remaining frozen over the blocks without retracting.",
+    "vlm_video_fps": 50.4,
+    "vlm_video_num_frames": 695,
+    "vlm_video_duration_seconds": 13.76984126984127,
+    "vlm_usage": {
+      "input_tokens": 21840,
+      "output_tokens": 374,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"stuck_or_no_progress\",\n \"onset_type\": \"operator_intervention\",\n \"failure_onset_seconds\": 8.0,\n \"failure_window_start_seconds\": 8.0,\n \"failure_window_end_seconds\": 13.77,\n \"confidence\": \"high\",\n \"failure_reason\": \"After releasing the yellow block onto the wooden block, the robot arm stops moving and remains hovering directly over the stacked blocks without retracting upwards.\",\n \"failure_reason_summary\": \"robot stuck over stacked blocks; failure to retract gripper\",\n \"recovery_action\": \"Lift the end-effector vertically away from the stacked blocks to clear contact and allow the task goal to register.\",\n \"recovery_action_summary\": \"raise arm vertically; clear workspace\",\n \"prevention_action\": \"Ensure the controller commands an upward retreat trajectory immediately after opening the gripper to complete the stacking routine.\",\n \"prevention_action_summary\": \"lift end-effector after release; clear stacked blocks\",\n \"justification\": \"The robot successfully grasps the yellow block and places it onto the wooden block around 7.5 to 8.0 seconds into the rollout. However, after opening the gripper at 8.0 seconds, the arm freezes and remains stationary on top of the blocks for the remainder of the episode.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 254,
+    "failure_sim_step": 254,
+    "failure_policy_step": 42,
+    "failure_chunk_index": 2,
+    "first_post_failure_policy_step": 43,
+    "failure_type": "stuck_or_no_progress",
+    "correction_action": "Lift the end-effector vertically away from the stacked blocks to clear contact and allow the task goal to register."
+  }
+}
\ No newline at end of file
diff --git a/blocks/ep045/labels.npz b/blocks/ep045/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..55378b26cc352099e772485a711880d7bd3eb024
--- /dev/null
+++ b/blocks/ep045/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cc44dde79b45dc2d0d0cf1852768ef788514848b805a40da67e3df2989d4d1bd
+size 3934
diff --git a/blocks/ep046/example.md b/blocks/ep046/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..94e55ec738b63f1bdba8e16e7a9a4c1777fdef8a
--- /dev/null
+++ b/blocks/ep046/example.md
@@ -0,0 +1,36 @@
+# t000__ep046__blocks
+
+- **suite / variant:** `` / ``
+- **instruction:** Stack the yellow block on the wooden block
+- **success:** False
+- **control steps:** 836  (16.6s at 50 Hz)
+- **policy inferences:** 144  (replan_steps=6)
+- **video:** `blocks/ep046/rollout.mp4`
+- **features:** `144` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | lowering gripper toward yellow block |
+| 3.0–6.0 | 151–301 | 25–50 | attempting grasp and lifting empty |
+| 6.0–9.0 | 302–452 | 50–75 | moving empty gripper over wooden block |
+| 9.0–12.0 | 453–603 | 75–100 | holding position above wooden block |
+| 12.0–15.0 | 604–754 | 100–125 | remaining idle above wooden block |
+| 15.0–16.6 | 755–835 | 125–139 | remaining idle at final position |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 4.007936507936508 s (00:04.0)
+- **frame / control_step:** 202 (policy_step=33, chunk_index=4)
+- **coarse onset:** 4.2 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the yellow block but failed to close securely around it, lifting an empty gripper and moving to the placement target without the block.
+  Summary: failed grasp on yellow block; moving empty gripper to target
+- **recovery:** Move the gripper back to the yellow block on the table, execute a secure grasp, and place it on top of the wooden block.
+  Summary: re-align gripper; grasp yellow block; place on wooden block
+- **prevention:** Center the gripper accurately around the yellow block and verify a firm grasp before lifting and moving toward the wooden block.
+  Summary: center gripper on block; secure hold before transport
+- **justification:** At around 4.2 seconds, the gripper attempts to close on the yellow block but fails to secure it. By 5.0 seconds, the robot lifts its gripper empty and proceeds to execute the placement motion over the wooden block without carrying the block.
diff --git a/blocks/ep046/labels.json b/blocks/ep046/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..93a6f8d0366cb6d8b18fc456f500b204670afea8
--- /dev/null
+++ b/blocks/ep046/labels.json
@@ -0,0 +1,131 @@
+{
+  "rollout_id": "t000__ep046__blocks",
+  "scene_variant": "",
+  "task_id": 0,
+  "task_file": "",
+  "instruction": "Stack the yellow block on the wooden block",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:17:57+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.996031746031746,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "lowering gripper toward yellow block",
+      "description": "lowering gripper toward yellow block"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.996031746031746,
+      "t_end_sec": 5.992063492063492,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "attempting grasp and lifting empty",
+      "description": "attempting grasp and lifting empty"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.992063492063492,
+      "t_end_sec": 8.988095238095239,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "moving empty gripper over wooden block",
+      "description": "moving empty gripper over wooden block"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.988095238095239,
+      "t_end_sec": 11.984126984126984,
+      "control_step_start": 453,
+      "control_step_end": 603,
+      "policy_step_start": 75,
+      "policy_step_end": 100,
+      "phrase": "holding position above wooden block",
+      "description": "holding position above wooden block"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 11.984126984126984,
+      "t_end_sec": 14.980158730158731,
+      "control_step_start": 604,
+      "control_step_end": 754,
+      "policy_step_start": 100,
+      "policy_step_end": 125,
+      "phrase": "remaining idle above wooden block",
+      "description": "remaining idle above wooden block"
+    },
+    {
+      "segment_index": 5,
+      "t_start_sec": 14.980158730158731,
+      "t_end_sec": 16.58730158730159,
+      "control_step_start": 755,
+      "control_step_end": 835,
+      "policy_step_start": 125,
+      "policy_step_end": 139,
+      "phrase": "remaining idle at final position",
+      "description": "remaining idle at final position"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 4.007936507936508,
+    "vlm_failure_onset_timestamp": "00:04.0",
+    "vlm_failure_onset_frame": 202,
+    "vlm_failure_onset_step": 202,
+    "vlm_coarse_failure_onset_seconds": 4.2,
+    "vlm_failure_window_start_seconds": 4.0,
+    "vlm_failure_window_end_seconds": 5.0,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the yellow block but failed to close securely around it, lifting an empty gripper and moving to the placement target without the block.",
+    "vlm_failure_reason_summary": "failed grasp on yellow block; moving empty gripper to target",
+    "vlm_recovery_action": "Move the gripper back to the yellow block on the table, execute a secure grasp, and place it on top of the wooden block.",
+    "vlm_recovery_action_summary": "re-align gripper; grasp yellow block; place on wooden block",
+    "vlm_prevention_action": "Center the gripper accurately around the yellow block and verify a firm grasp before lifting and moving toward the wooden block.",
+    "vlm_prevention_action_summary": "center gripper on block; secure hold before transport",
+    "vlm_justification": "At around 4.2 seconds, the gripper attempts to close on the yellow block but fails to secure it. By 5.0 seconds, the robot lifts its gripper empty and proceeds to execute the placement motion over the wooden block without carrying the block.",
+    "vlm_temporal_justification": "The robot closes its gripper around the yellow block and begins to lift at 02:20. At 02:21, as the gripper moves upward, it is clearly visible that the yellow block remains stationary on the table, leaving the gripper empty.",
+    "vlm_video_fps": 50.4,
+    "vlm_video_num_frames": 836,
+    "vlm_video_duration_seconds": 16.567460317460316,
+    "vlm_usage": {
+      "input_tokens": 22026,
+      "output_tokens": 391,
+      "cached_tokens": 1157
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 4.2,\n \"failure_window_start_seconds\": 4.0,\n \"failure_window_end_seconds\": 5.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the yellow block but failed to close securely around it, lifting an empty gripper and moving to the placement target without the block.\",\n \"failure_reason_summary\": \"failed grasp on yellow block; moving empty gripper to target\",\n \"recovery_action\": \"Move the gripper back to the yellow block on the table, execute a secure grasp, and place it on top of the wooden block.\",\n \"recovery_action_summary\": \"re-align gripper; grasp yellow block; place on wooden block\",\n \"prevention_action\": \"Center the gripper accurately around the yellow block and verify a firm grasp before lifting and moving toward the wooden block.\",\n \"prevention_action_summary\": \"center gripper on block; secure hold before transport\",\n \"justification\": \"At around 4.2 seconds, the gripper attempts to close on the yellow block but fails to secure it. By 5.0 seconds, the robot lifts its gripper empty and proceeds to execute the placement motion over the wooden block without carrying the block.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 202,
+    "failure_sim_step": 202,
+    "failure_policy_step": 33,
+    "failure_chunk_index": 4,
+    "first_post_failure_policy_step": 34,
+    "failure_type": "grasp_failure",
+    "correction_action": "Move the gripper back to the yellow block on the table, execute a secure grasp, and place it on top of the wooden block."
+  }
+}
\ No newline at end of file
diff --git a/blocks/ep046/labels.npz b/blocks/ep046/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..4d4cb7d483c612984188b186ec76ddddd873a9d8
--- /dev/null
+++ b/blocks/ep046/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6833ed31c6e1f8c6cc34eda6d3755048ba6998f5bf1172458167483824399787
+size 3961
diff --git a/blocks/ep047/example.md b/blocks/ep047/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..e86d411073e41a3115726495c3461e5ff9659a6c
--- /dev/null
+++ b/blocks/ep047/example.md
@@ -0,0 +1,36 @@
+# t000__ep047__blocks
+
+- **suite / variant:** `` / ``
+- **instruction:** Stack the yellow block on the wooden block
+- **success:** False
+- **control steps:** 833  (16.3s at 51 Hz)
+- **policy inferences:** 142  (replan_steps=6)
+- **video:** `blocks/ep047/rollout.mp4`
+- **features:** `142` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–152 | 0–25 | reaching toward yellow block |
+| 3.0–6.0 | 153–305 | 25–50 | grasping the yellow block |
+| 6.0–9.0 | 306–458 | 51–76 | lifting yellow block toward target |
+| 9.0–12.0 | 459–611 | 76–101 | lowering block and slipping off |
+| 12.0–15.0 | 612–764 | 102–127 | hovering gripper near tabletop |
+| 15.0–16.3 | 765–832 | 127–138 | idle above table surface |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 8.784313725490197 s (00:08.8)
+- **frame / control_step:** 448 (policy_step=74, chunk_index=4)
+- **coarse onset:** 10.5 s (refined=True)
+- **mode:** placement_or_insertion_failure
+- **reason:** The robot successfully grasped the yellow block and carried it to the wooden block, but misaligned the final placement, causing the yellow block to slip off the wooden block and fall onto the table.
+  Summary: misaligned placement; block slipped off target
+- **recovery:** Re-grasp the fallen yellow block from the tabletop, align it carefully above the wooden block, and gently stack it on top.
+  Summary: re-grasp yellow block; align and stack on wooden block
+- **prevention:** Align the yellow block directly over the center of the wooden block and lower it smoothly before releasing the grip.
+  Summary: center yellow block over target; release carefully
+- **justification:** At around 10.5 seconds, the robot lowers the yellow block onto the edge of the wooden block rather than centering it. This poor placement causes the yellow block to tilt and slide off onto the table, leaving the task unfulfilled.
diff --git a/blocks/ep047/labels.json b/blocks/ep047/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..367d4e475483fb7f5412680b4c05065298523e7d
--- /dev/null
+++ b/blocks/ep047/labels.json
@@ -0,0 +1,131 @@
+{
+  "rollout_id": "t000__ep047__blocks",
+  "scene_variant": "",
+  "task_id": 0,
+  "task_file": "",
+  "instruction": "Stack the yellow block on the wooden block",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:18:40+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0,
+      "control_step_start": 0,
+      "control_step_end": 152,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching toward yellow block",
+      "description": "reaching toward yellow block"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0,
+      "t_end_sec": 6.0,
+      "control_step_start": 153,
+      "control_step_end": 305,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "grasping the yellow block",
+      "description": "grasping the yellow block"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.0,
+      "t_end_sec": 9.0,
+      "control_step_start": 306,
+      "control_step_end": 458,
+      "policy_step_start": 51,
+      "policy_step_end": 76,
+      "phrase": "lifting yellow block toward target",
+      "description": "lifting yellow block toward target"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.0,
+      "t_end_sec": 12.0,
+      "control_step_start": 459,
+      "control_step_end": 611,
+      "policy_step_start": 76,
+      "policy_step_end": 101,
+      "phrase": "lowering block and slipping off",
+      "description": "lowering block and slipping off"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 12.0,
+      "t_end_sec": 15.0,
+      "control_step_start": 612,
+      "control_step_end": 764,
+      "policy_step_start": 102,
+      "policy_step_end": 127,
+      "phrase": "hovering gripper near tabletop",
+      "description": "hovering gripper near tabletop"
+    },
+    {
+      "segment_index": 5,
+      "t_start_sec": 15.0,
+      "t_end_sec": 16.333333333333332,
+      "control_step_start": 765,
+      "control_step_end": 832,
+      "policy_step_start": 127,
+      "policy_step_end": 138,
+      "phrase": "idle above table surface",
+      "description": "idle above table surface"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "placement_or_insertion_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 8.784313725490197,
+    "vlm_failure_onset_timestamp": "00:08.8",
+    "vlm_failure_onset_frame": 448,
+    "vlm_failure_onset_step": 448,
+    "vlm_coarse_failure_onset_seconds": 10.5,
+    "vlm_failure_window_start_seconds": 9.5,
+    "vlm_failure_window_end_seconds": 11.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot successfully grasped the yellow block and carried it to the wooden block, but misaligned the final placement, causing the yellow block to slip off the wooden block and fall onto the table.",
+    "vlm_failure_reason_summary": "misaligned placement; block slipped off target",
+    "vlm_recovery_action": "Re-grasp the fallen yellow block from the tabletop, align it carefully above the wooden block, and gently stack it on top.",
+    "vlm_recovery_action_summary": "re-grasp yellow block; align and stack on wooden block",
+    "vlm_prevention_action": "Align the yellow block directly over the center of the wooden block and lower it smoothly before releasing the grip.",
+    "vlm_prevention_action_summary": "center yellow block over target; release carefully",
+    "vlm_justification": "At around 10.5 seconds, the robot lowers the yellow block onto the edge of the wooden block rather than centering it. This poor placement causes the yellow block to tilt and slide off onto the table, leaving the task unfulfilled.",
+    "vlm_temporal_justification": "At 01:05 (second 65 of the clip), the robot gripper opens to release the yellow block, but places it off-center and tilted on the edge of the wooden block rather than flush on top, causing it to eventually slip off onto the table.",
+    "vlm_video_fps": 51.0,
+    "vlm_video_num_frames": 833,
+    "vlm_video_duration_seconds": 16.313725490196077,
+    "vlm_usage": {
+      "input_tokens": 22231,
+      "output_tokens": 397,
+      "cached_tokens": 1089
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"placement_or_insertion_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 10.5,\n \"failure_window_start_seconds\": 9.5,\n \"failure_window_end_seconds\": 11.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot successfully grasped the yellow block and carried it to the wooden block, but misaligned the final placement, causing the yellow block to slip off the wooden block and fall onto the table.\",\n \"failure_reason_summary\": \"misaligned placement; block slipped off target\",\n \"recovery_action\": \"Re-grasp the fallen yellow block from the tabletop, align it carefully above the wooden block, and gently stack it on top.\",\n \"recovery_action_summary\": \"re-grasp yellow block; align and stack on wooden block\",\n \"prevention_action\": \"Align the yellow block directly over the center of the wooden block and lower it smoothly before releasing the grip.\",\n \"prevention_action_summary\": \"center yellow block over target; release carefully\",\n \"justification\": \"At around 10.5 seconds, the robot lowers the yellow block onto the edge of the wooden block rather than centering it. This poor placement causes the yellow block to tilt and slide off onto the table, leaving the task unfulfilled.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 448,
+    "failure_sim_step": 448,
+    "failure_policy_step": 74,
+    "failure_chunk_index": 4,
+    "first_post_failure_policy_step": 75,
+    "failure_type": "placement_or_insertion_failure",
+    "correction_action": "Re-grasp the fallen yellow block from the tabletop, align it carefully above the wooden block, and gently stack it on top."
+  }
+}
\ No newline at end of file
diff --git a/blocks/ep047/labels.npz b/blocks/ep047/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..b2a470695fe8d2f9fa641632c5700f51eea91f3d
--- /dev/null
+++ b/blocks/ep047/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2ff2fed78c58fa450cbd9bff6e1afa2aca8e4ee510d3d66ff4be2cdcfbc12ca2
+size 3964
diff --git a/blocks/ep048/example.md b/blocks/ep048/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..ddd3f674143336dc527f0a91e081a3b496049e29
--- /dev/null
+++ b/blocks/ep048/example.md
@@ -0,0 +1,39 @@
+# t000__ep048__blocks
+
+- **suite / variant:** `` / ``
+- **instruction:** Stack the yellow block on the wooden block
+- **success:** False
+- **control steps:** 1281  (25.4s at 50 Hz)
+- **policy inferences:** 221  (replan_steps=6)
+- **video:** `blocks/ep048/rollout.mp4`
+- **features:** `221` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching and grasping yellow block |
+| 3.0–6.0 | 152–303 | 25–50 | lifting and moving yellow block |
+| 6.0–9.0 | 304–455 | 50–75 | lowering and releasing yellow block off-target |
+| 9.0–12.0 | 456–607 | 76–101 | pressing gripper near blocks |
+| 12.0–15.0 | 608–759 | 101–126 | hovering near blocks |
+| 15.0–18.1 | 760–911 | 126–151 | lingering over blocks |
+| 18.1–21.1 | 912–1063 | 152–177 | remaining stationary over blocks |
+| 21.1–24.1 | 1064–1215 | 177–202 | hovering idle near blocks |
+| 24.1–25.4 | 1216–1280 | 202–213 | idle / no progress |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 7.00990099009901 s (00:07.0)
+- **frame / control_step:** 354 (policy_step=59, chunk_index=0)
+- **coarse onset:** 8.0 s (refined=True)
+- **mode:** placement_or_insertion_failure
+- **reason:** The robot carries the yellow block to the wooden block but misaligns its position during release, placing the yellow block beside the wooden block instead of stacking it on top.
+  Summary: misaligned placement; yellow block placed beside wooden block
+- **recovery:** Re-grasp the yellow block, raise it above the wooden block, align it directly over the top surface, and place it down cleanly.
+  Summary: re-grasp yellow block; align and stack on wooden block
+- **prevention:** Position the yellow block vertically directly above the center of the wooden block before releasing the gripper to ensure a stable stacked placement.
+  Summary: center yellow block above wooden block; release directly on top
+- **justification:** The robot successfully grasps the yellow block and brings it to the target area, but at around 8.0 seconds, it misaligns the placement attempt, causing the yellow block to slip down beside the wooden block rather than resting on top.
diff --git a/blocks/ep048/labels.json b/blocks/ep048/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..9f03e9c104273f8697953539a8934203ff8a8378
--- /dev/null
+++ b/blocks/ep048/labels.json
@@ -0,0 +1,164 @@
+{
+  "rollout_id": "t000__ep048__blocks",
+  "scene_variant": "",
+  "task_id": 0,
+  "task_file": "",
+  "instruction": "Stack the yellow block on the wooden block",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:19:17+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00990099009901,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching and grasping yellow block",
+      "description": "reaching and grasping yellow block"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00990099009901,
+      "t_end_sec": 6.01980198019802,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "lifting and moving yellow block",
+      "description": "lifting and moving yellow block"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01980198019802,
+      "t_end_sec": 9.029702970297029,
+      "control_step_start": 304,
+      "control_step_end": 455,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "lowering and releasing yellow block off-target",
+      "description": "lowering and releasing yellow block off-target"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.029702970297029,
+      "t_end_sec": 12.03960396039604,
+      "control_step_start": 456,
+      "control_step_end": 607,
+      "policy_step_start": 76,
+      "policy_step_end": 101,
+      "phrase": "pressing gripper near blocks",
+      "description": "pressing gripper near blocks"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 12.03960396039604,
+      "t_end_sec": 15.049504950495049,
+      "control_step_start": 608,
+      "control_step_end": 759,
+      "policy_step_start": 101,
+      "policy_step_end": 126,
+      "phrase": "hovering near blocks",
+      "description": "hovering near blocks"
+    },
+    {
+      "segment_index": 5,
+      "t_start_sec": 15.049504950495049,
+      "t_end_sec": 18.059405940594058,
+      "control_step_start": 760,
+      "control_step_end": 911,
+      "policy_step_start": 126,
+      "policy_step_end": 151,
+      "phrase": "lingering over blocks",
+      "description": "lingering over blocks"
+    },
+    {
+      "segment_index": 6,
+      "t_start_sec": 18.059405940594058,
+      "t_end_sec": 21.06930693069307,
+      "control_step_start": 912,
+      "control_step_end": 1063,
+      "policy_step_start": 152,
+      "policy_step_end": 177,
+      "phrase": "remaining stationary over blocks",
+      "description": "remaining stationary over blocks"
+    },
+    {
+      "segment_index": 7,
+      "t_start_sec": 21.06930693069307,
+      "t_end_sec": 24.07920792079208,
+      "control_step_start": 1064,
+      "control_step_end": 1215,
+      "policy_step_start": 177,
+      "policy_step_end": 202,
+      "phrase": "hovering idle near blocks",
+      "description": "hovering idle near blocks"
+    },
+    {
+      "segment_index": 8,
+      "t_start_sec": 24.07920792079208,
+      "t_end_sec": 25.366336633663366,
+      "control_step_start": 1216,
+      "control_step_end": 1280,
+      "policy_step_start": 202,
+      "policy_step_end": 213,
+      "phrase": "idle / no progress",
+      "description": "idle / no progress"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "placement_or_insertion_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 7.00990099009901,
+    "vlm_failure_onset_timestamp": "00:07.0",
+    "vlm_failure_onset_frame": 354,
+    "vlm_failure_onset_step": 354,
+    "vlm_coarse_failure_onset_seconds": 8.0,
+    "vlm_failure_window_start_seconds": 7.0,
+    "vlm_failure_window_end_seconds": 9.0,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot carries the yellow block to the wooden block but misaligns its position during release, placing the yellow block beside the wooden block instead of stacking it on top.",
+    "vlm_failure_reason_summary": "misaligned placement; yellow block placed beside wooden block",
+    "vlm_recovery_action": "Re-grasp the yellow block, raise it above the wooden block, align it directly over the top surface, and place it down cleanly.",
+    "vlm_recovery_action_summary": "re-grasp yellow block; align and stack on wooden block",
+    "vlm_prevention_action": "Position the yellow block vertically directly above the center of the wooden block before releasing the gripper to ensure a stable stacked placement.",
+    "vlm_prevention_action_summary": "center yellow block above wooden block; release directly on top",
+    "vlm_justification": "The robot successfully grasps the yellow block and brings it to the target area, but at around 8.0 seconds, it misaligns the placement attempt, causing the yellow block to slip down beside the wooden block rather than resting on top.",
+    "vlm_temporal_justification": "At second 102 (01:42 in the clip), the robot lowers the yellow block down to the table surface directly beside the wooden block instead of aligning it on top, making the placement misalignment visibly apparent and committed.",
+    "vlm_video_fps": 50.5,
+    "vlm_video_num_frames": 1281,
+    "vlm_video_duration_seconds": 25.346534653465348,
+    "vlm_usage": {
+      "input_tokens": 22667,
+      "output_tokens": 394,
+      "cached_tokens": 1701
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"placement_or_insertion_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 8.0,\n \"failure_window_start_seconds\": 7.0,\n \"failure_window_end_seconds\": 9.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot carries the yellow block to the wooden block but misaligns its position during release, placing the yellow block beside the wooden block instead of stacking it on top.\",\n \"failure_reason_summary\": \"misaligned placement; yellow block placed beside wooden block\",\n \"recovery_action\": \"Re-grasp the yellow block, raise it above the wooden block, align it directly over the top surface, and place it down cleanly.\",\n \"recovery_action_summary\": \"re-grasp yellow block; align and stack on wooden block\",\n \"prevention_action\": \"Position the yellow block vertically directly above the center of the wooden block before releasing the gripper to ensure a stable stacked placement.\",\n \"prevention_action_summary\": \"center yellow block above wooden block; release directly on top\",\n \"justification\": \"The robot successfully grasps the yellow block and brings it to the target area, but at around 8.0 seconds, it misaligns the placement attempt, causing the yellow block to slip down beside the wooden block rather than resting on top.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 354,
+    "failure_sim_step": 354,
+    "failure_policy_step": 59,
+    "failure_chunk_index": 0,
+    "first_post_failure_policy_step": 60,
+    "failure_type": "placement_or_insertion_failure",
+    "correction_action": "Re-grasp the yellow block, raise it above the wooden block, align it directly over the top surface, and place it down cleanly."
+  }
+}
\ No newline at end of file
diff --git a/blocks/ep048/labels.npz b/blocks/ep048/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..8269ec97ccc3335af6da389768da8cbb11c67583
--- /dev/null
+++ b/blocks/ep048/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c909e0d17765e7b0986c797db21cf0f4424aa90761efcee2cd3850bc13213f60
+size 4095
diff --git a/blocks/ep049/example.md b/blocks/ep049/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..a1c688e107a1ff0c9e386e034576e327212e0438
--- /dev/null
+++ b/blocks/ep049/example.md
@@ -0,0 +1,36 @@
+# t000__ep049__blocks
+
+- **suite / variant:** `` / ``
+- **instruction:** Stack the yellow block on the wooden block
+- **success:** False
+- **control steps:** 844  (16.8s at 50 Hz)
+- **policy inferences:** 147  (replan_steps=6)
+- **video:** `blocks/ep049/rollout.mp4`
+- **features:** `147` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | lowering toward yellow block |
+| 3.0–6.0 | 151–301 | 25–50 | aligning and grasping yellow block |
+| 6.0–9.0 | 302–452 | 50–75 | carrying yellow block toward wooden block |
+| 9.0–12.0 | 453–603 | 75–100 | lowering yellow block and tilting gripper |
+| 12.0–15.0 | 604–754 | 100–125 | pressing tilted gripper on wooden block |
+| 15.0–16.8 | 755–843 | 125–140 | holding tilted position over wooden block |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 11.135458167330677 s (00:11.1)
+- **frame / control_step:** 559 (policy_step=93, chunk_index=1)
+- **coarse onset:** 10.8 s (refined=True)
+- **mode:** placement_or_insertion_failure
+- **reason:** The robot brought the yellow block over the wooden block but severely tilted its end-effector during release, causing the yellow block to sit misaligned and unstable on the wooden block rather than stacking cleanly.
+  Summary: improper alignment during stacking; tilted gripper release
+- **recovery:** Lift the end-effector, re-grasp the yellow block, position it directly above the wooden block in an upright orientation, and gently set it down to complete the stack.
+  Summary: re-grasp yellow block; align and stack upright
+- **prevention:** Maintain a vertical end-effector orientation when approaching the wooden block and release the yellow block gently once aligned.
+  Summary: maintain vertical orientation; release block gently
+- **justification:** None
diff --git a/blocks/ep049/labels.json b/blocks/ep049/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..5429f43eee97643a535f136f9c0b2959f8a3b7f7
--- /dev/null
+++ b/blocks/ep049/labels.json
@@ -0,0 +1,131 @@
+{
+  "rollout_id": "t000__ep049__blocks",
+  "scene_variant": "",
+  "task_id": 0,
+  "task_file": "",
+  "instruction": "Stack the yellow block on the wooden block",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:19:50+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00796812749004,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "lowering toward yellow block",
+      "description": "lowering toward yellow block"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00796812749004,
+      "t_end_sec": 6.01593625498008,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "aligning and grasping yellow block",
+      "description": "aligning and grasping yellow block"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01593625498008,
+      "t_end_sec": 9.02390438247012,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "carrying yellow block toward wooden block",
+      "description": "carrying yellow block toward wooden block"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.02390438247012,
+      "t_end_sec": 12.03187250996016,
+      "control_step_start": 453,
+      "control_step_end": 603,
+      "policy_step_start": 75,
+      "policy_step_end": 100,
+      "phrase": "lowering yellow block and tilting gripper",
+      "description": "lowering yellow block and tilting gripper"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 12.03187250996016,
+      "t_end_sec": 15.0398406374502,
+      "control_step_start": 604,
+      "control_step_end": 754,
+      "policy_step_start": 100,
+      "policy_step_end": 125,
+      "phrase": "pressing tilted gripper on wooden block",
+      "description": "pressing tilted gripper on wooden block"
+    },
+    {
+      "segment_index": 5,
+      "t_start_sec": 15.0398406374502,
+      "t_end_sec": 16.812749003984063,
+      "control_step_start": 755,
+      "control_step_end": 843,
+      "policy_step_start": 125,
+      "policy_step_end": 140,
+      "phrase": "holding tilted position over wooden block",
+      "description": "holding tilted position over wooden block"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "placement_or_insertion_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 11.135458167330677,
+    "vlm_failure_onset_timestamp": "00:11.1",
+    "vlm_failure_onset_frame": 559,
+    "vlm_failure_onset_step": 559,
+    "vlm_coarse_failure_onset_seconds": 10.8,
+    "vlm_failure_window_start_seconds": 10.0,
+    "vlm_failure_window_end_seconds": 11.8,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot brought the yellow block over the wooden block but severely tilted its end-effector during release, causing the yellow block to sit misaligned and unstable on the wooden block rather than stacking cleanly.",
+    "vlm_failure_reason_summary": "improper alignment during stacking; tilted gripper release",
+    "vlm_recovery_action": "Lift the end-effector, re-grasp the yellow block, position it directly above the wooden block in an upright orientation, and gently set it down to complete the stack.",
+    "vlm_recovery_action_summary": "re-grasp yellow block; align and stack upright",
+    "vlm_prevention_action": "Maintain a vertical end-effector orientation when approaching the wooden block and release the yellow block gently once aligned.",
+    "vlm_prevention_action_summary": "maintain vertical orientation; release block gently",
+    "vlm_justification": null,
+    "vlm_temporal_justification": "At second 168 (02:48 in the clip), the robot's end-effector begins tilting unnaturally forward while descending over the wooden block, causing the yellow block to approach at a tilted angle and land misaligned and unstable.",
+    "vlm_video_fps": 50.2,
+    "vlm_video_num_frames": 844,
+    "vlm_video_duration_seconds": 16.792828685258964,
+    "vlm_usage": {
+      "input_tokens": 22043,
+      "output_tokens": 347,
+      "cached_tokens": 1157
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"placement_or_insertion_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 10.8,\n \"failure_window_start_seconds\": 10.0,\n \"failure_window_end_seconds\": 11.8,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot brought the yellow block over the wooden block but severely tilted its end-effector during release, causing the yellow block to sit misaligned and unstable on the wooden block rather than stacking cleanly.\",\n \"failure_reason_summary\": \"improper alignment during stacking; tilted gripper release\",\n \"recovery_action\": \"Lift the end-effector, re-grasp the yellow block, position it directly above the wooden block in an upright orientation, and gently set it down to complete the stack.\",\n \"recovery_action_summary\": \"re-grasp yellow block; align and stack upright\",\n \"prevention_action\": \"Maintain a vertical end-effector orientation when approaching the wooden block and release the yellow block gently once aligned.\",\n \"prevention_action_summary\": \"maintain vertical orientation; release block gently\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 559,
+    "failure_sim_step": 559,
+    "failure_policy_step": 93,
+    "failure_chunk_index": 1,
+    "first_post_failure_policy_step": 94,
+    "failure_type": "placement_or_insertion_failure",
+    "correction_action": "Lift the end-effector, re-grasp the yellow block, position it directly above the wooden block in an upright orientation, and gently set it down to complete the stack."
+  }
+}
\ No newline at end of file
diff --git a/blocks/ep049/labels.npz b/blocks/ep049/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..7f83b8d2e38d031fe87dcd395928fe1c455bfc2d
--- /dev/null
+++ b/blocks/ep049/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c5d43e9e8c913a82b85ae88b80935e8be73e5b7411f74cb10d048e858326a767
+size 4050
diff --git a/closer_drawer/ep000/example.md b/closer_drawer/ep000/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..8322bfe37dffeb17b1a3208a06cce3abca515378
--- /dev/null
+++ b/closer_drawer/ep000/example.md
@@ -0,0 +1,23 @@
+# t006__ep000__closer_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Close the drawer
+- **success:** True
+- **control steps:** 624  (12.3s at 51 Hz)
+- **policy inferences:** 108  (replan_steps=5)
+- **video:** `closer_drawer/ep000/rollout.mp4`
+- **features:** `108` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–30 | repositioning arm above table |
+| 3.0–6.0 | 152–303 | 30–60 | lowering arm toward drawer |
+| 6.0–9.0 | 304–455 | 60–91 | approaching open drawer |
+| 9.0–12.0 | 456–607 | 91–121 | pushing drawer closed |
+| 12.0–12.3 | 608–623 | 121–124 | completing drawer closure |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/closer_drawer/ep000/labels.json b/closer_drawer/ep000/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..a10b439096ba9dfbddabe599e449a2edb5a3260c
--- /dev/null
+++ b/closer_drawer/ep000/labels.json
@@ -0,0 +1,90 @@
+{
+  "rollout_id": "t006__ep000__closer_drawer",
+  "scene_variant": "",
+  "task_id": 6,
+  "task_file": "",
+  "instruction": "Close the drawer",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:20:00+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9921259842519685,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 30,
+      "phrase": "repositioning arm above table",
+      "description": "repositioning arm above table"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9921259842519685,
+      "t_end_sec": 5.984251968503937,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 30,
+      "policy_step_end": 60,
+      "phrase": "lowering arm toward drawer",
+      "description": "lowering arm toward drawer"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.984251968503937,
+      "t_end_sec": 8.976377952755906,
+      "control_step_start": 304,
+      "control_step_end": 455,
+      "policy_step_start": 60,
+      "policy_step_end": 91,
+      "phrase": "approaching open drawer",
+      "description": "approaching open drawer"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.976377952755906,
+      "t_end_sec": 11.968503937007874,
+      "control_step_start": 456,
+      "control_step_end": 607,
+      "policy_step_start": 91,
+      "policy_step_end": 121,
+      "phrase": "pushing drawer closed",
+      "description": "pushing drawer closed"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 11.968503937007874,
+      "t_end_sec": 12.283464566929135,
+      "control_step_start": 608,
+      "control_step_end": 623,
+      "policy_step_start": 121,
+      "policy_step_end": 124,
+      "phrase": "completing drawer closure",
+      "description": "completing drawer closure"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/closer_drawer/ep000/labels.npz b/closer_drawer/ep000/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..f847f56568f53324c23f932d87671dbdc0b50d5b
--- /dev/null
+++ b/closer_drawer/ep000/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2b8313a5ac6520cad79897c4e0d01dd88c1c5a9d8c94ad39c744858e0431b8d2
+size 3165
diff --git a/closer_drawer/ep001/example.md b/closer_drawer/ep001/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..7888d6a114d17d6e0c4500b72035c688009ca5af
--- /dev/null
+++ b/closer_drawer/ep001/example.md
@@ -0,0 +1,23 @@
+# t006__ep001__closer_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Close the drawer
+- **success:** True
+- **control steps:** 646  (12.8s at 50 Hz)
+- **policy inferences:** 112  (replan_steps=6)
+- **video:** `closer_drawer/ep001/rollout.mp4`
+- **features:** `112` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | moving arm towards workspace |
+| 3.0–6.0 | 151–301 | 25–50 | hovering above table |
+| 6.0–9.0 | 302–452 | 50–75 | approaching open drawer |
+| 9.0–12.0 | 453–603 | 75–100 | pushing drawer closed |
+| 12.0–12.8 | 604–645 | 100–107 | retracting arm |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/closer_drawer/ep001/labels.json b/closer_drawer/ep001/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..be0dc8ae8bbbd0f845f23d356cc7ae573955da93
--- /dev/null
+++ b/closer_drawer/ep001/labels.json
@@ -0,0 +1,90 @@
+{
+  "rollout_id": "t006__ep001__closer_drawer",
+  "scene_variant": "",
+  "task_id": 6,
+  "task_file": "",
+  "instruction": "Close the drawer",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:20:09+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.996031746031746,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "moving arm towards workspace",
+      "description": "moving arm towards workspace"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.996031746031746,
+      "t_end_sec": 5.992063492063492,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "hovering above table",
+      "description": "hovering above table"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.992063492063492,
+      "t_end_sec": 8.988095238095239,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "approaching open drawer",
+      "description": "approaching open drawer"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.988095238095239,
+      "t_end_sec": 11.984126984126984,
+      "control_step_start": 453,
+      "control_step_end": 603,
+      "policy_step_start": 75,
+      "policy_step_end": 100,
+      "phrase": "pushing drawer closed",
+      "description": "pushing drawer closed"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 11.984126984126984,
+      "t_end_sec": 12.817460317460318,
+      "control_step_start": 604,
+      "control_step_end": 645,
+      "policy_step_start": 100,
+      "policy_step_end": 107,
+      "phrase": "retracting arm",
+      "description": "retracting arm"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/closer_drawer/ep001/labels.npz b/closer_drawer/ep001/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..b25451ac178e869ee37d027d51e9f700a0769991
--- /dev/null
+++ b/closer_drawer/ep001/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:306cc4537fa05d98b2d498eee5d230e61d2450a4eb2fdcebad7ea45f21fa1ecb
+size 3154
diff --git a/closer_drawer/ep002/example.md b/closer_drawer/ep002/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..0c931efdffebf9cf11975c1ea83129ffc6f39093
--- /dev/null
+++ b/closer_drawer/ep002/example.md
@@ -0,0 +1,23 @@
+# t006__ep002__closer_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Close the drawer
+- **success:** True
+- **control steps:** 683  (13.6s at 50 Hz)
+- **policy inferences:** 118  (replan_steps=6)
+- **video:** `closer_drawer/ep002/rollout.mp4`
+- **features:** `118` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | repositioning robot arm |
+| 3.0–6.0 | 151–301 | 25–50 | lowering gripper toward table |
+| 6.0–9.0 | 302–452 | 50–75 | approaching the open drawer |
+| 9.0–12.0 | 453–603 | 75–100 | pushing drawer closed |
+| 12.0–13.6 | 604–682 | 100–113 | retracting arm away |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/closer_drawer/ep002/labels.json b/closer_drawer/ep002/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..adfd56fc776e78978d0c63fb6826e9d4f90fd929
--- /dev/null
+++ b/closer_drawer/ep002/labels.json
@@ -0,0 +1,90 @@
+{
+  "rollout_id": "t006__ep002__closer_drawer",
+  "scene_variant": "",
+  "task_id": 6,
+  "task_file": "",
+  "instruction": "Close the drawer",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:20:20+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.001988071570577,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "repositioning robot arm",
+      "description": "repositioning robot arm"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.001988071570577,
+      "t_end_sec": 6.003976143141154,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "lowering gripper toward table",
+      "description": "lowering gripper toward table"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.003976143141154,
+      "t_end_sec": 9.00596421471173,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "approaching the open drawer",
+      "description": "approaching the open drawer"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.00596421471173,
+      "t_end_sec": 12.007952286282308,
+      "control_step_start": 453,
+      "control_step_end": 603,
+      "policy_step_start": 75,
+      "policy_step_end": 100,
+      "phrase": "pushing drawer closed",
+      "description": "pushing drawer closed"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 12.007952286282308,
+      "t_end_sec": 13.578528827037774,
+      "control_step_start": 604,
+      "control_step_end": 682,
+      "policy_step_start": 100,
+      "policy_step_end": 113,
+      "phrase": "retracting arm away",
+      "description": "retracting arm away"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/closer_drawer/ep002/labels.npz b/closer_drawer/ep002/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..182a5d2c470eee86efdcd0cfafa8563a6ad22e27
--- /dev/null
+++ b/closer_drawer/ep002/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7083da038f8a2453f043c9d9c79ab5a3cea26bf3cb2cf1d2ce6ac632e94bf851
+size 3184
diff --git a/closer_drawer/ep003/example.md b/closer_drawer/ep003/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..9e8bc8037c0ac302e03cd563d6684e90cdc7151c
--- /dev/null
+++ b/closer_drawer/ep003/example.md
@@ -0,0 +1,22 @@
+# t006__ep003__closer_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Close the drawer
+- **success:** True
+- **control steps:** 575  (11.4s at 50 Hz)
+- **policy inferences:** 99  (replan_steps=6)
+- **video:** `closer_drawer/ep003/rollout.mp4`
+- **features:** `99` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | moving towards the drawer |
+| 3.0–6.0 | 151–301 | 25–50 | lowering gripper near drawer |
+| 6.0–9.0 | 302–452 | 50–75 | pushing top drawer closed |
+| 9.0–11.4 | 453–574 | 75–95 | retracting arm upwards |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/closer_drawer/ep003/labels.json b/closer_drawer/ep003/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..771e8a752546358b8e81b88138c589b9034203c2
--- /dev/null
+++ b/closer_drawer/ep003/labels.json
@@ -0,0 +1,79 @@
+{
+  "rollout_id": "t006__ep003__closer_drawer",
+  "scene_variant": "",
+  "task_id": 6,
+  "task_file": "",
+  "instruction": "Close the drawer",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:20:29+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.996031746031746,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "moving towards the drawer",
+      "description": "moving towards the drawer"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.996031746031746,
+      "t_end_sec": 5.992063492063492,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "lowering gripper near drawer",
+      "description": "lowering gripper near drawer"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.992063492063492,
+      "t_end_sec": 8.988095238095239,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "pushing top drawer closed",
+      "description": "pushing top drawer closed"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.988095238095239,
+      "t_end_sec": 11.408730158730158,
+      "control_step_start": 453,
+      "control_step_end": 574,
+      "policy_step_start": 75,
+      "policy_step_end": 95,
+      "phrase": "retracting arm upwards",
+      "description": "retracting arm upwards"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/closer_drawer/ep003/labels.npz b/closer_drawer/ep003/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..70f0735efd9d3fd3d511442456e98d6b90fd0367
--- /dev/null
+++ b/closer_drawer/ep003/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f670560fb99fd321abe26919d58e7b0f488547c333a6e8e08648eec4c565eb9f
+size 3127
diff --git a/closer_drawer/ep004/example.md b/closer_drawer/ep004/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..22b2b267640adbcd4037bbbe8dd4a1b9b7cbcfe7
--- /dev/null
+++ b/closer_drawer/ep004/example.md
@@ -0,0 +1,23 @@
+# t006__ep004__closer_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Close the drawer
+- **success:** True
+- **control steps:** 704  (14.0s at 50 Hz)
+- **policy inferences:** 122  (replan_steps=6)
+- **video:** `closer_drawer/ep004/rollout.mp4`
+- **features:** `122` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | moving towards drawer |
+| 3.0–6.0 | 151–301 | 25–50 | positioning gripper near drawer |
+| 6.0–9.0 | 302–452 | 50–75 | approaching and pushing drawer |
+| 9.0–12.0 | 453–603 | 75–100 | closing drawer and retracting |
+| 12.0–14.0 | 604–703 | 100–117 | retracting robot arm |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/closer_drawer/ep004/labels.json b/closer_drawer/ep004/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..c654d6272b41c143fcead4e8caaefdfdca73bd02
--- /dev/null
+++ b/closer_drawer/ep004/labels.json
@@ -0,0 +1,90 @@
+{
+  "rollout_id": "t006__ep004__closer_drawer",
+  "scene_variant": "",
+  "task_id": 6,
+  "task_file": "",
+  "instruction": "Close the drawer",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:20:39+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.001988071570577,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "moving towards drawer",
+      "description": "moving towards drawer"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.001988071570577,
+      "t_end_sec": 6.003976143141154,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "positioning gripper near drawer",
+      "description": "positioning gripper near drawer"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.003976143141154,
+      "t_end_sec": 9.00596421471173,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "approaching and pushing drawer",
+      "description": "approaching and pushing drawer"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.00596421471173,
+      "t_end_sec": 12.007952286282308,
+      "control_step_start": 453,
+      "control_step_end": 603,
+      "policy_step_start": 75,
+      "policy_step_end": 100,
+      "phrase": "closing drawer and retracting",
+      "description": "closing drawer and retracting"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 12.007952286282308,
+      "t_end_sec": 13.996023856858848,
+      "control_step_start": 604,
+      "control_step_end": 703,
+      "policy_step_start": 100,
+      "policy_step_end": 117,
+      "phrase": "retracting robot arm",
+      "description": "retracting robot arm"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/closer_drawer/ep004/labels.npz b/closer_drawer/ep004/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..f2f1e9ef48ce3248c94b99d00ed9e84072237589
--- /dev/null
+++ b/closer_drawer/ep004/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5c8e7e8c7ce37dbcdbe1977faf254ad47efe79ce55e7889fc5a5d27b55b85d79
+size 3164
diff --git a/closer_drawer/ep005/example.md b/closer_drawer/ep005/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..1e30e4325c273a252426533868b334ebd5d26240
--- /dev/null
+++ b/closer_drawer/ep005/example.md
@@ -0,0 +1,23 @@
+# t006__ep005__closer_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Close the drawer
+- **success:** True
+- **control steps:** 715  (14.2s at 50 Hz)
+- **policy inferences:** 124  (replan_steps=6)
+- **video:** `closer_drawer/ep005/rollout.mp4`
+- **features:** `124` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | repositioning robot arm |
+| 3.0–6.0 | 151–301 | 25–50 | lowering arm toward table |
+| 6.0–9.0 | 302–452 | 50–75 | approaching the open drawer |
+| 9.0–12.0 | 453–603 | 75–100 | pushing drawer shut |
+| 12.0–14.2 | 604–714 | 100–119 | retracting arm |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/closer_drawer/ep005/labels.json b/closer_drawer/ep005/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..cead7cd6af97066e4758b7fe10d781efadac1665
--- /dev/null
+++ b/closer_drawer/ep005/labels.json
@@ -0,0 +1,90 @@
+{
+  "rollout_id": "t006__ep005__closer_drawer",
+  "scene_variant": "",
+  "task_id": 6,
+  "task_file": "",
+  "instruction": "Close the drawer",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:20:49+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.996031746031746,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "repositioning robot arm",
+      "description": "repositioning robot arm"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.996031746031746,
+      "t_end_sec": 5.992063492063492,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "lowering arm toward table",
+      "description": "lowering arm toward table"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.992063492063492,
+      "t_end_sec": 8.988095238095239,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "approaching the open drawer",
+      "description": "approaching the open drawer"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.988095238095239,
+      "t_end_sec": 11.984126984126984,
+      "control_step_start": 453,
+      "control_step_end": 603,
+      "policy_step_start": 75,
+      "policy_step_end": 100,
+      "phrase": "pushing drawer shut",
+      "description": "pushing drawer shut"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 11.984126984126984,
+      "t_end_sec": 14.186507936507937,
+      "control_step_start": 604,
+      "control_step_end": 714,
+      "policy_step_start": 100,
+      "policy_step_end": 119,
+      "phrase": "retracting arm",
+      "description": "retracting arm"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/closer_drawer/ep005/labels.npz b/closer_drawer/ep005/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..52c1345f61d4f7c7df5a3a375d8e84d217d5552e
--- /dev/null
+++ b/closer_drawer/ep005/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bf2fe4465587bcfc62dec2033b0a1ac45fe259a7a211ddf3768105088860bf72
+size 3162
diff --git a/closer_drawer/ep006/example.md b/closer_drawer/ep006/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..b0a45d27f86f875ac1857e246ffb3ed12a7ae855
--- /dev/null
+++ b/closer_drawer/ep006/example.md
@@ -0,0 +1,23 @@
+# t006__ep006__closer_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Close the drawer
+- **success:** True
+- **control steps:** 692  (13.4s at 52 Hz)
+- **policy inferences:** 118  (replan_steps=6)
+- **video:** `closer_drawer/ep006/rollout.mp4`
+- **features:** `118` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–153 | 0–25 | repositioning arm above desk |
+| 3.0–6.0 | 154–307 | 25–51 | lowering towards drawer |
+| 6.0–9.0 | 308–461 | 51–76 | approaching drawer handle |
+| 9.0–12.0 | 462–615 | 77–102 | pushing drawer shut |
+| 12.0–13.4 | 616–691 | 102–115 | retracting arm |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/closer_drawer/ep006/labels.json b/closer_drawer/ep006/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..7c67ad642e707d8b2d0fc01ba8c3441eef4fc8cc
--- /dev/null
+++ b/closer_drawer/ep006/labels.json
@@ -0,0 +1,90 @@
+{
+  "rollout_id": "t006__ep006__closer_drawer",
+  "scene_variant": "",
+  "task_id": 6,
+  "task_file": "",
+  "instruction": "Close the drawer",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:20:59+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9902912621359223,
+      "control_step_start": 0,
+      "control_step_end": 153,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "repositioning arm above desk",
+      "description": "repositioning arm above desk"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9902912621359223,
+      "t_end_sec": 5.980582524271845,
+      "control_step_start": 154,
+      "control_step_end": 307,
+      "policy_step_start": 25,
+      "policy_step_end": 51,
+      "phrase": "lowering towards drawer",
+      "description": "lowering towards drawer"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.980582524271845,
+      "t_end_sec": 8.970873786407767,
+      "control_step_start": 308,
+      "control_step_end": 461,
+      "policy_step_start": 51,
+      "policy_step_end": 76,
+      "phrase": "approaching drawer handle",
+      "description": "approaching drawer handle"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.970873786407767,
+      "t_end_sec": 11.96116504854369,
+      "control_step_start": 462,
+      "control_step_end": 615,
+      "policy_step_start": 77,
+      "policy_step_end": 102,
+      "phrase": "pushing drawer shut",
+      "description": "pushing drawer shut"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 11.96116504854369,
+      "t_end_sec": 13.436893203883495,
+      "control_step_start": 616,
+      "control_step_end": 691,
+      "policy_step_start": 102,
+      "policy_step_end": 115,
+      "phrase": "retracting arm",
+      "description": "retracting arm"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/closer_drawer/ep006/labels.npz b/closer_drawer/ep006/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..0038ce9663dc3b4632290100e79cf91890d64f2b
--- /dev/null
+++ b/closer_drawer/ep006/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9cbc50a015c60cdb67c75f036d5512eee29eda95db98d930f5414f5bcaffeb86
+size 3158
diff --git a/closer_drawer/ep007/example.md b/closer_drawer/ep007/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..bf46870e91cf6b64a4ec0c21d0f2ef4467277ec2
--- /dev/null
+++ b/closer_drawer/ep007/example.md
@@ -0,0 +1,23 @@
+# t006__ep007__closer_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Close the drawer
+- **success:** True
+- **control steps:** 687  (13.7s at 50 Hz)
+- **policy inferences:** 118  (replan_steps=6)
+- **video:** `closer_drawer/ep007/rollout.mp4`
+- **features:** `118` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | reaching toward drawer |
+| 3.0–6.0 | 151–301 | 25–50 | approaching open drawer |
+| 6.0–9.0 | 302–452 | 50–75 | aligning gripper with drawer |
+| 9.0–12.0 | 453–603 | 75–100 | pushing drawer closed |
+| 12.0–13.7 | 604–686 | 100–114 | retracting arm upward |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/closer_drawer/ep007/labels.json b/closer_drawer/ep007/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..8a1b55b79ca0fdf05cb824e71648ecf8c081d4e0
--- /dev/null
+++ b/closer_drawer/ep007/labels.json
@@ -0,0 +1,90 @@
+{
+  "rollout_id": "t006__ep007__closer_drawer",
+  "scene_variant": "",
+  "task_id": 6,
+  "task_file": "",
+  "instruction": "Close the drawer",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:21:07+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00796812749004,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching toward drawer",
+      "description": "reaching toward drawer"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00796812749004,
+      "t_end_sec": 6.01593625498008,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "approaching open drawer",
+      "description": "approaching open drawer"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01593625498008,
+      "t_end_sec": 9.02390438247012,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "aligning gripper with drawer",
+      "description": "aligning gripper with drawer"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.02390438247012,
+      "t_end_sec": 12.03187250996016,
+      "control_step_start": 453,
+      "control_step_end": 603,
+      "policy_step_start": 75,
+      "policy_step_end": 100,
+      "phrase": "pushing drawer closed",
+      "description": "pushing drawer closed"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 12.03187250996016,
+      "t_end_sec": 13.685258964143426,
+      "control_step_start": 604,
+      "control_step_end": 686,
+      "policy_step_start": 100,
+      "policy_step_end": 114,
+      "phrase": "retracting arm upward",
+      "description": "retracting arm upward"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/closer_drawer/ep007/labels.npz b/closer_drawer/ep007/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..b6081210edff79e887072b762eaa06f0c86e6df3
--- /dev/null
+++ b/closer_drawer/ep007/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:26d581670af98031eb230cccb507b93d35a8d38a3b59b4ec0ec2a800d48650dc
+size 3155
diff --git a/closer_drawer/ep008/example.md b/closer_drawer/ep008/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..d7eb2f7eb2c28c89d42e4d3142d4efc68b3e247e
--- /dev/null
+++ b/closer_drawer/ep008/example.md
@@ -0,0 +1,23 @@
+# t006__ep008__closer_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Close the drawer
+- **success:** True
+- **control steps:** 681  (13.5s at 50 Hz)
+- **policy inferences:** 117  (replan_steps=6)
+- **video:** `closer_drawer/ep008/rollout.mp4`
+- **features:** `117` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | moving arm toward drawer |
+| 3.0–6.0 | 151–301 | 25–50 | lowering gripper near drawer |
+| 6.0–9.0 | 302–452 | 50–75 | aligning gripper with drawer |
+| 9.0–12.0 | 453–603 | 75–100 | pushing drawer closed |
+| 12.0–13.5 | 604–680 | 100–113 | retracting arm from drawer |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/closer_drawer/ep008/labels.json b/closer_drawer/ep008/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..a323b362c23a399a173d494ff5e7fa720f3f5b58
--- /dev/null
+++ b/closer_drawer/ep008/labels.json
@@ -0,0 +1,90 @@
+{
+  "rollout_id": "t006__ep008__closer_drawer",
+  "scene_variant": "",
+  "task_id": 6,
+  "task_file": "",
+  "instruction": "Close the drawer",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:21:17+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.996031746031746,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "moving arm toward drawer",
+      "description": "moving arm toward drawer"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.996031746031746,
+      "t_end_sec": 5.992063492063492,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "lowering gripper near drawer",
+      "description": "lowering gripper near drawer"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.992063492063492,
+      "t_end_sec": 8.988095238095239,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "aligning gripper with drawer",
+      "description": "aligning gripper with drawer"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.988095238095239,
+      "t_end_sec": 11.984126984126984,
+      "control_step_start": 453,
+      "control_step_end": 603,
+      "policy_step_start": 75,
+      "policy_step_end": 100,
+      "phrase": "pushing drawer closed",
+      "description": "pushing drawer closed"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 11.984126984126984,
+      "t_end_sec": 13.511904761904763,
+      "control_step_start": 604,
+      "control_step_end": 680,
+      "policy_step_start": 100,
+      "policy_step_end": 113,
+      "phrase": "retracting arm from drawer",
+      "description": "retracting arm from drawer"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/closer_drawer/ep008/labels.npz b/closer_drawer/ep008/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..c3e10d20813ddfd726fb070e72a67e88fdeff037
--- /dev/null
+++ b/closer_drawer/ep008/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:11a681b6fa6c368443751e60b2b72acd2c96108b3c3ae73abbfd8fbdc1cc571c
+size 3151
diff --git a/closer_drawer/ep009/example.md b/closer_drawer/ep009/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..b8294acf260ac3743e2fa5cc1ede5c9bb9c359fa
--- /dev/null
+++ b/closer_drawer/ep009/example.md
@@ -0,0 +1,23 @@
+# t006__ep009__closer_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Close the drawer
+- **success:** True
+- **control steps:** 671  (13.3s at 51 Hz)
+- **policy inferences:** 115  (replan_steps=6)
+- **video:** `closer_drawer/ep009/rollout.mp4`
+- **features:** `115` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | moving arm towards drawer |
+| 3.0–6.0 | 152–303 | 25–50 | approaching the open drawer |
+| 6.0–9.0 | 304–455 | 50–75 | aligning gripper with drawer handle |
+| 9.0–12.0 | 456–607 | 76–101 | pushing drawer shut |
+| 12.0–13.3 | 608–670 | 101–111 | retracting arm away |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/closer_drawer/ep009/labels.json b/closer_drawer/ep009/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..3e1373101cf6242d222cd935a8e4c683b9c84aee
--- /dev/null
+++ b/closer_drawer/ep009/labels.json
@@ -0,0 +1,90 @@
+{
+  "rollout_id": "t006__ep009__closer_drawer",
+  "scene_variant": "",
+  "task_id": 6,
+  "task_file": "",
+  "instruction": "Close the drawer",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:21:26+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0039525691699605,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "moving arm towards drawer",
+      "description": "moving arm towards drawer"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0039525691699605,
+      "t_end_sec": 6.007905138339921,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "approaching the open drawer",
+      "description": "approaching the open drawer"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.007905138339921,
+      "t_end_sec": 9.011857707509881,
+      "control_step_start": 304,
+      "control_step_end": 455,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "aligning gripper with drawer handle",
+      "description": "aligning gripper with drawer handle"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.011857707509881,
+      "t_end_sec": 12.015810276679842,
+      "control_step_start": 456,
+      "control_step_end": 607,
+      "policy_step_start": 76,
+      "policy_step_end": 101,
+      "phrase": "pushing drawer shut",
+      "description": "pushing drawer shut"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 12.015810276679842,
+      "t_end_sec": 13.26086956521739,
+      "control_step_start": 608,
+      "control_step_end": 670,
+      "policy_step_start": 101,
+      "policy_step_end": 111,
+      "phrase": "retracting arm away",
+      "description": "retracting arm away"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/closer_drawer/ep009/labels.npz b/closer_drawer/ep009/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..3e2587f2ac9082ac42e2084c88ae452ef2dd2fc5
--- /dev/null
+++ b/closer_drawer/ep009/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:792ffb607b806d345224a28cb06474c31ad57750d496e8038cd9f5a64f921243
+size 3178
diff --git a/closer_drawer/ep010/example.md b/closer_drawer/ep010/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..757e4cff6696598688c03ab3119397751f456eee
--- /dev/null
+++ b/closer_drawer/ep010/example.md
@@ -0,0 +1,33 @@
+# t006__ep010__closer_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Close the drawer
+- **success:** False
+- **control steps:** 409  (8.0s at 51 Hz)
+- **policy inferences:** 70  (replan_steps=6)
+- **video:** `closer_drawer/ep010/rollout.mp4`
+- **features:** `70` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–152 | 0–25 | approaching the open drawer |
+| 3.0–6.0 | 153–305 | 25–50 | positioning gripper near drawer |
+| 6.0–8.0 | 306–408 | 51–68 | pushing drawer shut |
+
+## Failure
+
+- **yes**
+- **onset type:** timeout
+- **onset:** 8.0 s (00:08.0)
+- **frame / control_step:** 408 (policy_step=68, chunk_index=0)
+- **coarse onset:** 8.0 s (refined=False)
+- **mode:** timeout_or_insufficient_progress
+- **reason:** The robot executed the drawer closing sequence correctly by approaching the handle and pushing inward, but moved too slowly to fully close the drawer before the episode timeout.
+  Summary: insufficient execution speed; incomplete closure before timeout
+- **recovery:** Continue pushing the drawer inward until it is completely flush with the cabinet.
+  Summary: push drawer completely shut
+- **prevention:** Approach and push the drawer more quickly and directly to finish the task within the time limit.
+  Summary: increase motion speed; push drawer shut faster
+- **justification:** The robot made no observable mistakes or improper contact; it approached the drawer and began pushing it closed, but ran out of time at 8.0 seconds before completing the action.
diff --git a/closer_drawer/ep010/labels.json b/closer_drawer/ep010/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..1efd0a6db0392121b0e6a0f48d1ebddc716bd84f
--- /dev/null
+++ b/closer_drawer/ep010/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t006__ep010__closer_drawer",
+  "scene_variant": "",
+  "task_id": 6,
+  "task_file": "",
+  "instruction": "Close the drawer",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:21:46+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0,
+      "control_step_start": 0,
+      "control_step_end": 152,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "approaching the open drawer",
+      "description": "approaching the open drawer"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0,
+      "t_end_sec": 6.0,
+      "control_step_start": 153,
+      "control_step_end": 305,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "positioning gripper near drawer",
+      "description": "positioning gripper near drawer"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.0,
+      "t_end_sec": 8.019607843137255,
+      "control_step_start": 306,
+      "control_step_end": 408,
+      "policy_step_start": 51,
+      "policy_step_end": 68,
+      "phrase": "pushing drawer shut",
+      "description": "pushing drawer shut"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "timeout_or_insufficient_progress",
+    "vlm_failure_onset_type": "timeout",
+    "vlm_failure_onset_seconds": 8.0,
+    "vlm_failure_onset_timestamp": "00:08.0",
+    "vlm_failure_onset_frame": 408,
+    "vlm_failure_onset_step": 408,
+    "vlm_coarse_failure_onset_seconds": 8.0,
+    "vlm_failure_window_start_seconds": 7.5,
+    "vlm_failure_window_end_seconds": 8.0,
+    "vlm_temporal_refined": false,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot executed the drawer closing sequence correctly by approaching the handle and pushing inward, but moved too slowly to fully close the drawer before the episode timeout.",
+    "vlm_failure_reason_summary": "insufficient execution speed; incomplete closure before timeout",
+    "vlm_recovery_action": "Continue pushing the drawer inward until it is completely flush with the cabinet.",
+    "vlm_recovery_action_summary": "push drawer completely shut",
+    "vlm_prevention_action": "Approach and push the drawer more quickly and directly to finish the task within the time limit.",
+    "vlm_prevention_action_summary": "increase motion speed; push drawer shut faster",
+    "vlm_justification": "The robot made no observable mistakes or improper contact; it approached the drawer and began pushing it closed, but ran out of time at 8.0 seconds before completing the action.",
+    "vlm_temporal_justification": "No earlier failure event identified; using the final frame.",
+    "vlm_video_fps": 51.0,
+    "vlm_video_num_frames": 409,
+    "vlm_video_duration_seconds": 8.0,
+    "vlm_usage": {
+      "input_tokens": 1986,
+      "output_tokens": 261,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"timeout_or_insufficient_progress\",\n \"onset_type\": \"timeout\",\n \"failure_onset_seconds\": 8.0,\n \"failure_window_start_seconds\": 7.5,\n \"failure_window_end_seconds\": 8.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot executed the drawer closing sequence correctly by approaching the handle and pushing inward, but moved too slowly to fully close the drawer before the episode timeout.\",\n \"failure_reason_summary\": \"insufficient execution speed; incomplete closure before timeout\",\n \"recovery_action\": \"Continue pushing the drawer inward until it is completely flush with the cabinet.\",\n \"recovery_action_summary\": \"push drawer completely shut\",\n \"prevention_action\": \"Approach and push the drawer more quickly and directly to finish the task within the time limit.\",\n \"prevention_action_summary\": \"increase motion speed; push drawer shut faster\",\n \"justification\": \"The robot made no observable mistakes or improper contact; it approached the drawer and began pushing it closed, but ran out of time at 8.0 seconds before completing the action.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 408,
+    "failure_sim_step": 408,
+    "failure_policy_step": 68,
+    "failure_chunk_index": 0,
+    "first_post_failure_policy_step": 69,
+    "failure_type": "timeout_or_insufficient_progress",
+    "correction_action": "Continue pushing the drawer inward until it is completely flush with the cabinet."
+  }
+}
\ No newline at end of file
diff --git a/closer_drawer/ep010/labels.npz b/closer_drawer/ep010/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..f15f3d2ac7a53ad1166a3fab6285e1958df3d369
--- /dev/null
+++ b/closer_drawer/ep010/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:80b5330ade4bf9a5fefa4534894c0ea12f7ab356baeff253293542090ce507a3
+size 3768
diff --git a/closer_drawer/ep011/example.md b/closer_drawer/ep011/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..57ca80f5227cef867a9a1841bc88fe904d7706c5
--- /dev/null
+++ b/closer_drawer/ep011/example.md
@@ -0,0 +1,21 @@
+# t006__ep011__closer_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Close the drawer
+- **success:** True
+- **control steps:** 336  (6.7s at 50 Hz)
+- **policy inferences:** 58  (replan_steps=6)
+- **video:** `closer_drawer/ep011/rollout.mp4`
+- **features:** `58` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | reaching for the drawer |
+| 3.0–6.0 | 151–301 | 25–50 | pushing the drawer closed |
+| 6.0–6.7 | 302–335 | 50–55 | retracting from the drawer |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/closer_drawer/ep011/labels.json b/closer_drawer/ep011/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..726c00b493797e1b0c9e06035039d1dba8fc6bee
--- /dev/null
+++ b/closer_drawer/ep011/labels.json
@@ -0,0 +1,68 @@
+{
+  "rollout_id": "t006__ep011__closer_drawer",
+  "scene_variant": "",
+  "task_id": 6,
+  "task_file": "",
+  "instruction": "Close the drawer",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:21:55+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00796812749004,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for the drawer",
+      "description": "reaching for the drawer"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00796812749004,
+      "t_end_sec": 6.01593625498008,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "pushing the drawer closed",
+      "description": "pushing the drawer closed"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01593625498008,
+      "t_end_sec": 6.6932270916334655,
+      "control_step_start": 302,
+      "control_step_end": 335,
+      "policy_step_start": 50,
+      "policy_step_end": 55,
+      "phrase": "retracting from the drawer",
+      "description": "retracting from the drawer"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/closer_drawer/ep011/labels.npz b/closer_drawer/ep011/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..6402253782630387c6377e99b863971ea6982e8d
--- /dev/null
+++ b/closer_drawer/ep011/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ea7c47817d71c78d00f678abb9b162cc37dc1b9d539a1c2443c4e0deaa06b755
+size 3074
diff --git a/closer_drawer/ep012/example.md b/closer_drawer/ep012/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..7fcf3c60f11e371d73136661a5d1ad5589597049
--- /dev/null
+++ b/closer_drawer/ep012/example.md
@@ -0,0 +1,21 @@
+# t006__ep012__closer_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Close the drawer
+- **success:** True
+- **control steps:** 350  (6.9s at 50 Hz)
+- **policy inferences:** 61  (replan_steps=6)
+- **video:** `closer_drawer/ep012/rollout.mp4`
+- **features:** `61` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching for the drawer |
+| 3.0–6.0 | 152–303 | 25–50 | pushing the drawer closed |
+| 6.0–6.9 | 304–349 | 50–58 | retracting the robotic arm |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/closer_drawer/ep012/labels.json b/closer_drawer/ep012/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..74eec2ef97c05fe19317d18bcffe52bfcf4cae20
--- /dev/null
+++ b/closer_drawer/ep012/labels.json
@@ -0,0 +1,68 @@
+{
+  "rollout_id": "t006__ep012__closer_drawer",
+  "scene_variant": "",
+  "task_id": 6,
+  "task_file": "",
+  "instruction": "Close the drawer",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:22:02+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00990099009901,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for the drawer",
+      "description": "reaching for the drawer"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00990099009901,
+      "t_end_sec": 6.01980198019802,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "pushing the drawer closed",
+      "description": "pushing the drawer closed"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01980198019802,
+      "t_end_sec": 6.930693069306931,
+      "control_step_start": 304,
+      "control_step_end": 349,
+      "policy_step_start": 50,
+      "policy_step_end": 58,
+      "phrase": "retracting the robotic arm",
+      "description": "retracting the robotic arm"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/closer_drawer/ep012/labels.npz b/closer_drawer/ep012/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..9a37c35806cab84d39d699f8a7df9819506417f1
--- /dev/null
+++ b/closer_drawer/ep012/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:26af5f11a92b1234a6cf20c200baf53d8d62bcd6e69eb276b147b4a450abd192
+size 3083
diff --git a/closer_drawer/ep013/example.md b/closer_drawer/ep013/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..71693e03b39e26f3d598e67859e9dde2db2b626d
--- /dev/null
+++ b/closer_drawer/ep013/example.md
@@ -0,0 +1,32 @@
+# t006__ep013__closer_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Close the drawer
+- **success:** False
+- **control steps:** 282  (5.6s at 50 Hz)
+- **policy inferences:** 49  (replan_steps=6)
+- **video:** `closer_drawer/ep013/rollout.mp4`
+- **features:** `49` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | descending toward drawer |
+| 3.0–5.6 | 151–281 | 25–46 | positioning gripper near drawer |
+
+## Failure
+
+- **yes**
+- **onset type:** timeout
+- **onset:** 5.58648111332008 s (00:05.6)
+- **frame / control_step:** 281 (policy_step=46, chunk_index=5)
+- **coarse onset:** 5.59 s (refined=False)
+- **mode:** timeout_or_insufficient_progress
+- **reason:** The robot arm approached the open drawer and aligned near the handle, but failed to push the drawer closed before the episode time limit expired.
+  Summary: insufficient execution speed; drawer not closed before timeout
+- **recovery:** Align the gripper directly with the drawer handle and push the drawer inward until fully closed.
+  Summary: align with drawer handle; push drawer closed
+- **prevention:** Move more directly towards the drawer lip and push it shut earlier in the trajectory.
+  Summary: approach drawer directly; push shut promptly
+- **justification:** The robot moved towards the correct object and made no discrete error or collision, but progressed too slowly to execute the drawer closing motion before the 5.59-second timeout.
diff --git a/closer_drawer/ep013/labels.json b/closer_drawer/ep013/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..0b2463f1a8637c17fe237eb4d16473f648736885
--- /dev/null
+++ b/closer_drawer/ep013/labels.json
@@ -0,0 +1,87 @@
+{
+  "rollout_id": "t006__ep013__closer_drawer",
+  "scene_variant": "",
+  "task_id": 6,
+  "task_file": "",
+  "instruction": "Close the drawer",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:22:18+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.001988071570577,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "descending toward drawer",
+      "description": "descending toward drawer"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.001988071570577,
+      "t_end_sec": 5.606361829025845,
+      "control_step_start": 151,
+      "control_step_end": 281,
+      "policy_step_start": 25,
+      "policy_step_end": 46,
+      "phrase": "positioning gripper near drawer",
+      "description": "positioning gripper near drawer"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "timeout_or_insufficient_progress",
+    "vlm_failure_onset_type": "timeout",
+    "vlm_failure_onset_seconds": 5.58648111332008,
+    "vlm_failure_onset_timestamp": "00:05.6",
+    "vlm_failure_onset_frame": 281,
+    "vlm_failure_onset_step": 281,
+    "vlm_coarse_failure_onset_seconds": 5.59,
+    "vlm_failure_window_start_seconds": 5.09,
+    "vlm_failure_window_end_seconds": 5.59,
+    "vlm_temporal_refined": false,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot arm approached the open drawer and aligned near the handle, but failed to push the drawer closed before the episode time limit expired.",
+    "vlm_failure_reason_summary": "insufficient execution speed; drawer not closed before timeout",
+    "vlm_recovery_action": "Align the gripper directly with the drawer handle and push the drawer inward until fully closed.",
+    "vlm_recovery_action_summary": "align with drawer handle; push drawer closed",
+    "vlm_prevention_action": "Move more directly towards the drawer lip and push it shut earlier in the trajectory.",
+    "vlm_prevention_action_summary": "approach drawer directly; push shut promptly",
+    "vlm_justification": "The robot moved towards the correct object and made no discrete error or collision, but progressed too slowly to execute the drawer closing motion before the 5.59-second timeout.",
+    "vlm_temporal_justification": "No earlier failure event identified; using the final frame.",
+    "vlm_video_fps": 50.3,
+    "vlm_video_num_frames": 282,
+    "vlm_video_duration_seconds": 5.58648111332008,
+    "vlm_usage": {
+      "input_tokens": 1860,
+      "output_tokens": 266,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"timeout_or_insufficient_progress\",\n \"onset_type\": \"timeout\",\n \"failure_onset_seconds\": 5.59,\n \"failure_window_start_seconds\": 5.09,\n \"failure_window_end_seconds\": 5.59,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot arm approached the open drawer and aligned near the handle, but failed to push the drawer closed before the episode time limit expired.\",\n \"failure_reason_summary\": \"insufficient execution speed; drawer not closed before timeout\",\n \"recovery_action\": \"Align the gripper directly with the drawer handle and push the drawer inward until fully closed.\",\n \"recovery_action_summary\": \"align with drawer handle; push drawer closed\",\n \"prevention_action\": \"Move more directly towards the drawer lip and push it shut earlier in the trajectory.\",\n \"prevention_action_summary\": \"approach drawer directly; push shut promptly\",\n \"justification\": \"The robot moved towards the correct object and made no discrete error or collision, but progressed too slowly to execute the drawer closing motion before the 5.59-second timeout.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 281,
+    "failure_sim_step": 281,
+    "failure_policy_step": 46,
+    "failure_chunk_index": 5,
+    "first_post_failure_policy_step": 47,
+    "failure_type": "timeout_or_insufficient_progress",
+    "correction_action": "Align the gripper directly with the drawer handle and push the drawer inward until fully closed."
+  }
+}
\ No newline at end of file
diff --git a/closer_drawer/ep013/labels.npz b/closer_drawer/ep013/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..db48ba5422606a44786153ebb92def0cabe12ba8
--- /dev/null
+++ b/closer_drawer/ep013/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b5ef269d66158f3031ac16d1c421a127475e00fecf778643069762723f2a116c
+size 3699
diff --git a/closer_drawer/ep014/example.md b/closer_drawer/ep014/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..1af80b242d35f00eb98e14c0dc0ab5767b2855fd
--- /dev/null
+++ b/closer_drawer/ep014/example.md
@@ -0,0 +1,21 @@
+# t006__ep014__closer_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Close the drawer
+- **success:** True
+- **control steps:** 374  (7.4s at 51 Hz)
+- **policy inferences:** 64  (replan_steps=6)
+- **video:** `closer_drawer/ep014/rollout.mp4`
+- **features:** `64` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching towards open drawer |
+| 3.0–6.0 | 152–303 | 25–50 | pushing drawer closed |
+| 6.0–7.4 | 304–373 | 50–62 | retracting arm |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/closer_drawer/ep014/labels.json b/closer_drawer/ep014/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..2590244ddea1b9f9313fb42c7c93521d8a267f14
--- /dev/null
+++ b/closer_drawer/ep014/labels.json
@@ -0,0 +1,68 @@
+{
+  "rollout_id": "t006__ep014__closer_drawer",
+  "scene_variant": "",
+  "task_id": 6,
+  "task_file": "",
+  "instruction": "Close the drawer",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:22:25+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9980276134122286,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching towards open drawer",
+      "description": "reaching towards open drawer"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9980276134122286,
+      "t_end_sec": 5.996055226824457,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "pushing drawer closed",
+      "description": "pushing drawer closed"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.996055226824457,
+      "t_end_sec": 7.376725838264299,
+      "control_step_start": 304,
+      "control_step_end": 373,
+      "policy_step_start": 50,
+      "policy_step_end": 62,
+      "phrase": "retracting arm",
+      "description": "retracting arm"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/closer_drawer/ep014/labels.npz b/closer_drawer/ep014/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..ee4c5629d5a3bde8322a8515eebf84276d468a62
--- /dev/null
+++ b/closer_drawer/ep014/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a1c6146a2cc6fc703436866326c4e636e23acba28a3d6d77f4488603bdc1196d
+size 3080
diff --git a/closer_drawer/ep015/example.md b/closer_drawer/ep015/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..d155defc055737d7e4640dad38846e46746d1b37
--- /dev/null
+++ b/closer_drawer/ep015/example.md
@@ -0,0 +1,20 @@
+# t006__ep015__closer_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Close the drawer
+- **success:** True
+- **control steps:** 307  (6.0s at 51 Hz)
+- **policy inferences:** 53  (replan_steps=6)
+- **video:** `closer_drawer/ep015/rollout.mp4`
+- **features:** `53` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–153 | 0–25 | reaching for the drawer |
+| 3.0–6.0 | 154–306 | 25–51 | pushing the drawer closed |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/closer_drawer/ep015/labels.json b/closer_drawer/ep015/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..5385308831deca0560d88a0be7b23b225d337256
--- /dev/null
+++ b/closer_drawer/ep015/labels.json
@@ -0,0 +1,57 @@
+{
+  "rollout_id": "t006__ep015__closer_drawer",
+  "scene_variant": "",
+  "task_id": 6,
+  "task_file": "",
+  "instruction": "Close the drawer",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:22:33+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0078125,
+      "control_step_start": 0,
+      "control_step_end": 153,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for the drawer",
+      "description": "reaching for the drawer"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0078125,
+      "t_end_sec": 5.99609375,
+      "control_step_start": 154,
+      "control_step_end": 306,
+      "policy_step_start": 25,
+      "policy_step_end": 51,
+      "phrase": "pushing the drawer closed",
+      "description": "pushing the drawer closed"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/closer_drawer/ep015/labels.npz b/closer_drawer/ep015/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..063a308c1c468cba13388a24a30516566edc900f
--- /dev/null
+++ b/closer_drawer/ep015/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f7e72a56cc3cf10a9c38a6208dae2bf4d702080eb5d29244a9c88bdf9051ea8c
+size 3031
diff --git a/closer_drawer/ep016/example.md b/closer_drawer/ep016/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..3dc62175f245c099dc118fdad1de8b249368437b
--- /dev/null
+++ b/closer_drawer/ep016/example.md
@@ -0,0 +1,20 @@
+# t006__ep016__closer_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Close the drawer
+- **success:** True
+- **control steps:** 285  (5.6s at 50 Hz)
+- **policy inferences:** 49  (replan_steps=6)
+- **video:** `closer_drawer/ep016/rollout.mp4`
+- **features:** `49` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching for the drawer |
+| 3.0–5.6 | 152–284 | 25–47 | pushing the drawer closed |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/closer_drawer/ep016/labels.json b/closer_drawer/ep016/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..4715025f98d496e6f7f2fcd6812b019b4cb133e6
--- /dev/null
+++ b/closer_drawer/ep016/labels.json
@@ -0,0 +1,57 @@
+{
+  "rollout_id": "t006__ep016__closer_drawer",
+  "scene_variant": "",
+  "task_id": 6,
+  "task_file": "",
+  "instruction": "Close the drawer",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:22:42+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00990099009901,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for the drawer",
+      "description": "reaching for the drawer"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00990099009901,
+      "t_end_sec": 5.643564356435643,
+      "control_step_start": 152,
+      "control_step_end": 284,
+      "policy_step_start": 25,
+      "policy_step_end": 47,
+      "phrase": "pushing the drawer closed",
+      "description": "pushing the drawer closed"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/closer_drawer/ep016/labels.npz b/closer_drawer/ep016/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..ed3d25edabb0a88c5835f350f6d7e9e253da2a52
--- /dev/null
+++ b/closer_drawer/ep016/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:43da583ff5b1d242089742eb8492b48f02d4bb404268a332d5d29e35c9cbbcc4
+size 3033
diff --git a/closer_drawer/ep017/example.md b/closer_drawer/ep017/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..a5bbd18f6d6d7661198e4131b4d05890c44fd4ca
--- /dev/null
+++ b/closer_drawer/ep017/example.md
@@ -0,0 +1,21 @@
+# t006__ep017__closer_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Close the drawer
+- **success:** True
+- **control steps:** 346  (6.9s at 50 Hz)
+- **policy inferences:** 60  (replan_steps=5)
+- **video:** `closer_drawer/ep017/rollout.mp4`
+- **features:** `60` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–149 | 0–29 | reaching for the drawer |
+| 3.0–6.0 | 150–299 | 30–59 | pushing the drawer closed |
+| 6.0–6.9 | 300–345 | 60–69 | retracting the robotic arm |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/closer_drawer/ep017/labels.json b/closer_drawer/ep017/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..4f59216e2eca06db77cdd40eeaca981856a9692c
--- /dev/null
+++ b/closer_drawer/ep017/labels.json
@@ -0,0 +1,68 @@
+{
+  "rollout_id": "t006__ep017__closer_drawer",
+  "scene_variant": "",
+  "task_id": 6,
+  "task_file": "",
+  "instruction": "Close the drawer",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:22:48+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9940119760479043,
+      "control_step_start": 0,
+      "control_step_end": 149,
+      "policy_step_start": 0,
+      "policy_step_end": 29,
+      "phrase": "reaching for the drawer",
+      "description": "reaching for the drawer"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9940119760479043,
+      "t_end_sec": 5.9880239520958085,
+      "control_step_start": 150,
+      "control_step_end": 299,
+      "policy_step_start": 30,
+      "policy_step_end": 59,
+      "phrase": "pushing the drawer closed",
+      "description": "pushing the drawer closed"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.9880239520958085,
+      "t_end_sec": 6.9061876247504985,
+      "control_step_start": 300,
+      "control_step_end": 345,
+      "policy_step_start": 60,
+      "policy_step_end": 69,
+      "phrase": "retracting the robotic arm",
+      "description": "retracting the robotic arm"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/closer_drawer/ep017/labels.npz b/closer_drawer/ep017/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..7c4c794d1cde7185f965f6e90fb8249706c946d7
--- /dev/null
+++ b/closer_drawer/ep017/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:83a9027859cc23c498a7b635a87facbf0e4cc47245f03d59c19605789157d365
+size 3084
diff --git a/closer_drawer/ep018/example.md b/closer_drawer/ep018/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..ad4bd82183e67c008fdf05c246d8c8634cefa03b
--- /dev/null
+++ b/closer_drawer/ep018/example.md
@@ -0,0 +1,32 @@
+# t006__ep018__closer_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Close the drawer
+- **success:** False
+- **control steps:** 247  (4.9s at 51 Hz)
+- **policy inferences:** 42  (replan_steps=6)
+- **video:** `closer_drawer/ep018/rollout.mp4`
+- **features:** `42` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | approaching open drawer |
+| 3.0–4.9 | 152–246 | 25–41 | closing gripper in air |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 3.116370808678501 s (00:03.1)
+- **frame / control_step:** 158 (policy_step=26, chunk_index=2)
+- **coarse onset:** 2.8 s (refined=True)
+- **mode:** open_close_failure
+- **reason:** The robot descended to the side of the open drawer lip and closed its gripper on empty air rather than contacting or grasping the drawer to push it shut.
+  Summary: missed drawer lip; closed gripper on air
+- **recovery:** Open the gripper, align it with the open drawer lip or handle, and push the drawer straight back into its housing until fully closed.
+  Summary: re-align with drawer lip; push drawer closed
+- **prevention:** Align the end-effector with the exposed drawer handle before descending to ensure firm contact and push the drawer fully shut.
+  Summary: align gripper with drawer handle; push drawer shut
+- **justification:** Around 2.8 seconds into the video, the robot closes its gripper in empty space to the right of the open drawer, failing to contact the drawer at all.
diff --git a/closer_drawer/ep018/labels.json b/closer_drawer/ep018/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..9ce24c1f8323b227a8b2e8a23601c0beca24f281
--- /dev/null
+++ b/closer_drawer/ep018/labels.json
@@ -0,0 +1,87 @@
+{
+  "rollout_id": "t006__ep018__closer_drawer",
+  "scene_variant": "",
+  "task_id": 6,
+  "task_file": "",
+  "instruction": "Close the drawer",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:23:23+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9980276134122286,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "approaching open drawer",
+      "description": "approaching open drawer"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9980276134122286,
+      "t_end_sec": 4.871794871794871,
+      "control_step_start": 152,
+      "control_step_end": 246,
+      "policy_step_start": 25,
+      "policy_step_end": 41,
+      "phrase": "closing gripper in air",
+      "description": "closing gripper in air"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "open_close_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 3.116370808678501,
+    "vlm_failure_onset_timestamp": "00:03.1",
+    "vlm_failure_onset_frame": 158,
+    "vlm_failure_onset_step": 158,
+    "vlm_coarse_failure_onset_seconds": 2.8,
+    "vlm_failure_window_start_seconds": 2.3,
+    "vlm_failure_window_end_seconds": 3.3,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot descended to the side of the open drawer lip and closed its gripper on empty air rather than contacting or grasping the drawer to push it shut.",
+    "vlm_failure_reason_summary": "missed drawer lip; closed gripper on air",
+    "vlm_recovery_action": "Open the gripper, align it with the open drawer lip or handle, and push the drawer straight back into its housing until fully closed.",
+    "vlm_recovery_action_summary": "re-align with drawer lip; push drawer closed",
+    "vlm_prevention_action": "Align the end-effector with the exposed drawer handle before descending to ensure firm contact and push the drawer fully shut.",
+    "vlm_prevention_action_summary": "align gripper with drawer handle; push drawer shut",
+    "vlm_justification": "Around 2.8 seconds into the video, the robot closes its gripper in empty space to the right of the open drawer, failing to contact the drawer at all.",
+    "vlm_temporal_justification": "At second 02:38 (frame 158), the robot arm completes its descent to the right of the open drawer lip and begins closing its gripper on empty air, missing the drawer handle/lip entirely.",
+    "vlm_video_fps": 50.7,
+    "vlm_video_num_frames": 247,
+    "vlm_video_duration_seconds": 4.85207100591716,
+    "vlm_usage": {
+      "input_tokens": 17718,
+      "output_tokens": 365,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"open_close_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 2.8,\n \"failure_window_start_seconds\": 2.3,\n \"failure_window_end_seconds\": 3.3,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot descended to the side of the open drawer lip and closed its gripper on empty air rather than contacting or grasping the drawer to push it shut.\",\n \"failure_reason_summary\": \"missed drawer lip; closed gripper on air\",\n \"recovery_action\": \"Open the gripper, align it with the open drawer lip or handle, and push the drawer straight back into its housing until fully closed.\",\n \"recovery_action_summary\": \"re-align with drawer lip; push drawer closed\",\n \"prevention_action\": \"Align the end-effector with the exposed drawer handle before descending to ensure firm contact and push the drawer fully shut.\",\n \"prevention_action_summary\": \"align gripper with drawer handle; push drawer shut\",\n \"justification\": \"Around 2.8 seconds into the video, the robot closes its gripper in empty space to the right of the open drawer, failing to contact the drawer at all.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 158,
+    "failure_sim_step": 158,
+    "failure_policy_step": 26,
+    "failure_chunk_index": 2,
+    "first_post_failure_policy_step": 27,
+    "failure_type": "open_close_failure",
+    "correction_action": "Open the gripper, align it with the open drawer lip or handle, and push the drawer straight back into its housing until fully closed."
+  }
+}
\ No newline at end of file
diff --git a/closer_drawer/ep018/labels.npz b/closer_drawer/ep018/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..a164fe31589f77bffd89306b625706612a790237
--- /dev/null
+++ b/closer_drawer/ep018/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3d5ca47c0ed6556b7c51f5adf34a53a859a4daf24b852ed5fb622133dee0344a
+size 3721
diff --git a/closer_drawer/ep019/example.md b/closer_drawer/ep019/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..421827402360483bfa578470c55cfbed198ea6a4
--- /dev/null
+++ b/closer_drawer/ep019/example.md
@@ -0,0 +1,32 @@
+# t006__ep019__closer_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Close the drawer
+- **success:** False
+- **control steps:** 226  (4.5s at 50 Hz)
+- **policy inferences:** 40  (replan_steps=6)
+- **video:** `closer_drawer/ep019/rollout.mp4`
+- **features:** `40` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | lowering arm toward table surface |
+| 3.0–4.5 | 152–225 | 25–37 | hovering above table surface |
+
+## Failure
+
+- **yes**
+- **onset type:** operator_intervention
+- **onset:** 1.0693069306930694 s (00:01.1)
+- **frame / control_step:** 54 (policy_step=9, chunk_index=0)
+- **coarse onset:** 1.8 s (refined=True)
+- **mode:** stuck_or_no_progress
+- **reason:** The robot arm descends vertically into open space near the center of the table rather than approaching the drawer on the left, making no visible progress toward closing it.
+  Summary: unproductive descent in open space; no progress toward drawer
+- **recovery:** Reposition the arm toward the extended drawer handle and push the drawer fully back into its housing.
+  Summary: align with drawer handle; push drawer closed
+- **prevention:** Direct the end-effector trajectory toward the drawer handle immediately from the start pose rather than moving straight down.
+  Summary: target drawer handle directly; execute closing motion
+- **justification:** From early in the episode, the robot lowers its end-effector straight down in the center of the workspace, completely ignoring the drawer on the left and making no progress toward completing the task.
diff --git a/closer_drawer/ep019/labels.json b/closer_drawer/ep019/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..cccfd74b574112c541b2087c6f1602a893c7ef5c
--- /dev/null
+++ b/closer_drawer/ep019/labels.json
@@ -0,0 +1,87 @@
+{
+  "rollout_id": "t006__ep019__closer_drawer",
+  "scene_variant": "",
+  "task_id": 6,
+  "task_file": "",
+  "instruction": "Close the drawer",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:23:57+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00990099009901,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "lowering arm toward table surface",
+      "description": "lowering arm toward table surface"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00990099009901,
+      "t_end_sec": 4.475247524752476,
+      "control_step_start": 152,
+      "control_step_end": 225,
+      "policy_step_start": 25,
+      "policy_step_end": 37,
+      "phrase": "hovering above table surface",
+      "description": "hovering above table surface"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "stuck_or_no_progress",
+    "vlm_failure_onset_type": "operator_intervention",
+    "vlm_failure_onset_seconds": 1.0693069306930694,
+    "vlm_failure_onset_timestamp": "00:01.1",
+    "vlm_failure_onset_frame": 54,
+    "vlm_failure_onset_step": 54,
+    "vlm_coarse_failure_onset_seconds": 1.8,
+    "vlm_failure_window_start_seconds": 1.3,
+    "vlm_failure_window_end_seconds": 2.3,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot arm descends vertically into open space near the center of the table rather than approaching the drawer on the left, making no visible progress toward closing it.",
+    "vlm_failure_reason_summary": "unproductive descent in open space; no progress toward drawer",
+    "vlm_recovery_action": "Reposition the arm toward the extended drawer handle and push the drawer fully back into its housing.",
+    "vlm_recovery_action_summary": "align with drawer handle; push drawer closed",
+    "vlm_prevention_action": "Direct the end-effector trajectory toward the drawer handle immediately from the start pose rather than moving straight down.",
+    "vlm_prevention_action_summary": "target drawer handle directly; execute closing motion",
+    "vlm_justification": "From early in the episode, the robot lowers its end-effector straight down in the center of the workspace, completely ignoring the drawer on the left and making no progress toward completing the task.",
+    "vlm_temporal_justification": "At clip second 54 (original rollout frame 54), the robot arm begins moving downward from its home position towards the empty space near the center of the table, rather than approaching the blue drawer on the left.",
+    "vlm_video_fps": 50.5,
+    "vlm_video_num_frames": 226,
+    "vlm_video_duration_seconds": 4.455445544554456,
+    "vlm_usage": {
+      "input_tokens": 16334,
+      "output_tokens": 356,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"stuck_or_no_progress\",\n \"onset_type\": \"operator_intervention\",\n \"failure_onset_seconds\": 1.8,\n \"failure_window_start_seconds\": 1.3,\n \"failure_window_end_seconds\": 2.3,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot arm descends vertically into open space near the center of the table rather than approaching the drawer on the left, making no visible progress toward closing it.\",\n \"failure_reason_summary\": \"unproductive descent in open space; no progress toward drawer\",\n \"recovery_action\": \"Reposition the arm toward the extended drawer handle and push the drawer fully back into its housing.\",\n \"recovery_action_summary\": \"align with drawer handle; push drawer closed\",\n \"prevention_action\": \"Direct the end-effector trajectory toward the drawer handle immediately from the start pose rather than moving straight down.\",\n \"prevention_action_summary\": \"target drawer handle directly; execute closing motion\",\n \"justification\": \"From early in the episode, the robot lowers its end-effector straight down in the center of the workspace, completely ignoring the drawer on the left and making no progress toward completing the task.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 54,
+    "failure_sim_step": 54,
+    "failure_policy_step": 9,
+    "failure_chunk_index": 0,
+    "first_post_failure_policy_step": 10,
+    "failure_type": "stuck_or_no_progress",
+    "correction_action": "Reposition the arm toward the extended drawer handle and push the drawer fully back into its housing."
+  }
+}
\ No newline at end of file
diff --git a/closer_drawer/ep019/labels.npz b/closer_drawer/ep019/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..d1bfe71fd521d29030f734f3b5dc6da5fdf18a7f
--- /dev/null
+++ b/closer_drawer/ep019/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:324d1f60f29b84dca756dc26c32f5aa2ceec659bc88498d2845c4f184f38faa2
+size 3759
diff --git a/closer_drawer/ep020/example.md b/closer_drawer/ep020/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..74ec32f1e3896b88345ba1c8588aaf42f6a4a31d
--- /dev/null
+++ b/closer_drawer/ep020/example.md
@@ -0,0 +1,21 @@
+# t006__ep020__closer_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Close the drawer
+- **success:** True
+- **control steps:** 305  (6.1s at 50 Hz)
+- **policy inferences:** 52  (replan_steps=6)
+- **video:** `closer_drawer/ep020/rollout.mp4`
+- **features:** `52` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | reaching towards open drawer |
+| 3.0–6.0 | 151–301 | 25–50 | pushing the drawer shut |
+| 6.0–6.1 | 302–304 | 50–50 | idle near closed drawer |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/closer_drawer/ep020/labels.json b/closer_drawer/ep020/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..fa67122905b10a788d29ec7753b2703f5e147a4e
--- /dev/null
+++ b/closer_drawer/ep020/labels.json
@@ -0,0 +1,68 @@
+{
+  "rollout_id": "t006__ep020__closer_drawer",
+  "scene_variant": "",
+  "task_id": 6,
+  "task_file": "",
+  "instruction": "Close the drawer",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:24:04+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.996031746031746,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching towards open drawer",
+      "description": "reaching towards open drawer"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.996031746031746,
+      "t_end_sec": 5.992063492063492,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "pushing the drawer shut",
+      "description": "pushing the drawer shut"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.992063492063492,
+      "t_end_sec": 6.051587301587301,
+      "control_step_start": 302,
+      "control_step_end": 304,
+      "policy_step_start": 50,
+      "policy_step_end": 50,
+      "phrase": "idle near closed drawer",
+      "description": "idle near closed drawer"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/closer_drawer/ep020/labels.npz b/closer_drawer/ep020/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..3addd2ccde30812fe64f8da0bb7c933737fa2459
--- /dev/null
+++ b/closer_drawer/ep020/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:aab54120b80aeae4f3505606436ba1f89ec6f8c60c9e68a8cddaf3e28c3c1ce8
+size 3079
diff --git a/closer_drawer/ep021/example.md b/closer_drawer/ep021/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..bb5a3422ba39219b87ce89c5b86d528fc622817c
--- /dev/null
+++ b/closer_drawer/ep021/example.md
@@ -0,0 +1,33 @@
+# t006__ep021__closer_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Close the drawer
+- **success:** False
+- **control steps:** 343  (6.8s at 51 Hz)
+- **policy inferences:** 59  (replan_steps=6)
+- **video:** `closer_drawer/ep021/rollout.mp4`
+- **features:** `59` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | approaching open drawer |
+| 3.0–6.0 | 152–303 | 25–50 | pushing against drawer edge |
+| 6.0–6.8 | 304–342 | 50–57 | hovering near drawer |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 4.703557312252964 s (00:04.7)
+- **frame / control_step:** 238 (policy_step=39, chunk_index=4)
+- **coarse onset:** 4.5 s (refined=True)
+- **mode:** open_close_failure
+- **reason:** The robot gripper contacted the edge of the drawer at an awkward angle and slipped off without applying enough force to slide the drawer shut.
+  Summary: misaligned contact with drawer; failed to push drawer shut
+- **recovery:** Re-align the gripper with the open drawer lip and push forward horizontally until the drawer is fully closed.
+  Summary: re-align gripper; push drawer closed
+- **prevention:** Align the end-effector directly against the front lip of the drawer before executing a forward pushing motion.
+  Summary: align gripper with drawer handle; push firmly shut
+- **justification:** Around 4.5 seconds, the robot arm makes contact with the open drawer, but the gripper misaligns and fails to slide the drawer into its slot, leaving the drawer open at the end of the episode.
diff --git a/closer_drawer/ep021/labels.json b/closer_drawer/ep021/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..c6c8022ee831feea3536a0b5cb7b2926398128b1
--- /dev/null
+++ b/closer_drawer/ep021/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t006__ep021__closer_drawer",
+  "scene_variant": "",
+  "task_id": 6,
+  "task_file": "",
+  "instruction": "Close the drawer",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:24:35+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0039525691699605,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "approaching open drawer",
+      "description": "approaching open drawer"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0039525691699605,
+      "t_end_sec": 6.007905138339921,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "pushing against drawer edge",
+      "description": "pushing against drawer edge"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.007905138339921,
+      "t_end_sec": 6.778656126482213,
+      "control_step_start": 304,
+      "control_step_end": 342,
+      "policy_step_start": 50,
+      "policy_step_end": 57,
+      "phrase": "hovering near drawer",
+      "description": "hovering near drawer"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "open_close_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 4.703557312252964,
+    "vlm_failure_onset_timestamp": "00:04.7",
+    "vlm_failure_onset_frame": 238,
+    "vlm_failure_onset_step": 238,
+    "vlm_coarse_failure_onset_seconds": 4.5,
+    "vlm_failure_window_start_seconds": 4.0,
+    "vlm_failure_window_end_seconds": 5.0,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot gripper contacted the edge of the drawer at an awkward angle and slipped off without applying enough force to slide the drawer shut.",
+    "vlm_failure_reason_summary": "misaligned contact with drawer; failed to push drawer shut",
+    "vlm_recovery_action": "Re-align the gripper with the open drawer lip and push forward horizontally until the drawer is fully closed.",
+    "vlm_recovery_action_summary": "re-align gripper; push drawer closed",
+    "vlm_prevention_action": "Align the end-effector directly against the front lip of the drawer before executing a forward pushing motion.",
+    "vlm_prevention_action_summary": "align gripper with drawer handle; push firmly shut",
+    "vlm_justification": "Around 4.5 seconds, the robot arm makes contact with the open drawer, but the gripper misaligns and fails to slide the drawer into its slot, leaving the drawer open at the end of the episode.",
+    "vlm_temporal_justification": "At 02:42 (second 162), the robot's gripper makes contact with the edge of the extended drawer at an awkward angle, causing it to slip past the lip without exerting horizontal force to push the drawer shut.",
+    "vlm_video_fps": 50.6,
+    "vlm_video_num_frames": 343,
+    "vlm_video_duration_seconds": 6.758893280632411,
+    "vlm_usage": {
+      "input_tokens": 19103,
+      "output_tokens": 363,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"open_close_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 4.5,\n \"failure_window_start_seconds\": 4.0,\n \"failure_window_end_seconds\": 5.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot gripper contacted the edge of the drawer at an awkward angle and slipped off without applying enough force to slide the drawer shut.\",\n \"failure_reason_summary\": \"misaligned contact with drawer; failed to push drawer shut\",\n \"recovery_action\": \"Re-align the gripper with the open drawer lip and push forward horizontally until the drawer is fully closed.\",\n \"recovery_action_summary\": \"re-align gripper; push drawer closed\",\n \"prevention_action\": \"Align the end-effector directly against the front lip of the drawer before executing a forward pushing motion.\",\n \"prevention_action_summary\": \"align gripper with drawer handle; push firmly shut\",\n \"justification\": \"Around 4.5 seconds, the robot arm makes contact with the open drawer, but the gripper misaligns and fails to slide the drawer into its slot, leaving the drawer open at the end of the episode.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 238,
+    "failure_sim_step": 238,
+    "failure_policy_step": 39,
+    "failure_chunk_index": 4,
+    "first_post_failure_policy_step": 40,
+    "failure_type": "open_close_failure",
+    "correction_action": "Re-align the gripper with the open drawer lip and push forward horizontally until the drawer is fully closed."
+  }
+}
\ No newline at end of file
diff --git a/closer_drawer/ep021/labels.npz b/closer_drawer/ep021/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..4b05171b71c48c21936e766d3eea625fe4178f74
--- /dev/null
+++ b/closer_drawer/ep021/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d9a71cae3b22b77dae8b103564973d934ae6f3bcf2c538f47be5a7a689f30266
+size 3751
diff --git a/closer_drawer/ep022/example.md b/closer_drawer/ep022/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..13e855df8a12d33a282c58b9d875fcb9d0cc1c02
--- /dev/null
+++ b/closer_drawer/ep022/example.md
@@ -0,0 +1,34 @@
+# t006__ep022__closer_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Close the drawer
+- **success:** False
+- **control steps:** 515  (10.2s at 50 Hz)
+- **policy inferences:** 88  (replan_steps=6)
+- **video:** `closer_drawer/ep022/rollout.mp4`
+- **features:** `88` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | approaching the cyan drawer |
+| 3.0–6.0 | 151–301 | 25–50 | lowering gripper to drawer edge |
+| 6.0–9.0 | 302–452 | 50–75 | pressing drawer edge unsuccessfully |
+| 9.0–10.2 | 453–514 | 75–85 | holding position at drawer |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 5.436507936507937 s (00:05.4)
+- **frame / control_step:** 274 (policy_step=45, chunk_index=4)
+- **coarse onset:** 5.1 s (refined=True)
+- **mode:** open_close_failure
+- **reason:** The robot reaches the light blue drawer but fails to securely engage or push the drawer handle, leaving the drawer open.
+  Summary: failed to push drawer; drawer remains open
+- **recovery:** Re-align the gripper with the drawer rim or handle and apply forward pushing force to slide the drawer fully closed.
+  Summary: re-align gripper; push drawer closed
+- **prevention:** Approach the drawer handle at the correct height and orientation to reliably exert closing force on the drawer.
+  Summary: align with drawer handle; push smoothly
+- **justification:** The robot moves to the cyan drawer and touches the rim around 5.1 seconds, but fails to apply effective force or grab the drawer to slide it shut. As a result, the drawer stays open for the remainder of the trajectory.
diff --git a/closer_drawer/ep022/labels.json b/closer_drawer/ep022/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..e642dce0bc781117cc703a71c973d27021c1344a
--- /dev/null
+++ b/closer_drawer/ep022/labels.json
@@ -0,0 +1,109 @@
+{
+  "rollout_id": "t006__ep022__closer_drawer",
+  "scene_variant": "",
+  "task_id": 6,
+  "task_file": "",
+  "instruction": "Close the drawer",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:25:08+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.996031746031746,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "approaching the cyan drawer",
+      "description": "approaching the cyan drawer"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.996031746031746,
+      "t_end_sec": 5.992063492063492,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "lowering gripper to drawer edge",
+      "description": "lowering gripper to drawer edge"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.992063492063492,
+      "t_end_sec": 8.988095238095239,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "pressing drawer edge unsuccessfully",
+      "description": "pressing drawer edge unsuccessfully"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.988095238095239,
+      "t_end_sec": 10.218253968253968,
+      "control_step_start": 453,
+      "control_step_end": 514,
+      "policy_step_start": 75,
+      "policy_step_end": 85,
+      "phrase": "holding position at drawer",
+      "description": "holding position at drawer"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "open_close_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 5.436507936507937,
+    "vlm_failure_onset_timestamp": "00:05.4",
+    "vlm_failure_onset_frame": 274,
+    "vlm_failure_onset_step": 274,
+    "vlm_coarse_failure_onset_seconds": 5.1,
+    "vlm_failure_window_start_seconds": 4.5,
+    "vlm_failure_window_end_seconds": 7.0,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot reaches the light blue drawer but fails to securely engage or push the drawer handle, leaving the drawer open.",
+    "vlm_failure_reason_summary": "failed to push drawer; drawer remains open",
+    "vlm_recovery_action": "Re-align the gripper with the drawer rim or handle and apply forward pushing force to slide the drawer fully closed.",
+    "vlm_recovery_action_summary": "re-align gripper; push drawer closed",
+    "vlm_prevention_action": "Approach the drawer handle at the correct height and orientation to reliably exert closing force on the drawer.",
+    "vlm_prevention_action_summary": "align with drawer handle; push smoothly",
+    "vlm_justification": "The robot moves to the cyan drawer and touches the rim around 5.1 seconds, but fails to apply effective force or grab the drawer to slide it shut. As a result, the drawer stays open for the remainder of the trajectory.",
+    "vlm_temporal_justification": "At 02:48 into the clip, the robot gripper closes completely without grasping or engaging the edge of the drawer, missing the drawer wall and failing to perform the closing action.",
+    "vlm_video_fps": 50.4,
+    "vlm_video_num_frames": 515,
+    "vlm_video_duration_seconds": 10.1984126984127,
+    "vlm_usage": {
+      "input_tokens": 21562,
+      "output_tokens": 349,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"open_close_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 5.1,\n \"failure_window_start_seconds\": 4.5,\n \"failure_window_end_seconds\": 7.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot reaches the light blue drawer but fails to securely engage or push the drawer handle, leaving the drawer open.\",\n \"failure_reason_summary\": \"failed to push drawer; drawer remains open\",\n \"recovery_action\": \"Re-align the gripper with the drawer rim or handle and apply forward pushing force to slide the drawer fully closed.\",\n \"recovery_action_summary\": \"re-align gripper; push drawer closed\",\n \"prevention_action\": \"Approach the drawer handle at the correct height and orientation to reliably exert closing force on the drawer.\",\n \"prevention_action_summary\": \"align with drawer handle; push smoothly\",\n \"justification\": \"The robot moves to the cyan drawer and touches the rim around 5.1 seconds, but fails to apply effective force or grab the drawer to slide it shut. As a result, the drawer stays open for the remainder of the trajectory.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 274,
+    "failure_sim_step": 274,
+    "failure_policy_step": 45,
+    "failure_chunk_index": 4,
+    "first_post_failure_policy_step": 46,
+    "failure_type": "open_close_failure",
+    "correction_action": "Re-align the gripper with the drawer rim or handle and apply forward pushing force to slide the drawer fully closed."
+  }
+}
\ No newline at end of file
diff --git a/closer_drawer/ep022/labels.npz b/closer_drawer/ep022/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..1bd9e3a327a5a89fb96c14d1991a1c9b8731348e
--- /dev/null
+++ b/closer_drawer/ep022/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:11ccd61e5fabc568f42ccc02e8f76d8c49a5a354c4d31e69bc1dc4cd18d484b9
+size 3779
diff --git a/closer_drawer/ep023/example.md b/closer_drawer/ep023/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..ac5f11e54f68b7bc13f13b918f58b0eafe6e4a33
--- /dev/null
+++ b/closer_drawer/ep023/example.md
@@ -0,0 +1,24 @@
+# t006__ep023__closer_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Close the drawer
+- **success:** True
+- **control steps:** 851  (16.8s at 51 Hz)
+- **policy inferences:** 145  (replan_steps=6)
+- **video:** `closer_drawer/ep023/rollout.mp4`
+- **features:** `145` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | Reaching towards the box lid |
+| 3.0–6.0 | 152–303 | 25–50 | Approaching the open box flap |
+| 6.0–9.0 | 304–455 | 50–75 | Pushing the box lid closed |
+| 9.0–12.0 | 456–607 | 76–101 | Pressing the lid fully closed |
+| 12.0–15.0 | 608–759 | 101–126 | Retracting arm away from box |
+| 15.0–16.8 | 760–850 | 126–141 | Moving arm to resting position |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/closer_drawer/ep023/labels.json b/closer_drawer/ep023/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..34b7bbbee3c480edcd4cb877fb4656e692175694
--- /dev/null
+++ b/closer_drawer/ep023/labels.json
@@ -0,0 +1,101 @@
+{
+  "rollout_id": "t006__ep023__closer_drawer",
+  "scene_variant": "",
+  "task_id": 6,
+  "task_file": "",
+  "instruction": "Close the drawer",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:25:21+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0039525691699605,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "Reaching towards the box lid",
+      "description": "Reaching towards the box lid"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0039525691699605,
+      "t_end_sec": 6.007905138339921,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "Approaching the open box flap",
+      "description": "Approaching the open box flap"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.007905138339921,
+      "t_end_sec": 9.011857707509881,
+      "control_step_start": 304,
+      "control_step_end": 455,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "Pushing the box lid closed",
+      "description": "Pushing the box lid closed"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.011857707509881,
+      "t_end_sec": 12.015810276679842,
+      "control_step_start": 456,
+      "control_step_end": 607,
+      "policy_step_start": 76,
+      "policy_step_end": 101,
+      "phrase": "Pressing the lid fully closed",
+      "description": "Pressing the lid fully closed"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 12.015810276679842,
+      "t_end_sec": 15.019762845849803,
+      "control_step_start": 608,
+      "control_step_end": 759,
+      "policy_step_start": 101,
+      "policy_step_end": 126,
+      "phrase": "Retracting arm away from box",
+      "description": "Retracting arm away from box"
+    },
+    {
+      "segment_index": 5,
+      "t_start_sec": 15.019762845849803,
+      "t_end_sec": 16.818181818181817,
+      "control_step_start": 760,
+      "control_step_end": 850,
+      "policy_step_start": 126,
+      "policy_step_end": 141,
+      "phrase": "Moving arm to resting position",
+      "description": "Moving arm to resting position"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/closer_drawer/ep023/labels.npz b/closer_drawer/ep023/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..a221e2367807119fe86f5f9227b4fa0a19721148
--- /dev/null
+++ b/closer_drawer/ep023/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7b971424c285e7f05ae94aed993192fbefb2f1a4e0d5d86bbfa0a877df080301
+size 3228
diff --git a/closer_drawer/ep024/example.md b/closer_drawer/ep024/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..3354f30a2fd5b748be6ad292cdd93323b79696b2
--- /dev/null
+++ b/closer_drawer/ep024/example.md
@@ -0,0 +1,34 @@
+# t006__ep024__closer_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Close the drawer
+- **success:** False
+- **control steps:** 586  (11.6s at 50 Hz)
+- **policy inferences:** 101  (replan_steps=6)
+- **video:** `closer_drawer/ep024/rollout.mp4`
+- **features:** `101` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | lowering arm toward drawer |
+| 3.0–6.0 | 151–301 | 25–50 | descending toward drawer lip |
+| 6.0–9.0 | 302–452 | 50–75 | attempting to push drawer closed |
+| 9.0–11.6 | 453–585 | 75–97 | hovering above open drawer |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 7.738095238095238 s (00:07.7)
+- **frame / control_step:** 390 (policy_step=65, chunk_index=0)
+- **coarse onset:** 8.0 s (refined=True)
+- **mode:** open_close_failure
+- **reason:** The robot reaches for the drawer edge to close it, but fails to maintain proper contact or leverage, slipping off without pushing the drawer shut.
+  Summary: slipping off drawer edge; drawer remains open
+- **recovery:** Re-align the end-effector with the drawer edge or handle and apply direct forward force to slide the drawer completely closed.
+  Summary: re-align gripper with drawer; push drawer closed
+- **prevention:** Approach the drawer edge at a stable height and angle, maintaining consistent force forward to ensure the drawer shuts fully.
+  Summary: maintain contact on drawer lip; push completely closed
+- **justification:** Around 8.0 seconds into the rollout, the robot's gripper reaches the open drawer edge but fails to hook or push it effectively. The gripper slips off the lip at around 9.0 seconds and fails to close the drawer before time expires.
diff --git a/closer_drawer/ep024/labels.json b/closer_drawer/ep024/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..debaddb42050d6d01f41081d95abb3f320d90a9b
--- /dev/null
+++ b/closer_drawer/ep024/labels.json
@@ -0,0 +1,109 @@
+{
+  "rollout_id": "t006__ep024__closer_drawer",
+  "scene_variant": "",
+  "task_id": 6,
+  "task_file": "",
+  "instruction": "Close the drawer",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:25:53+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.996031746031746,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "lowering arm toward drawer",
+      "description": "lowering arm toward drawer"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.996031746031746,
+      "t_end_sec": 5.992063492063492,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "descending toward drawer lip",
+      "description": "descending toward drawer lip"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.992063492063492,
+      "t_end_sec": 8.988095238095239,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "attempting to push drawer closed",
+      "description": "attempting to push drawer closed"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.988095238095239,
+      "t_end_sec": 11.626984126984127,
+      "control_step_start": 453,
+      "control_step_end": 585,
+      "policy_step_start": 75,
+      "policy_step_end": 97,
+      "phrase": "hovering above open drawer",
+      "description": "hovering above open drawer"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "open_close_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 7.738095238095238,
+    "vlm_failure_onset_timestamp": "00:07.7",
+    "vlm_failure_onset_frame": 390,
+    "vlm_failure_onset_step": 390,
+    "vlm_coarse_failure_onset_seconds": 8.0,
+    "vlm_failure_window_start_seconds": 8.0,
+    "vlm_failure_window_end_seconds": 11.61,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot reaches for the drawer edge to close it, but fails to maintain proper contact or leverage, slipping off without pushing the drawer shut.",
+    "vlm_failure_reason_summary": "slipping off drawer edge; drawer remains open",
+    "vlm_recovery_action": "Re-align the end-effector with the drawer edge or handle and apply direct forward force to slide the drawer completely closed.",
+    "vlm_recovery_action_summary": "re-align gripper with drawer; push drawer closed",
+    "vlm_prevention_action": "Approach the drawer edge at a stable height and angle, maintaining consistent force forward to ensure the drawer shuts fully.",
+    "vlm_prevention_action_summary": "maintain contact on drawer lip; push completely closed",
+    "vlm_justification": "Around 8.0 seconds into the rollout, the robot's gripper reaches the open drawer edge but fails to hook or push it effectively. The gripper slips off the lip at around 9.0 seconds and fails to close the drawer before time expires.",
+    "vlm_temporal_justification": "At clip second 02:18 (frame 390 in the original sequence, corresponding to second 138 in the clip timer), the robotic gripper reaches the rear rim of the cyan box to close it, but fails to obtain a proper grip or leverage, slipping off as it attempts to execute the push.",
+    "vlm_video_fps": 50.4,
+    "vlm_video_num_frames": 586,
+    "vlm_video_duration_seconds": 11.607142857142858,
+    "vlm_usage": {
+      "input_tokens": 21693,
+      "output_tokens": 399,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"open_close_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 8.0,\n \"failure_window_start_seconds\": 8.0,\n \"failure_window_end_seconds\": 11.61,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot reaches for the drawer edge to close it, but fails to maintain proper contact or leverage, slipping off without pushing the drawer shut.\",\n \"failure_reason_summary\": \"slipping off drawer edge; drawer remains open\",\n \"recovery_action\": \"Re-align the end-effector with the drawer edge or handle and apply direct forward force to slide the drawer completely closed.\",\n \"recovery_action_summary\": \"re-align gripper with drawer; push drawer closed\",\n \"prevention_action\": \"Approach the drawer edge at a stable height and angle, maintaining consistent force forward to ensure the drawer shuts fully.\",\n \"prevention_action_summary\": \"maintain contact on drawer lip; push completely closed\",\n \"justification\": \"Around 8.0 seconds into the rollout, the robot's gripper reaches the open drawer edge but fails to hook or push it effectively. The gripper slips off the lip at around 9.0 seconds and fails to close the drawer before time expires.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 390,
+    "failure_sim_step": 390,
+    "failure_policy_step": 65,
+    "failure_chunk_index": 0,
+    "first_post_failure_policy_step": 66,
+    "failure_type": "open_close_failure",
+    "correction_action": "Re-align the end-effector with the drawer edge or handle and apply direct forward force to slide the drawer completely closed."
+  }
+}
\ No newline at end of file
diff --git a/closer_drawer/ep024/labels.npz b/closer_drawer/ep024/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..c1aed5e63ed3e40a65b3b8bcdc6b949d2df510ec
--- /dev/null
+++ b/closer_drawer/ep024/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8c9d29335d3f5c2da0948326860fa8962cded4db515d7f4d0611f0da453baf19
+size 3820
diff --git a/closer_drawer/ep025/example.md b/closer_drawer/ep025/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..ec0a52b25b2afe24d4d803aba3f14df654b2f72a
--- /dev/null
+++ b/closer_drawer/ep025/example.md
@@ -0,0 +1,35 @@
+# t006__ep025__closer_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Close the drawer
+- **success:** False
+- **control steps:** 617  (12.3s at 50 Hz)
+- **policy inferences:** 106  (replan_steps=6)
+- **video:** `closer_drawer/ep025/rollout.mp4`
+- **features:** `106` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | moving arm toward drawer |
+| 3.0–6.0 | 151–301 | 25–50 | positioning gripper above drawer edge |
+| 6.0–9.0 | 302–452 | 50–75 | closing gripper and missing drawer |
+| 9.0–12.0 | 453–603 | 75–100 | hovering near open drawer |
+| 12.0–12.3 | 604–616 | 100–102 | remaining idle near drawer |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 7.709163346613545 s (00:07.7)
+- **frame / control_step:** 387 (policy_step=64, chunk_index=3)
+- **coarse onset:** 7.8 s (refined=True)
+- **mode:** open_close_failure
+- **reason:** The robot attempted to grasp the edge of the drawer to close it, but missed the contact point and closed its gripper in the air.
+  Summary: missed drawer edge; drawer remains open
+- **recovery:** Re-align the gripper with the open drawer lip, establish a firm grip or push against the edge, and move it until the drawer is fully closed.
+  Summary: re-align gripper; engage drawer edge; close drawer
+- **prevention:** Accurately align the end-effector with the drawer edge or handle before closing the gripper to guarantee contact.
+  Summary: align with edge; grasp handle; push drawer shut
+- **justification:** At around 7.8 seconds, the end-effector closes near the top-right rim of the drawer without making contact or grasping the lip, failing to move the drawer and leaving it open.
diff --git a/closer_drawer/ep025/labels.json b/closer_drawer/ep025/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..e3795c7e58391aefe2e2bbaf47f16be8ed15188b
--- /dev/null
+++ b/closer_drawer/ep025/labels.json
@@ -0,0 +1,120 @@
+{
+  "rollout_id": "t006__ep025__closer_drawer",
+  "scene_variant": "",
+  "task_id": 6,
+  "task_file": "",
+  "instruction": "Close the drawer",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:26:29+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00796812749004,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "moving arm toward drawer",
+      "description": "moving arm toward drawer"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00796812749004,
+      "t_end_sec": 6.01593625498008,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "positioning gripper above drawer edge",
+      "description": "positioning gripper above drawer edge"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01593625498008,
+      "t_end_sec": 9.02390438247012,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "closing gripper and missing drawer",
+      "description": "closing gripper and missing drawer"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.02390438247012,
+      "t_end_sec": 12.03187250996016,
+      "control_step_start": 453,
+      "control_step_end": 603,
+      "policy_step_start": 75,
+      "policy_step_end": 100,
+      "phrase": "hovering near open drawer",
+      "description": "hovering near open drawer"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 12.03187250996016,
+      "t_end_sec": 12.290836653386453,
+      "control_step_start": 604,
+      "control_step_end": 616,
+      "policy_step_start": 100,
+      "policy_step_end": 102,
+      "phrase": "remaining idle near drawer",
+      "description": "remaining idle near drawer"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "open_close_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 7.709163346613545,
+    "vlm_failure_onset_timestamp": "00:07.7",
+    "vlm_failure_onset_frame": 387,
+    "vlm_failure_onset_step": 387,
+    "vlm_coarse_failure_onset_seconds": 7.8,
+    "vlm_failure_window_start_seconds": 6.8,
+    "vlm_failure_window_end_seconds": 8.8,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the edge of the drawer to close it, but missed the contact point and closed its gripper in the air.",
+    "vlm_failure_reason_summary": "missed drawer edge; drawer remains open",
+    "vlm_recovery_action": "Re-align the gripper with the open drawer lip, establish a firm grip or push against the edge, and move it until the drawer is fully closed.",
+    "vlm_recovery_action_summary": "re-align gripper; engage drawer edge; close drawer",
+    "vlm_prevention_action": "Accurately align the end-effector with the drawer edge or handle before closing the gripper to guarantee contact.",
+    "vlm_prevention_action_summary": "align with edge; grasp handle; push drawer shut",
+    "vlm_justification": "At around 7.8 seconds, the end-effector closes near the top-right rim of the drawer without making contact or grasping the lip, failing to move the drawer and leaving it open.",
+    "vlm_temporal_justification": "At timestamp 02:26 (146 seconds into the clip), the robot begins closing its gripper fingers in thin air behind the drawer edge, missing the grasping point on the drawer lip.",
+    "vlm_video_fps": 50.2,
+    "vlm_video_num_frames": 617,
+    "vlm_video_duration_seconds": 12.270916334661354,
+    "vlm_usage": {
+      "input_tokens": 21692,
+      "output_tokens": 365,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"open_close_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 7.8,\n \"failure_window_start_seconds\": 6.8,\n \"failure_window_end_seconds\": 8.8,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the edge of the drawer to close it, but missed the contact point and closed its gripper in the air.\",\n \"failure_reason_summary\": \"missed drawer edge; drawer remains open\",\n \"recovery_action\": \"Re-align the gripper with the open drawer lip, establish a firm grip or push against the edge, and move it until the drawer is fully closed.\",\n \"recovery_action_summary\": \"re-align gripper; engage drawer edge; close drawer\",\n \"prevention_action\": \"Accurately align the end-effector with the drawer edge or handle before closing the gripper to guarantee contact.\",\n \"prevention_action_summary\": \"align with edge; grasp handle; push drawer shut\",\n \"justification\": \"At around 7.8 seconds, the end-effector closes near the top-right rim of the drawer without making contact or grasping the lip, failing to move the drawer and leaving it open.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 387,
+    "failure_sim_step": 387,
+    "failure_policy_step": 64,
+    "failure_chunk_index": 3,
+    "first_post_failure_policy_step": 65,
+    "failure_type": "open_close_failure",
+    "correction_action": "Re-align the gripper with the open drawer lip, establish a firm grip or push against the edge, and move it until the drawer is fully closed."
+  }
+}
\ No newline at end of file
diff --git a/closer_drawer/ep025/labels.npz b/closer_drawer/ep025/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..c736b4a690e9dd67db628aca0908dab5f4c8a755
--- /dev/null
+++ b/closer_drawer/ep025/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7b14aecbfafbc6462d40202c9816562571ed5d664a7bb8f8f668d29bace83c07
+size 3851
diff --git a/closer_drawer/ep026/example.md b/closer_drawer/ep026/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..ed31c76605e61f182dbafe4b892c68039321d04d
--- /dev/null
+++ b/closer_drawer/ep026/example.md
@@ -0,0 +1,22 @@
+# t006__ep026__closer_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Close the drawer
+- **success:** True
+- **control steps:** 557  (10.9s at 51 Hz)
+- **policy inferences:** 96  (replan_steps=6)
+- **video:** `closer_drawer/ep026/rollout.mp4`
+- **features:** `96` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–152 | 0–25 | idle above the box |
+| 3.0–6.0 | 153–305 | 25–50 | approaching the open box lid |
+| 6.0–9.0 | 306–458 | 51–76 | pushing lid down to close box |
+| 9.0–10.9 | 459–556 | 76–92 | holding box lid closed |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/closer_drawer/ep026/labels.json b/closer_drawer/ep026/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..6ccfe9b8c7e78d2d61908f21193b2eae12b1006a
--- /dev/null
+++ b/closer_drawer/ep026/labels.json
@@ -0,0 +1,79 @@
+{
+  "rollout_id": "t006__ep026__closer_drawer",
+  "scene_variant": "",
+  "task_id": 6,
+  "task_file": "",
+  "instruction": "Close the drawer",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:26:37+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0,
+      "control_step_start": 0,
+      "control_step_end": 152,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "idle above the box",
+      "description": "idle above the box"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0,
+      "t_end_sec": 6.0,
+      "control_step_start": 153,
+      "control_step_end": 305,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "approaching the open box lid",
+      "description": "approaching the open box lid"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.0,
+      "t_end_sec": 9.0,
+      "control_step_start": 306,
+      "control_step_end": 458,
+      "policy_step_start": 51,
+      "policy_step_end": 76,
+      "phrase": "pushing lid down to close box",
+      "description": "pushing lid down to close box"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.0,
+      "t_end_sec": 10.92156862745098,
+      "control_step_start": 459,
+      "control_step_end": 556,
+      "policy_step_start": 76,
+      "policy_step_end": 92,
+      "phrase": "holding box lid closed",
+      "description": "holding box lid closed"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/closer_drawer/ep026/labels.npz b/closer_drawer/ep026/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..f83821c8467adf945564bbd06f6141e95d60c27d
--- /dev/null
+++ b/closer_drawer/ep026/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8e503adf9b5bb9f0304511f6ee09f4700f08c7ce8277ac293d4a8b1cb0fc05da
+size 3120
diff --git a/closer_drawer/ep027/example.md b/closer_drawer/ep027/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..346d8cbb9504f7c2e2fc676d5e86ef0657a8a923
--- /dev/null
+++ b/closer_drawer/ep027/example.md
@@ -0,0 +1,34 @@
+# t006__ep027__closer_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Close the drawer
+- **success:** False
+- **control steps:** 590  (11.8s at 50 Hz)
+- **policy inferences:** 101  (replan_steps=6)
+- **video:** `closer_drawer/ep027/rollout.mp4`
+- **features:** `101` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–149 | 0–24 | lowering arm toward drawer |
+| 3.0–6.0 | 150–299 | 25–49 | contacting upper drawer edge |
+| 6.0–9.0 | 300–449 | 50–74 | holding position at drawer edge |
+| 9.0–11.8 | 450–589 | 75–98 | remaining stationary at drawer edge |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 4.2 s (00:04.2)
+- **frame / control_step:** 210 (policy_step=35, chunk_index=0)
+- **coarse onset:** 4.5 s (refined=True)
+- **mode:** open_close_failure
+- **reason:** The robot gripper contacts the upper edge of the open drawer around 4.5 seconds, but fails to exert the proper force trajectory to slide the drawer shut, remaining static until the episode ends.
+  Summary: failed to push drawer shut; gripper stuck on drawer rim
+- **recovery:** Release the rim, reposition the end-effector against the rear side of the drawer frame, and push forward to close it completely.
+  Summary: reposition gripper; push drawer fully closed
+- **prevention:** Approach the drawer at a lower angle and maintain forward force along the drawer sliding axis to push it fully closed.
+  Summary: align with drawer axis; push forward to close drawer
+- **justification:** The robot successfully reaches the correct target object (the drawer) at 4.5s, but fails to perform the required action to alter its state from open to closed, classifying this as an open_close_failure.
diff --git a/closer_drawer/ep027/labels.json b/closer_drawer/ep027/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..d0b6d3d0d7bc4fe7ecaa71d92f2b8bf669e5ead8
--- /dev/null
+++ b/closer_drawer/ep027/labels.json
@@ -0,0 +1,109 @@
+{
+  "rollout_id": "t006__ep027__closer_drawer",
+  "scene_variant": "",
+  "task_id": 6,
+  "task_file": "",
+  "instruction": "Close the drawer",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:27:16+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0,
+      "control_step_start": 0,
+      "control_step_end": 149,
+      "policy_step_start": 0,
+      "policy_step_end": 24,
+      "phrase": "lowering arm toward drawer",
+      "description": "lowering arm toward drawer"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0,
+      "t_end_sec": 6.0,
+      "control_step_start": 150,
+      "control_step_end": 299,
+      "policy_step_start": 25,
+      "policy_step_end": 49,
+      "phrase": "contacting upper drawer edge",
+      "description": "contacting upper drawer edge"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.0,
+      "t_end_sec": 9.0,
+      "control_step_start": 300,
+      "control_step_end": 449,
+      "policy_step_start": 50,
+      "policy_step_end": 74,
+      "phrase": "holding position at drawer edge",
+      "description": "holding position at drawer edge"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.0,
+      "t_end_sec": 11.8,
+      "control_step_start": 450,
+      "control_step_end": 589,
+      "policy_step_start": 75,
+      "policy_step_end": 98,
+      "phrase": "remaining stationary at drawer edge",
+      "description": "remaining stationary at drawer edge"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "open_close_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 4.2,
+    "vlm_failure_onset_timestamp": "00:04.2",
+    "vlm_failure_onset_frame": 210,
+    "vlm_failure_onset_step": 210,
+    "vlm_coarse_failure_onset_seconds": 4.5,
+    "vlm_failure_window_start_seconds": 4.0,
+    "vlm_failure_window_end_seconds": 5.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot gripper contacts the upper edge of the open drawer around 4.5 seconds, but fails to exert the proper force trajectory to slide the drawer shut, remaining static until the episode ends.",
+    "vlm_failure_reason_summary": "failed to push drawer shut; gripper stuck on drawer rim",
+    "vlm_recovery_action": "Release the rim, reposition the end-effector against the rear side of the drawer frame, and push forward to close it completely.",
+    "vlm_recovery_action_summary": "reposition gripper; push drawer fully closed",
+    "vlm_prevention_action": "Approach the drawer at a lower angle and maintain forward force along the drawer sliding axis to push it fully closed.",
+    "vlm_prevention_action_summary": "align with drawer axis; push forward to close drawer",
+    "vlm_justification": "The robot successfully reaches the correct target object (the drawer) at 4.5s, but fails to perform the required action to alter its state from open to closed, classifying this as an open_close_failure.",
+    "vlm_temporal_justification": "At clip second 135 (02:15), the robot end-effector makes contact with the upper rear edge of the open drawer. From this point onward, it fails to push forward to close the drawer, remaining stuck and static for the remainder of the clip.",
+    "vlm_video_fps": 50.0,
+    "vlm_video_num_frames": 590,
+    "vlm_video_duration_seconds": 11.78,
+    "vlm_usage": {
+      "input_tokens": 21575,
+      "output_tokens": 393,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"open_close_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 4.5,\n \"failure_window_start_seconds\": 4.0,\n \"failure_window_end_seconds\": 5.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot gripper contacts the upper edge of the open drawer around 4.5 seconds, but fails to exert the proper force trajectory to slide the drawer shut, remaining static until the episode ends.\",\n \"failure_reason_summary\": \"failed to push drawer shut; gripper stuck on drawer rim\",\n \"recovery_action\": \"Release the rim, reposition the end-effector against the rear side of the drawer frame, and push forward to close it completely.\",\n \"recovery_action_summary\": \"reposition gripper; push drawer fully closed\",\n \"prevention_action\": \"Approach the drawer at a lower angle and maintain forward force along the drawer sliding axis to push it fully closed.\",\n \"prevention_action_summary\": \"align with drawer axis; push forward to close drawer\",\n \"justification\": \"The robot successfully reaches the correct target object (the drawer) at 4.5s, but fails to perform the required action to alter its state from open to closed, classifying this as an open_close_failure.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 210,
+    "failure_sim_step": 210,
+    "failure_policy_step": 35,
+    "failure_chunk_index": 0,
+    "first_post_failure_policy_step": 36,
+    "failure_type": "open_close_failure",
+    "correction_action": "Release the rim, reposition the end-effector against the rear side of the drawer frame, and push forward to close it completely."
+  }
+}
\ No newline at end of file
diff --git a/closer_drawer/ep027/labels.npz b/closer_drawer/ep027/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..1184d7073079dfd05422f6925300c83ed7d23179
--- /dev/null
+++ b/closer_drawer/ep027/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c89ee8c49c57c2b4511b31b1065451b56a1c9d87b179351a50bfaa27307343db
+size 3885
diff --git a/closer_drawer/ep028/example.md b/closer_drawer/ep028/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..c0ae8569544624131209b042373489248c517a16
--- /dev/null
+++ b/closer_drawer/ep028/example.md
@@ -0,0 +1,23 @@
+# t006__ep028__closer_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Close the drawer
+- **success:** True
+- **control steps:** 725  (14.4s at 50 Hz)
+- **policy inferences:** 125  (replan_steps=6)
+- **video:** `closer_drawer/ep028/rollout.mp4`
+- **features:** `125` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching toward the drawer |
+| 3.0–6.0 | 152–303 | 25–50 | lowering gripper to drawer |
+| 6.0–9.0 | 304–455 | 50–75 | aligning gripper with flap |
+| 9.0–12.0 | 456–607 | 76–101 | pushing drawer flap forward |
+| 12.0–14.4 | 608–724 | 101–120 | closing the drawer completely |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/closer_drawer/ep028/labels.json b/closer_drawer/ep028/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..35a3f7ad998a9fbd038dbdaf7e56fd84fd77a53f
--- /dev/null
+++ b/closer_drawer/ep028/labels.json
@@ -0,0 +1,90 @@
+{
+  "rollout_id": "t006__ep028__closer_drawer",
+  "scene_variant": "",
+  "task_id": 6,
+  "task_file": "",
+  "instruction": "Close the drawer",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:27:24+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00990099009901,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching toward the drawer",
+      "description": "reaching toward the drawer"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00990099009901,
+      "t_end_sec": 6.01980198019802,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "lowering gripper to drawer",
+      "description": "lowering gripper to drawer"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01980198019802,
+      "t_end_sec": 9.029702970297029,
+      "control_step_start": 304,
+      "control_step_end": 455,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "aligning gripper with flap",
+      "description": "aligning gripper with flap"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.029702970297029,
+      "t_end_sec": 12.03960396039604,
+      "control_step_start": 456,
+      "control_step_end": 607,
+      "policy_step_start": 76,
+      "policy_step_end": 101,
+      "phrase": "pushing drawer flap forward",
+      "description": "pushing drawer flap forward"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 12.03960396039604,
+      "t_end_sec": 14.356435643564357,
+      "control_step_start": 608,
+      "control_step_end": 724,
+      "policy_step_start": 101,
+      "policy_step_end": 120,
+      "phrase": "closing the drawer completely",
+      "description": "closing the drawer completely"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/closer_drawer/ep028/labels.npz b/closer_drawer/ep028/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..2c5ea622820e163008a25322571d9c77ebb3f2ac
--- /dev/null
+++ b/closer_drawer/ep028/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a2a2ca16a38a8a74d57682640190691dd6c5650b09763ce2e96f9c1e94126d52
+size 3170
diff --git a/closer_drawer/ep029/example.md b/closer_drawer/ep029/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..3b8bd48eeb0827665e93c8dc14020c8e1ee2d7f8
--- /dev/null
+++ b/closer_drawer/ep029/example.md
@@ -0,0 +1,32 @@
+# t006__ep029__closer_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Close the drawer
+- **success:** False
+- **control steps:** 251  (4.9s at 51 Hz)
+- **policy inferences:** 44  (replan_steps=5)
+- **video:** `closer_drawer/ep029/rollout.mp4`
+- **features:** `44` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–152 | 0–30 | reaching for drawer handle |
+| 3.0–4.9 | 153–250 | 30–50 | holding drawer without pushing closed |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 2.7111984282907664 s (00:02.7)
+- **frame / control_step:** 138 (policy_step=27, chunk_index=3)
+- **coarse onset:** 2.7 s (refined=True)
+- **mode:** open_close_failure
+- **reason:** The robot reached the lip of the open drawer but failed to push or slide the drawer shut.
+  Summary: failed to push drawer; drawer remains open
+- **recovery:** Re-align the gripper with the drawer face and push forward until the drawer is fully closed.
+  Summary: re-align gripper with drawer; push drawer closed
+- **prevention:** Target the open drawer handle cleanly and apply sufficient forward force to close it completely.
+  Summary: approach drawer handle; push drawer shut
+- **justification:** At 2.7s, the robot arm makes contact with the open drawer lip/handle, but fails to apply the necessary force or trajectory to push the drawer shut, leaving it open for the remainder of the rollout.
diff --git a/closer_drawer/ep029/labels.json b/closer_drawer/ep029/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..4b52760213f2a17dbe14a6bfaa3da1c60f52a28e
--- /dev/null
+++ b/closer_drawer/ep029/labels.json
@@ -0,0 +1,87 @@
+{
+  "rollout_id": "t006__ep029__closer_drawer",
+  "scene_variant": "",
+  "task_id": 6,
+  "task_file": "",
+  "instruction": "Close the drawer",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:27:58+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.005893909626719,
+      "control_step_start": 0,
+      "control_step_end": 152,
+      "policy_step_start": 0,
+      "policy_step_end": 30,
+      "phrase": "reaching for drawer handle",
+      "description": "reaching for drawer handle"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.005893909626719,
+      "t_end_sec": 4.931237721021611,
+      "control_step_start": 153,
+      "control_step_end": 250,
+      "policy_step_start": 30,
+      "policy_step_end": 50,
+      "phrase": "holding drawer without pushing closed",
+      "description": "holding drawer without pushing closed"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "open_close_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 2.7111984282907664,
+    "vlm_failure_onset_timestamp": "00:02.7",
+    "vlm_failure_onset_frame": 138,
+    "vlm_failure_onset_step": 138,
+    "vlm_coarse_failure_onset_seconds": 2.7,
+    "vlm_failure_window_start_seconds": 2.0,
+    "vlm_failure_window_end_seconds": 4.91,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot reached the lip of the open drawer but failed to push or slide the drawer shut.",
+    "vlm_failure_reason_summary": "failed to push drawer; drawer remains open",
+    "vlm_recovery_action": "Re-align the gripper with the drawer face and push forward until the drawer is fully closed.",
+    "vlm_recovery_action_summary": "re-align gripper with drawer; push drawer closed",
+    "vlm_prevention_action": "Target the open drawer handle cleanly and apply sufficient forward force to close it completely.",
+    "vlm_prevention_action_summary": "approach drawer handle; push drawer shut",
+    "vlm_justification": "At 2.7s, the robot arm makes contact with the open drawer lip/handle, but fails to apply the necessary force or trajectory to push the drawer shut, leaving it open for the remainder of the rollout.",
+    "vlm_temporal_justification": "At frame 138 (02:18), the robot gripper finishes descending and closes around the drawer lip, but fails to push or slide the drawer closed, remaining stationary for the rest of the execution.",
+    "vlm_video_fps": 50.9,
+    "vlm_video_num_frames": 251,
+    "vlm_video_duration_seconds": 4.911591355599215,
+    "vlm_usage": {
+      "input_tokens": 17959,
+      "output_tokens": 344,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"open_close_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 2.7,\n \"failure_window_start_seconds\": 2.0,\n \"failure_window_end_seconds\": 4.91,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot reached the lip of the open drawer but failed to push or slide the drawer shut.\",\n \"failure_reason_summary\": \"failed to push drawer; drawer remains open\",\n \"recovery_action\": \"Re-align the gripper with the drawer face and push forward until the drawer is fully closed.\",\n \"recovery_action_summary\": \"re-align gripper with drawer; push drawer closed\",\n \"prevention_action\": \"Target the open drawer handle cleanly and apply sufficient forward force to close it completely.\",\n \"prevention_action_summary\": \"approach drawer handle; push drawer shut\",\n \"justification\": \"At 2.7s, the robot arm makes contact with the open drawer lip/handle, but fails to apply the necessary force or trajectory to push the drawer shut, leaving it open for the remainder of the rollout.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 138,
+    "failure_sim_step": 138,
+    "failure_policy_step": 27,
+    "failure_chunk_index": 3,
+    "first_post_failure_policy_step": 28,
+    "failure_type": "open_close_failure",
+    "correction_action": "Re-align the gripper with the drawer face and push forward until the drawer is fully closed."
+  }
+}
\ No newline at end of file
diff --git a/closer_drawer/ep029/labels.npz b/closer_drawer/ep029/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..f25ee622c99ba6bb5e75e132760a224399afa603
--- /dev/null
+++ b/closer_drawer/ep029/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b8c6cc4ed9c10113630772816a6b4c5628c8bbeff4d82fb5a16381bb30d68539
+size 3618
diff --git a/closer_drawer/ep030/example.md b/closer_drawer/ep030/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..6d468f07b77c51f9902a52d8f8ba2756108183d2
--- /dev/null
+++ b/closer_drawer/ep030/example.md
@@ -0,0 +1,34 @@
+# t006__ep030__closer_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Close the drawer
+- **success:** False
+- **control steps:** 569  (11.2s at 51 Hz)
+- **policy inferences:** 97  (replan_steps=6)
+- **video:** `closer_drawer/ep030/rollout.mp4`
+- **features:** `97` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | lowering gripper toward drawer |
+| 3.0–6.0 | 152–303 | 25–50 | pushing drawer lip and slipping |
+| 6.0–9.0 | 304–455 | 50–75 | retracting gripper away from drawer |
+| 9.0–11.2 | 456–568 | 76–94 | hovering idle above table |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 4.930966469428007 s (00:04.9)
+- **frame / control_step:** 250 (policy_step=41, chunk_index=4)
+- **coarse onset:** 5.5 s (refined=True)
+- **mode:** open_close_failure
+- **reason:** The robot attempted to push the extended drawer shut, but its gripper slipped off the edge of the drawer lip without pushing it closed.
+  Summary: gripper slipped off drawer edge; failed to close drawer
+- **recovery:** Re-align the gripper with the front lip of the extended drawer and push straight along the slide axis until the drawer is fully closed.
+  Summary: re-align with drawer handle; push drawer shut
+- **prevention:** Align the end-effector firmly against the center of the drawer face or lip before pushing, maintaining contact along the linear sliding direction.
+  Summary: align with drawer edge; apply straight push force
+- **justification:** At around 5.5 seconds, the gripper contacts the top drawer lip to push it closed, but slips off at 6.0 seconds. The drawer remains fully open for the remainder of the episode.
diff --git a/closer_drawer/ep030/labels.json b/closer_drawer/ep030/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..59816c8a8cb970c9ea73d4cae1adf60e321cc42b
--- /dev/null
+++ b/closer_drawer/ep030/labels.json
@@ -0,0 +1,109 @@
+{
+  "rollout_id": "t006__ep030__closer_drawer",
+  "scene_variant": "",
+  "task_id": 6,
+  "task_file": "",
+  "instruction": "Close the drawer",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:28:37+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9980276134122286,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "lowering gripper toward drawer",
+      "description": "lowering gripper toward drawer"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9980276134122286,
+      "t_end_sec": 5.996055226824457,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "pushing drawer lip and slipping",
+      "description": "pushing drawer lip and slipping"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.996055226824457,
+      "t_end_sec": 8.994082840236686,
+      "control_step_start": 304,
+      "control_step_end": 455,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "retracting gripper away from drawer",
+      "description": "retracting gripper away from drawer"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.994082840236686,
+      "t_end_sec": 11.222879684418146,
+      "control_step_start": 456,
+      "control_step_end": 568,
+      "policy_step_start": 76,
+      "policy_step_end": 94,
+      "phrase": "hovering idle above table",
+      "description": "hovering idle above table"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "open_close_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 4.930966469428007,
+    "vlm_failure_onset_timestamp": "00:04.9",
+    "vlm_failure_onset_frame": 250,
+    "vlm_failure_onset_step": 250,
+    "vlm_coarse_failure_onset_seconds": 5.5,
+    "vlm_failure_window_start_seconds": 5.0,
+    "vlm_failure_window_end_seconds": 7.0,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to push the extended drawer shut, but its gripper slipped off the edge of the drawer lip without pushing it closed.",
+    "vlm_failure_reason_summary": "gripper slipped off drawer edge; failed to close drawer",
+    "vlm_recovery_action": "Re-align the gripper with the front lip of the extended drawer and push straight along the slide axis until the drawer is fully closed.",
+    "vlm_recovery_action_summary": "re-align with drawer handle; push drawer shut",
+    "vlm_prevention_action": "Align the end-effector firmly against the center of the drawer face or lip before pushing, maintaining contact along the linear sliding direction.",
+    "vlm_prevention_action_summary": "align with drawer edge; apply straight push force",
+    "vlm_justification": "At around 5.5 seconds, the gripper contacts the top drawer lip to push it closed, but slips off at 6.0 seconds. The drawer remains fully open for the remainder of the episode.",
+    "vlm_temporal_justification": "At clip timestamp 02:03 (123 seconds into the clip), the robot gripper slips off the lip of the drawer handle and fails to push it closed, sliding past the drawer edge instead.",
+    "vlm_video_fps": 50.7,
+    "vlm_video_num_frames": 569,
+    "vlm_video_duration_seconds": 11.203155818540433,
+    "vlm_usage": {
+      "input_tokens": 21754,
+      "output_tokens": 370,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"open_close_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 5.5,\n \"failure_window_start_seconds\": 5.0,\n \"failure_window_end_seconds\": 7.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to push the extended drawer shut, but its gripper slipped off the edge of the drawer lip without pushing it closed.\",\n \"failure_reason_summary\": \"gripper slipped off drawer edge; failed to close drawer\",\n \"recovery_action\": \"Re-align the gripper with the front lip of the extended drawer and push straight along the slide axis until the drawer is fully closed.\",\n \"recovery_action_summary\": \"re-align with drawer handle; push drawer shut\",\n \"prevention_action\": \"Align the end-effector firmly against the center of the drawer face or lip before pushing, maintaining contact along the linear sliding direction.\",\n \"prevention_action_summary\": \"align with drawer edge; apply straight push force\",\n \"justification\": \"At around 5.5 seconds, the gripper contacts the top drawer lip to push it closed, but slips off at 6.0 seconds. The drawer remains fully open for the remainder of the episode.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 250,
+    "failure_sim_step": 250,
+    "failure_policy_step": 41,
+    "failure_chunk_index": 4,
+    "first_post_failure_policy_step": 42,
+    "failure_type": "open_close_failure",
+    "correction_action": "Re-align the gripper with the front lip of the extended drawer and push straight along the slide axis until the drawer is fully closed."
+  }
+}
\ No newline at end of file
diff --git a/closer_drawer/ep030/labels.npz b/closer_drawer/ep030/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..3267e2fb255ac27c87384875fc27fe0abd2377f6
--- /dev/null
+++ b/closer_drawer/ep030/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5a8dd146bf41d30bec12f1107b658712b34774d53309ce57d5e064e4a7b189d9
+size 3847
diff --git a/closer_drawer/ep031/example.md b/closer_drawer/ep031/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..3ced9591573d341960f8e7d2058dd9938d1f5682
--- /dev/null
+++ b/closer_drawer/ep031/example.md
@@ -0,0 +1,32 @@
+# t006__ep031__closer_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Close the drawer
+- **success:** False
+- **control steps:** 253  (5.0s at 51 Hz)
+- **policy inferences:** 44  (replan_steps=6)
+- **video:** `closer_drawer/ep031/rollout.mp4`
+- **features:** `44` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching down to drawer handle |
+| 3.0–5.0 | 152–252 | 25–42 | holding gripper near drawer |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 2.2134387351778657 s (00:02.2)
+- **frame / control_step:** 112 (policy_step=18, chunk_index=4)
+- **coarse onset:** 2.8 s (refined=True)
+- **mode:** open_close_failure
+- **reason:** The robot positions its gripper near the drawer lip but fails to exert force to push the drawer back into the main housing, leaving the drawer completely open.
+  Summary: failed to push drawer shut; drawer remained open
+- **recovery:** Re-position the end-effector against the front handle or lip of the open drawer and push it inward until the drawer is completely closed.
+  Summary: re-align with drawer handle; push drawer shut
+- **prevention:** Firmly contact the drawer handle or front panel and apply steady horizontal forward force to push the drawer into its closed position.
+  Summary: contact drawer handle; push horizontally to close
+- **justification:** Around 2.8 seconds into the video, the robot closes its end-effector near the open drawer lip, but fails to execute a pushing trajectory. As a result, the drawer stays open throughout the remainder of the rollout.
diff --git a/closer_drawer/ep031/labels.json b/closer_drawer/ep031/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..dfc9645c260f3959792ba566aee283662a7dffe2
--- /dev/null
+++ b/closer_drawer/ep031/labels.json
@@ -0,0 +1,87 @@
+{
+  "rollout_id": "t006__ep031__closer_drawer",
+  "scene_variant": "",
+  "task_id": 6,
+  "task_file": "",
+  "instruction": "Close the drawer",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:29:11+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0039525691699605,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching down to drawer handle",
+      "description": "reaching down to drawer handle"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0039525691699605,
+      "t_end_sec": 5.0,
+      "control_step_start": 152,
+      "control_step_end": 252,
+      "policy_step_start": 25,
+      "policy_step_end": 42,
+      "phrase": "holding gripper near drawer",
+      "description": "holding gripper near drawer"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "open_close_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 2.2134387351778657,
+    "vlm_failure_onset_timestamp": "00:02.2",
+    "vlm_failure_onset_frame": 112,
+    "vlm_failure_onset_step": 112,
+    "vlm_coarse_failure_onset_seconds": 2.8,
+    "vlm_failure_window_start_seconds": 2.3,
+    "vlm_failure_window_end_seconds": 3.3,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot positions its gripper near the drawer lip but fails to exert force to push the drawer back into the main housing, leaving the drawer completely open.",
+    "vlm_failure_reason_summary": "failed to push drawer shut; drawer remained open",
+    "vlm_recovery_action": "Re-position the end-effector against the front handle or lip of the open drawer and push it inward until the drawer is completely closed.",
+    "vlm_recovery_action_summary": "re-align with drawer handle; push drawer shut",
+    "vlm_prevention_action": "Firmly contact the drawer handle or front panel and apply steady horizontal forward force to push the drawer into its closed position.",
+    "vlm_prevention_action_summary": "contact drawer handle; push horizontally to close",
+    "vlm_justification": "Around 2.8 seconds into the video, the robot closes its end-effector near the open drawer lip, but fails to execute a pushing trajectory. As a result, the drawer stays open throughout the remainder of the rollout.",
+    "vlm_temporal_justification": "At second 112 (frame 112), the gripper finishes lowering into position right next to the drawer lip, but it fails to apply any forward force to push the drawer back into its housing.",
+    "vlm_video_fps": 50.6,
+    "vlm_video_num_frames": 253,
+    "vlm_video_duration_seconds": 4.980237154150197,
+    "vlm_usage": {
+      "input_tokens": 18096,
+      "output_tokens": 373,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"open_close_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 2.8,\n \"failure_window_start_seconds\": 2.3,\n \"failure_window_end_seconds\": 3.3,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot positions its gripper near the drawer lip but fails to exert force to push the drawer back into the main housing, leaving the drawer completely open.\",\n \"failure_reason_summary\": \"failed to push drawer shut; drawer remained open\",\n \"recovery_action\": \"Re-position the end-effector against the front handle or lip of the open drawer and push it inward until the drawer is completely closed.\",\n \"recovery_action_summary\": \"re-align with drawer handle; push drawer shut\",\n \"prevention_action\": \"Firmly contact the drawer handle or front panel and apply steady horizontal forward force to push the drawer into its closed position.\",\n \"prevention_action_summary\": \"contact drawer handle; push horizontally to close\",\n \"justification\": \"Around 2.8 seconds into the video, the robot closes its end-effector near the open drawer lip, but fails to execute a pushing trajectory. As a result, the drawer stays open throughout the remainder of the rollout.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 112,
+    "failure_sim_step": 112,
+    "failure_policy_step": 18,
+    "failure_chunk_index": 4,
+    "first_post_failure_policy_step": 19,
+    "failure_type": "open_close_failure",
+    "correction_action": "Re-position the end-effector against the front handle or lip of the open drawer and push it inward until the drawer is completely closed."
+  }
+}
\ No newline at end of file
diff --git a/closer_drawer/ep031/labels.npz b/closer_drawer/ep031/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..d247bdc19995ef72d60b792c52476ad85ec3bc3f
--- /dev/null
+++ b/closer_drawer/ep031/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:60f0da3427cbfdc3f31050331723aace790add052069990da5f993ac70dc57f1
+size 3770
diff --git a/closer_drawer/ep032/example.md b/closer_drawer/ep032/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..392d3f50ede1aaaa3a39342f009153ecab21327c
--- /dev/null
+++ b/closer_drawer/ep032/example.md
@@ -0,0 +1,20 @@
+# t006__ep032__closer_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Close the drawer
+- **success:** True
+- **control steps:** 279  (5.5s at 50 Hz)
+- **policy inferences:** 48  (replan_steps=6)
+- **video:** `closer_drawer/ep032/rollout.mp4`
+- **features:** `48` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | reaching and pushing drawer closed |
+| 3.0–5.5 | 151–278 | 25–46 | closing drawer and retracting arm |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/closer_drawer/ep032/labels.json b/closer_drawer/ep032/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..4246358bc0f395f579a04a034fc99777f7b27aea
--- /dev/null
+++ b/closer_drawer/ep032/labels.json
@@ -0,0 +1,57 @@
+{
+  "rollout_id": "t006__ep032__closer_drawer",
+  "scene_variant": "",
+  "task_id": 6,
+  "task_file": "",
+  "instruction": "Close the drawer",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:29:20+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.996031746031746,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching and pushing drawer closed",
+      "description": "reaching and pushing drawer closed"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.996031746031746,
+      "t_end_sec": 5.535714285714286,
+      "control_step_start": 151,
+      "control_step_end": 278,
+      "policy_step_start": 25,
+      "policy_step_end": 46,
+      "phrase": "closing drawer and retracting arm",
+      "description": "closing drawer and retracting arm"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/closer_drawer/ep032/labels.npz b/closer_drawer/ep032/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..1081674f6d9275e16445e3eb7b442e3d081d3111
--- /dev/null
+++ b/closer_drawer/ep032/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ab54ece7920f306ef0c6e340a47cbe135461686205ec3fad455c5b9a4586e3df
+size 3041
diff --git a/closer_drawer/ep033/example.md b/closer_drawer/ep033/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..78addfc63a11664c6de83d1dc41845d65ae81e30
--- /dev/null
+++ b/closer_drawer/ep033/example.md
@@ -0,0 +1,33 @@
+# t006__ep033__closer_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Close the drawer
+- **success:** False
+- **control steps:** 311  (6.2s at 50 Hz)
+- **policy inferences:** 54  (replan_steps=6)
+- **video:** `closer_drawer/ep033/rollout.mp4`
+- **features:** `54` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–149 | 0–24 | lowering arm toward open drawer |
+| 3.0–6.0 | 150–299 | 25–49 | hovering near drawer lip |
+| 6.0–6.2 | 300–310 | 50–51 | idle near open drawer |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 1.6167664670658681 s (00:01.6)
+- **frame / control_step:** 81 (policy_step=13, chunk_index=3)
+- **coarse onset:** 2.7 s (refined=True)
+- **mode:** open_close_failure
+- **reason:** The robot reached for the drawer lip but failed to align its end effector properly to hook or push the drawer, leaving it open.
+  Summary: failed grasp on drawer lip; drawer left open
+- **recovery:** Reposition the gripper to align with the front handle or upper lip of the drawer, grasp or push firmly, and slide the drawer completely shut.
+  Summary: realign gripper with drawer; push drawer shut
+- **prevention:** Align the end effector precisely with the drawer handle before attempting to push or pull the drawer shut.
+  Summary: align gripper with drawer handle; push drawer shut
+- **justification:** The robot approaches the drawer around 2.2s–2.7s, but fails to make proper contact with the drawer lip/handle to push or pull it, leaving the drawer unmoved and open until the rollout ends.
diff --git a/closer_drawer/ep033/labels.json b/closer_drawer/ep033/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..0dd4ff061a7b3d855211cfcbd5b835f1e84e47ac
--- /dev/null
+++ b/closer_drawer/ep033/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t006__ep033__closer_drawer",
+  "scene_variant": "",
+  "task_id": 6,
+  "task_file": "",
+  "instruction": "Close the drawer",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:29:53+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9940119760479043,
+      "control_step_start": 0,
+      "control_step_end": 149,
+      "policy_step_start": 0,
+      "policy_step_end": 24,
+      "phrase": "lowering arm toward open drawer",
+      "description": "lowering arm toward open drawer"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9940119760479043,
+      "t_end_sec": 5.9880239520958085,
+      "control_step_start": 150,
+      "control_step_end": 299,
+      "policy_step_start": 25,
+      "policy_step_end": 49,
+      "phrase": "hovering near drawer lip",
+      "description": "hovering near drawer lip"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.9880239520958085,
+      "t_end_sec": 6.2075848303393215,
+      "control_step_start": 300,
+      "control_step_end": 310,
+      "policy_step_start": 50,
+      "policy_step_end": 51,
+      "phrase": "idle near open drawer",
+      "description": "idle near open drawer"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "open_close_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 1.6167664670658681,
+    "vlm_failure_onset_timestamp": "00:01.6",
+    "vlm_failure_onset_frame": 81,
+    "vlm_failure_onset_step": 81,
+    "vlm_coarse_failure_onset_seconds": 2.7,
+    "vlm_failure_window_start_seconds": 2.2,
+    "vlm_failure_window_end_seconds": 3.2,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot reached for the drawer lip but failed to align its end effector properly to hook or push the drawer, leaving it open.",
+    "vlm_failure_reason_summary": "failed grasp on drawer lip; drawer left open",
+    "vlm_recovery_action": "Reposition the gripper to align with the front handle or upper lip of the drawer, grasp or push firmly, and slide the drawer completely shut.",
+    "vlm_recovery_action_summary": "realign gripper with drawer; push drawer shut",
+    "vlm_prevention_action": "Align the end effector precisely with the drawer handle before attempting to push or pull the drawer shut.",
+    "vlm_prevention_action_summary": "align gripper with drawer handle; push drawer shut",
+    "vlm_justification": "The robot approaches the drawer around 2.2s\u20132.7s, but fails to make proper contact with the drawer lip/handle to push or pull it, leaving the drawer unmoved and open until the rollout ends.",
+    "vlm_temporal_justification": "At second 81 (01:21 in the video), the robot's end effector descends directly over the open drawer cavity rather than aligning with the front lip or handle, causing it to lower inside the drawer and fail to close it.",
+    "vlm_video_fps": 50.1,
+    "vlm_video_num_frames": 311,
+    "vlm_video_duration_seconds": 6.187624750499002,
+    "vlm_usage": {
+      "input_tokens": 20234,
+      "output_tokens": 373,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"open_close_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 2.7,\n \"failure_window_start_seconds\": 2.2,\n \"failure_window_end_seconds\": 3.2,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot reached for the drawer lip but failed to align its end effector properly to hook or push the drawer, leaving it open.\",\n \"failure_reason_summary\": \"failed grasp on drawer lip; drawer left open\",\n \"recovery_action\": \"Reposition the gripper to align with the front handle or upper lip of the drawer, grasp or push firmly, and slide the drawer completely shut.\",\n \"recovery_action_summary\": \"realign gripper with drawer; push drawer shut\",\n \"prevention_action\": \"Align the end effector precisely with the drawer handle before attempting to push or pull the drawer shut.\",\n \"prevention_action_summary\": \"align gripper with drawer handle; push drawer shut\",\n \"justification\": \"The robot approaches the drawer around 2.2s\u20132.7s, but fails to make proper contact with the drawer lip/handle to push or pull it, leaving the drawer unmoved and open until the rollout ends.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 81,
+    "failure_sim_step": 81,
+    "failure_policy_step": 13,
+    "failure_chunk_index": 3,
+    "first_post_failure_policy_step": 14,
+    "failure_type": "open_close_failure",
+    "correction_action": "Reposition the gripper to align with the front handle or upper lip of the drawer, grasp or push firmly, and slide the drawer completely shut."
+  }
+}
\ No newline at end of file
diff --git a/closer_drawer/ep033/labels.npz b/closer_drawer/ep033/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..8e6dbda9326653293d5fca2d83579086e804e7df
--- /dev/null
+++ b/closer_drawer/ep033/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:59c87d799103c14144b002cb5535c7a025239491962b95c0f078a8fc430eccb4
+size 3730
diff --git a/closer_drawer/ep034/example.md b/closer_drawer/ep034/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..6b1410c890a52e268dc5970d7c486e31756df854
--- /dev/null
+++ b/closer_drawer/ep034/example.md
@@ -0,0 +1,32 @@
+# t006__ep034__closer_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Close the drawer
+- **success:** False
+- **control steps:** 181  (3.6s at 50 Hz)
+- **policy inferences:** 32  (replan_steps=6)
+- **video:** `closer_drawer/ep034/rollout.mp4`
+- **features:** `32` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | lowering toward drawer handle |
+| 3.0–3.6 | 151–180 | 25–30 | hovering over drawer |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 2.385685884691849 s (00:02.4)
+- **frame / control_step:** 120 (policy_step=20, chunk_index=0)
+- **coarse onset:** 2.0 s (refined=True)
+- **mode:** open_close_failure
+- **reason:** The robot attempted to grasp and push the open drawer handle, but failed to make secure contact or apply proper force, leaving the drawer completely open.
+  Summary: failed grip on drawer handle; drawer remains open
+- **recovery:** Realign the gripper with the front handle of the drawer and apply forward pushing force to slide the drawer shut.
+  Summary: realign gripper; push drawer closed
+- **prevention:** Approach the drawer handle with proper alignment and orientation so the fingers firmly engage the handle before initiating the closing motion.
+  Summary: align gripper with handle; push drawer shut
+- **justification:** At around 2.0 seconds, the gripper lowers toward the open drawer handle but fails to establish a proper grip or push contact. As a result, the drawer remains open for the rest of the rollout, resulting in an open_close_failure.
diff --git a/closer_drawer/ep034/labels.json b/closer_drawer/ep034/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..1c66b13f17623323ab0ca5c39769ccfc5b16af29
--- /dev/null
+++ b/closer_drawer/ep034/labels.json
@@ -0,0 +1,87 @@
+{
+  "rollout_id": "t006__ep034__closer_drawer",
+  "scene_variant": "",
+  "task_id": 6,
+  "task_file": "",
+  "instruction": "Close the drawer",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:30:21+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.001988071570577,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "lowering toward drawer handle",
+      "description": "lowering toward drawer handle"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.001988071570577,
+      "t_end_sec": 3.598409542743539,
+      "control_step_start": 151,
+      "control_step_end": 180,
+      "policy_step_start": 25,
+      "policy_step_end": 30,
+      "phrase": "hovering over drawer",
+      "description": "hovering over drawer"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "open_close_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 2.385685884691849,
+    "vlm_failure_onset_timestamp": "00:02.4",
+    "vlm_failure_onset_frame": 120,
+    "vlm_failure_onset_step": 120,
+    "vlm_coarse_failure_onset_seconds": 2.0,
+    "vlm_failure_window_start_seconds": 1.8,
+    "vlm_failure_window_end_seconds": 2.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp and push the open drawer handle, but failed to make secure contact or apply proper force, leaving the drawer completely open.",
+    "vlm_failure_reason_summary": "failed grip on drawer handle; drawer remains open",
+    "vlm_recovery_action": "Realign the gripper with the front handle of the drawer and apply forward pushing force to slide the drawer shut.",
+    "vlm_recovery_action_summary": "realign gripper; push drawer closed",
+    "vlm_prevention_action": "Approach the drawer handle with proper alignment and orientation so the fingers firmly engage the handle before initiating the closing motion.",
+    "vlm_prevention_action_summary": "align gripper with handle; push drawer shut",
+    "vlm_justification": "At around 2.0 seconds, the gripper lowers toward the open drawer handle but fails to establish a proper grip or push contact. As a result, the drawer remains open for the rest of the rollout, resulting in an open_close_failure.",
+    "vlm_temporal_justification": "At frame 120 (02:00 in the clip), the robot gripper closes around the drawer edge without securing a proper contact or grip required to push the drawer closed, causing the subsequent push attempt to fail.",
+    "vlm_video_fps": 50.3,
+    "vlm_video_num_frames": 181,
+    "vlm_video_duration_seconds": 3.5785288270377738,
+    "vlm_usage": {
+      "input_tokens": 13496,
+      "output_tokens": 369,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"open_close_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 2.0,\n \"failure_window_start_seconds\": 1.8,\n \"failure_window_end_seconds\": 2.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp and push the open drawer handle, but failed to make secure contact or apply proper force, leaving the drawer completely open.\",\n \"failure_reason_summary\": \"failed grip on drawer handle; drawer remains open\",\n \"recovery_action\": \"Realign the gripper with the front handle of the drawer and apply forward pushing force to slide the drawer shut.\",\n \"recovery_action_summary\": \"realign gripper; push drawer closed\",\n \"prevention_action\": \"Approach the drawer handle with proper alignment and orientation so the fingers firmly engage the handle before initiating the closing motion.\",\n \"prevention_action_summary\": \"align gripper with handle; push drawer shut\",\n \"justification\": \"At around 2.0 seconds, the gripper lowers toward the open drawer handle but fails to establish a proper grip or push contact. As a result, the drawer remains open for the rest of the rollout, resulting in an open_close_failure.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 120,
+    "failure_sim_step": 120,
+    "failure_policy_step": 20,
+    "failure_chunk_index": 0,
+    "first_post_failure_policy_step": 21,
+    "failure_type": "open_close_failure",
+    "correction_action": "Realign the gripper with the front handle of the drawer and apply forward pushing force to slide the drawer shut."
+  }
+}
\ No newline at end of file
diff --git a/closer_drawer/ep034/labels.npz b/closer_drawer/ep034/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..06107fa7eb13193a23f583396228aee0c6fc7654
--- /dev/null
+++ b/closer_drawer/ep034/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:566f59d93a0036e7661a86d9dd19289e3a73629b9ea21faab75b60a526e8214a
+size 3706
diff --git a/closer_drawer/ep035/example.md b/closer_drawer/ep035/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..bd37871a01ed288afbab0c44c4623173af3abbc8
--- /dev/null
+++ b/closer_drawer/ep035/example.md
@@ -0,0 +1,32 @@
+# t006__ep035__closer_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Close the drawer
+- **success:** False
+- **control steps:** 226  (4.5s at 50 Hz)
+- **policy inferences:** 39  (replan_steps=6)
+- **video:** `closer_drawer/ep035/rollout.mp4`
+- **features:** `39` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–149 | 0–24 | lowering gripper toward drawer handle |
+| 3.0–4.5 | 150–225 | 25–37 | hovering at drawer without pushing |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 2.5948103792415167 s (00:02.6)
+- **frame / control_step:** 130 (policy_step=21, chunk_index=4)
+- **coarse onset:** 2.2 s (refined=True)
+- **mode:** open_close_failure
+- **reason:** The robot reaches for the open drawer lip to close it, but fails to make proper contact or exert sufficient pushing force, leaving the drawer completely open.
+  Summary: failed contact with drawer edge; drawer remaining open
+- **recovery:** Re-align the gripper with the front edge or handle of the open drawer and apply inward force until the drawer slides fully shut.
+  Summary: align gripper with drawer; push drawer closed
+- **prevention:** Position the gripper precisely against the open drawer handle or front face before attempting the forward pushing motion to ensure the drawer is pushed shut.
+  Summary: align with drawer edge; push drawer smoothly shut
+- **justification:** At approximately 2.2 seconds, the end-effector reaches the open drawer lip but fails to grip or push it inward, hovering near the drawer without moving it for the rest of the rollout.
diff --git a/closer_drawer/ep035/labels.json b/closer_drawer/ep035/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..56a295563d5be9b8016ff09e07a9e4de189b7e1b
--- /dev/null
+++ b/closer_drawer/ep035/labels.json
@@ -0,0 +1,87 @@
+{
+  "rollout_id": "t006__ep035__closer_drawer",
+  "scene_variant": "",
+  "task_id": 6,
+  "task_file": "",
+  "instruction": "Close the drawer",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:30:53+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9940119760479043,
+      "control_step_start": 0,
+      "control_step_end": 149,
+      "policy_step_start": 0,
+      "policy_step_end": 24,
+      "phrase": "lowering gripper toward drawer handle",
+      "description": "lowering gripper toward drawer handle"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9940119760479043,
+      "t_end_sec": 4.510978043912176,
+      "control_step_start": 150,
+      "control_step_end": 225,
+      "policy_step_start": 25,
+      "policy_step_end": 37,
+      "phrase": "hovering at drawer without pushing",
+      "description": "hovering at drawer without pushing"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "open_close_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 2.5948103792415167,
+    "vlm_failure_onset_timestamp": "00:02.6",
+    "vlm_failure_onset_frame": 130,
+    "vlm_failure_onset_step": 130,
+    "vlm_coarse_failure_onset_seconds": 2.2,
+    "vlm_failure_window_start_seconds": 1.7,
+    "vlm_failure_window_end_seconds": 3.2,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot reaches for the open drawer lip to close it, but fails to make proper contact or exert sufficient pushing force, leaving the drawer completely open.",
+    "vlm_failure_reason_summary": "failed contact with drawer edge; drawer remaining open",
+    "vlm_recovery_action": "Re-align the gripper with the front edge or handle of the open drawer and apply inward force until the drawer slides fully shut.",
+    "vlm_recovery_action_summary": "align gripper with drawer; push drawer closed",
+    "vlm_prevention_action": "Position the gripper precisely against the open drawer handle or front face before attempting the forward pushing motion to ensure the drawer is pushed shut.",
+    "vlm_prevention_action_summary": "align with drawer edge; push drawer smoothly shut",
+    "vlm_justification": "At approximately 2.2 seconds, the end-effector reaches the open drawer lip but fails to grip or push it inward, hovering near the drawer without moving it for the rest of the rollout.",
+    "vlm_temporal_justification": "At clip second 130 (frame 130), the gripper reaches the drawer lip and attempts to contact/close it, but fails to hook or exert pushing force to move the drawer, leaving it open.",
+    "vlm_video_fps": 50.1,
+    "vlm_video_num_frames": 226,
+    "vlm_video_duration_seconds": 4.491017964071856,
+    "vlm_usage": {
+      "input_tokens": 16395,
+      "output_tokens": 368,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"open_close_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 2.2,\n \"failure_window_start_seconds\": 1.7,\n \"failure_window_end_seconds\": 3.2,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot reaches for the open drawer lip to close it, but fails to make proper contact or exert sufficient pushing force, leaving the drawer completely open.\",\n \"failure_reason_summary\": \"failed contact with drawer edge; drawer remaining open\",\n \"recovery_action\": \"Re-align the gripper with the front edge or handle of the open drawer and apply inward force until the drawer slides fully shut.\",\n \"recovery_action_summary\": \"align gripper with drawer; push drawer closed\",\n \"prevention_action\": \"Position the gripper precisely against the open drawer handle or front face before attempting the forward pushing motion to ensure the drawer is pushed shut.\",\n \"prevention_action_summary\": \"align with drawer edge; push drawer smoothly shut\",\n \"justification\": \"At approximately 2.2 seconds, the end-effector reaches the open drawer lip but fails to grip or push it inward, hovering near the drawer without moving it for the rest of the rollout.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 130,
+    "failure_sim_step": 130,
+    "failure_policy_step": 21,
+    "failure_chunk_index": 4,
+    "first_post_failure_policy_step": 22,
+    "failure_type": "open_close_failure",
+    "correction_action": "Re-align the gripper with the front edge or handle of the open drawer and apply inward force until the drawer slides fully shut."
+  }
+}
\ No newline at end of file
diff --git a/closer_drawer/ep035/labels.npz b/closer_drawer/ep035/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..7ce554545ec95aa756c55c655b7c1128ded821b3
--- /dev/null
+++ b/closer_drawer/ep035/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:509a12017c582ea7b98751b90995a484a148ef362944ebbc77f5452ada3fe061
+size 3787
diff --git a/closer_drawer/ep036/example.md b/closer_drawer/ep036/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..51b96bb2c05151aca9dba53042c41078f187f5b4
--- /dev/null
+++ b/closer_drawer/ep036/example.md
@@ -0,0 +1,32 @@
+# t006__ep036__closer_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Close the drawer
+- **success:** False
+- **control steps:** 209  (4.1s at 51 Hz)
+- **policy inferences:** 35  (replan_steps=6)
+- **video:** `closer_drawer/ep036/rollout.mp4`
+- **features:** `35` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | descending toward drawer; closing gripper |
+| 3.0–4.1 | 152–208 | 25–34 | holding near drawer lip; remaining stationary |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 2.066929133858268 s (00:02.1)
+- **frame / control_step:** 105 (policy_step=17, chunk_index=3)
+- **coarse onset:** 2.0 s (refined=True)
+- **mode:** open_close_failure
+- **reason:** The robot end effector approaches the open drawer but fails to grasp or contact the handle properly to push the drawer back into its closed position.
+  Summary: failed to push drawer shut; drawer left open
+- **recovery:** Realign the gripper with the front edge or handle of the open drawer, establish contact, and push the drawer inward until fully closed.
+  Summary: realign gripper with drawer lip; push drawer closed
+- **prevention:** Position the gripper precisely at the drawer handle before moving forward to push the drawer smoothly into the closed state.
+  Summary: align gripper with drawer handle; push drawer fully shut
+- **justification:** At approximately 2.0 seconds, the robot reaches the open drawer handle area but fails to exert the necessary force or correct positioning to slide the drawer shut, leaving the task incomplete.
diff --git a/closer_drawer/ep036/labels.json b/closer_drawer/ep036/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..e30eac6aaac35a10c6ea3add46cceb376d6590cc
--- /dev/null
+++ b/closer_drawer/ep036/labels.json
@@ -0,0 +1,87 @@
+{
+  "rollout_id": "t006__ep036__closer_drawer",
+  "scene_variant": "",
+  "task_id": 6,
+  "task_file": "",
+  "instruction": "Close the drawer",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:31:22+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9921259842519685,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "descending toward drawer; closing gripper",
+      "description": "descending toward drawer; closing gripper"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9921259842519685,
+      "t_end_sec": 4.1141732283464565,
+      "control_step_start": 152,
+      "control_step_end": 208,
+      "policy_step_start": 25,
+      "policy_step_end": 34,
+      "phrase": "holding near drawer lip; remaining stationary",
+      "description": "holding near drawer lip; remaining stationary"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "open_close_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 2.066929133858268,
+    "vlm_failure_onset_timestamp": "00:02.1",
+    "vlm_failure_onset_frame": 105,
+    "vlm_failure_onset_step": 105,
+    "vlm_coarse_failure_onset_seconds": 2.0,
+    "vlm_failure_window_start_seconds": 2.0,
+    "vlm_failure_window_end_seconds": 4.09,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot end effector approaches the open drawer but fails to grasp or contact the handle properly to push the drawer back into its closed position.",
+    "vlm_failure_reason_summary": "failed to push drawer shut; drawer left open",
+    "vlm_recovery_action": "Realign the gripper with the front edge or handle of the open drawer, establish contact, and push the drawer inward until fully closed.",
+    "vlm_recovery_action_summary": "realign gripper with drawer lip; push drawer closed",
+    "vlm_prevention_action": "Position the gripper precisely at the drawer handle before moving forward to push the drawer smoothly into the closed state.",
+    "vlm_prevention_action_summary": "align gripper with drawer handle; push drawer fully shut",
+    "vlm_justification": "At approximately 2.0 seconds, the robot reaches the open drawer handle area but fails to exert the necessary force or correct positioning to slide the drawer shut, leaving the task incomplete.",
+    "vlm_temporal_justification": "At second 105 (corresponding to frame 105 of the original rollout), the robot's end effector finishes descending into the open drawer cavity rather than properly aligning with or grasping the front handle/edge required to push the drawer closed.",
+    "vlm_video_fps": 50.8,
+    "vlm_video_num_frames": 209,
+    "vlm_video_duration_seconds": 4.094488188976378,
+    "vlm_usage": {
+      "input_tokens": 15258,
+      "output_tokens": 368,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"open_close_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 2.0,\n \"failure_window_start_seconds\": 2.0,\n \"failure_window_end_seconds\": 4.09,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot end effector approaches the open drawer but fails to grasp or contact the handle properly to push the drawer back into its closed position.\",\n \"failure_reason_summary\": \"failed to push drawer shut; drawer left open\",\n \"recovery_action\": \"Realign the gripper with the front edge or handle of the open drawer, establish contact, and push the drawer inward until fully closed.\",\n \"recovery_action_summary\": \"realign gripper with drawer lip; push drawer closed\",\n \"prevention_action\": \"Position the gripper precisely at the drawer handle before moving forward to push the drawer smoothly into the closed state.\",\n \"prevention_action_summary\": \"align gripper with drawer handle; push drawer fully shut\",\n \"justification\": \"At approximately 2.0 seconds, the robot reaches the open drawer handle area but fails to exert the necessary force or correct positioning to slide the drawer shut, leaving the task incomplete.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 105,
+    "failure_sim_step": 105,
+    "failure_policy_step": 17,
+    "failure_chunk_index": 3,
+    "first_post_failure_policy_step": 18,
+    "failure_type": "open_close_failure",
+    "correction_action": "Realign the gripper with the front edge or handle of the open drawer, establish contact, and push the drawer inward until fully closed."
+  }
+}
\ No newline at end of file
diff --git a/closer_drawer/ep036/labels.npz b/closer_drawer/ep036/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..5cbe6ae8cc1ebd38a12c6e59ba2788c81e62ea16
--- /dev/null
+++ b/closer_drawer/ep036/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e972156e3b00a544b23e5b54b7fba16e66fa0651bdf5d7a35837faaf1bf9b949
+size 3774
diff --git a/closer_drawer/ep037/example.md b/closer_drawer/ep037/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..0d2904a63628d236b7e8b8199369d653956ae434
--- /dev/null
+++ b/closer_drawer/ep037/example.md
@@ -0,0 +1,32 @@
+# t006__ep037__closer_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Close the drawer
+- **success:** False
+- **control steps:** 259  (5.1s at 51 Hz)
+- **policy inferences:** 44  (replan_steps=6)
+- **video:** `closer_drawer/ep037/rollout.mp4`
+- **features:** `44` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–152 | 0–25 | approaching and grasping drawer lip |
+| 3.0–5.1 | 153–258 | 25–43 | holding drawer lip; no progress |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 2.700587084148728 s (00:02.7)
+- **frame / control_step:** 138 (policy_step=23, chunk_index=0)
+- **coarse onset:** 3.0 s (refined=True)
+- **mode:** open_close_failure
+- **reason:** The robot reached down and gripped the lip of the open drawer around 3.0 seconds, but failed to push the drawer horizontally back into its housing, leaving the drawer completely open.
+  Summary: drawer grasped but not pushed shut; container remains open
+- **recovery:** Reposition the end-effector to push against the front face of the drawer and exert forward pressure until the drawer slides completely into the box.
+  Summary: reposition gripper; push drawer fully closed
+- **prevention:** Maintain a straight inward trajectory toward the drawer box after contacting the drawer handle/lip to push the drawer fully closed.
+  Summary: push drawer directly inward; ensure full closure
+- **justification:** The robot reaches and contacts the open drawer around 3.0 seconds, but does not execute the required pushing motion to close it, making 3.0 seconds the onset of the failure behavior.
diff --git a/closer_drawer/ep037/labels.json b/closer_drawer/ep037/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..6b620df547401534a51d874f2de9dec90a441e93
--- /dev/null
+++ b/closer_drawer/ep037/labels.json
@@ -0,0 +1,87 @@
+{
+  "rollout_id": "t006__ep037__closer_drawer",
+  "scene_variant": "",
+  "task_id": 6,
+  "task_file": "",
+  "instruction": "Close the drawer",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:31:53+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.99412915851272,
+      "control_step_start": 0,
+      "control_step_end": 152,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "approaching and grasping drawer lip",
+      "description": "approaching and grasping drawer lip"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.99412915851272,
+      "t_end_sec": 5.068493150684931,
+      "control_step_start": 153,
+      "control_step_end": 258,
+      "policy_step_start": 25,
+      "policy_step_end": 43,
+      "phrase": "holding drawer lip; no progress",
+      "description": "holding drawer lip; no progress"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "open_close_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 2.700587084148728,
+    "vlm_failure_onset_timestamp": "00:02.7",
+    "vlm_failure_onset_frame": 138,
+    "vlm_failure_onset_step": 138,
+    "vlm_coarse_failure_onset_seconds": 3.0,
+    "vlm_failure_window_start_seconds": 2.5,
+    "vlm_failure_window_end_seconds": 5.05,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot reached down and gripped the lip of the open drawer around 3.0 seconds, but failed to push the drawer horizontally back into its housing, leaving the drawer completely open.",
+    "vlm_failure_reason_summary": "drawer grasped but not pushed shut; container remains open",
+    "vlm_recovery_action": "Reposition the end-effector to push against the front face of the drawer and exert forward pressure until the drawer slides completely into the box.",
+    "vlm_recovery_action_summary": "reposition gripper; push drawer fully closed",
+    "vlm_prevention_action": "Maintain a straight inward trajectory toward the drawer box after contacting the drawer handle/lip to push the drawer fully closed.",
+    "vlm_prevention_action_summary": "push drawer directly inward; ensure full closure",
+    "vlm_justification": "The robot reaches and contacts the open drawer around 3.0 seconds, but does not execute the required pushing motion to close it, making 3.0 seconds the onset of the failure behavior.",
+    "vlm_temporal_justification": "The robot successfully grips the lip of the open drawer around timestamp 02:12 - 02:15. However, at 02:18, instead of pushing the drawer horizontally back into its housing to close it, the robot arm stalls in place without completing the action.",
+    "vlm_video_fps": 51.1,
+    "vlm_video_num_frames": 259,
+    "vlm_video_duration_seconds": 5.048923679060665,
+    "vlm_usage": {
+      "input_tokens": 18481,
+      "output_tokens": 391,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"open_close_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 3.0,\n \"failure_window_start_seconds\": 2.5,\n \"failure_window_end_seconds\": 5.05,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot reached down and gripped the lip of the open drawer around 3.0 seconds, but failed to push the drawer horizontally back into its housing, leaving the drawer completely open.\",\n \"failure_reason_summary\": \"drawer grasped but not pushed shut; container remains open\",\n \"recovery_action\": \"Reposition the end-effector to push against the front face of the drawer and exert forward pressure until the drawer slides completely into the box.\",\n \"recovery_action_summary\": \"reposition gripper; push drawer fully closed\",\n \"prevention_action\": \"Maintain a straight inward trajectory toward the drawer box after contacting the drawer handle/lip to push the drawer fully closed.\",\n \"prevention_action_summary\": \"push drawer directly inward; ensure full closure\",\n \"justification\": \"The robot reaches and contacts the open drawer around 3.0 seconds, but does not execute the required pushing motion to close it, making 3.0 seconds the onset of the failure behavior.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 138,
+    "failure_sim_step": 138,
+    "failure_policy_step": 23,
+    "failure_chunk_index": 0,
+    "first_post_failure_policy_step": 24,
+    "failure_type": "open_close_failure",
+    "correction_action": "Reposition the end-effector to push against the front face of the drawer and exert forward pressure until the drawer slides completely into the box."
+  }
+}
\ No newline at end of file
diff --git a/closer_drawer/ep037/labels.npz b/closer_drawer/ep037/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..d1dcf8f32c864cc209a639221064a8f99a78bd6c
--- /dev/null
+++ b/closer_drawer/ep037/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d7d506480d0a614a50622d4c19a8e758e774d9acc8bb25b6f57e58a4bb04a812
+size 3805
diff --git a/closer_drawer/ep038/example.md b/closer_drawer/ep038/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..d6f64e8d6d93ab70d409a68e0c13d23aa8763679
--- /dev/null
+++ b/closer_drawer/ep038/example.md
@@ -0,0 +1,21 @@
+# t006__ep038__closer_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Close the drawer
+- **success:** True
+- **control steps:** 311  (6.2s at 50 Hz)
+- **policy inferences:** 53  (replan_steps=6)
+- **video:** `closer_drawer/ep038/rollout.mp4`
+- **features:** `53` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching for drawer |
+| 3.0–6.0 | 152–303 | 25–50 | pushing drawer closed |
+| 6.0–6.2 | 304–310 | 50–51 | holding position near drawer |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/closer_drawer/ep038/labels.json b/closer_drawer/ep038/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..cfa9d4498a4a160a8e54d9b15d2d1774ad5be8b9
--- /dev/null
+++ b/closer_drawer/ep038/labels.json
@@ -0,0 +1,68 @@
+{
+  "rollout_id": "t006__ep038__closer_drawer",
+  "scene_variant": "",
+  "task_id": 6,
+  "task_file": "",
+  "instruction": "Close the drawer",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:31:59+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00990099009901,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for drawer",
+      "description": "reaching for drawer"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00990099009901,
+      "t_end_sec": 6.01980198019802,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "pushing drawer closed",
+      "description": "pushing drawer closed"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01980198019802,
+      "t_end_sec": 6.158415841584159,
+      "control_step_start": 304,
+      "control_step_end": 310,
+      "policy_step_start": 50,
+      "policy_step_end": 51,
+      "phrase": "holding position near drawer",
+      "description": "holding position near drawer"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/closer_drawer/ep038/labels.npz b/closer_drawer/ep038/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..be4dd310bd6e25b4a664c85e7c5307f84e3b6ccd
--- /dev/null
+++ b/closer_drawer/ep038/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:96c4e76b266360920fa9fdeaf2c085552ef0ebfafa25afa90d103ec64e3499ac
+size 3077
diff --git a/closer_drawer/ep039/example.md b/closer_drawer/ep039/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..a38b2246fee09f59b4ca05aaebd08248b9939789
--- /dev/null
+++ b/closer_drawer/ep039/example.md
@@ -0,0 +1,20 @@
+# t006__ep039__closer_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Close the drawer
+- **success:** True
+- **control steps:** 257  (5.1s at 50 Hz)
+- **policy inferences:** 44  (replan_steps=6)
+- **video:** `closer_drawer/ep039/rollout.mp4`
+- **features:** `44` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–149 | 0–24 | reaching and closing drawer |
+| 3.0–5.1 | 150–256 | 25–42 | holding position above drawer |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/closer_drawer/ep039/labels.json b/closer_drawer/ep039/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..885c29616e871012864f0a9bc9b7748de433091a
--- /dev/null
+++ b/closer_drawer/ep039/labels.json
@@ -0,0 +1,57 @@
+{
+  "rollout_id": "t006__ep039__closer_drawer",
+  "scene_variant": "",
+  "task_id": 6,
+  "task_file": "",
+  "instruction": "Close the drawer",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:32:08+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9940119760479043,
+      "control_step_start": 0,
+      "control_step_end": 149,
+      "policy_step_start": 0,
+      "policy_step_end": 24,
+      "phrase": "reaching and closing drawer",
+      "description": "reaching and closing drawer"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9940119760479043,
+      "t_end_sec": 5.129740518962076,
+      "control_step_start": 150,
+      "control_step_end": 256,
+      "policy_step_start": 25,
+      "policy_step_end": 42,
+      "phrase": "holding position above drawer",
+      "description": "holding position above drawer"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/closer_drawer/ep039/labels.npz b/closer_drawer/ep039/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..f06a147911f021e922965c0f6c8eacbef0af9c8f
--- /dev/null
+++ b/closer_drawer/ep039/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0e5231d1504e5199cd2f5cfceec8d337737336452e73ddff74e849f76066605a
+size 3043
diff --git a/closer_drawer/ep040/example.md b/closer_drawer/ep040/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..2270c9e304103991501afb260144c76853d4449c
--- /dev/null
+++ b/closer_drawer/ep040/example.md
@@ -0,0 +1,20 @@
+# t006__ep040__closer_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Close the drawer
+- **success:** True
+- **control steps:** 206  (4.1s at 50 Hz)
+- **policy inferences:** 35  (replan_steps=5)
+- **video:** `closer_drawer/ep040/rollout.mp4`
+- **features:** `35` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–30 | reaching for and pushing drawer |
+| 3.0–4.1 | 151–205 | 30–41 | pushing drawer fully closed |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/closer_drawer/ep040/labels.json b/closer_drawer/ep040/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..e19ea7b94313812068ef4cbe2aa2b2b6fe700f63
--- /dev/null
+++ b/closer_drawer/ep040/labels.json
@@ -0,0 +1,57 @@
+{
+  "rollout_id": "t006__ep040__closer_drawer",
+  "scene_variant": "",
+  "task_id": 6,
+  "task_file": "",
+  "instruction": "Close the drawer",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:32:15+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.001988071570577,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 30,
+      "phrase": "reaching for and pushing drawer",
+      "description": "reaching for and pushing drawer"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.001988071570577,
+      "t_end_sec": 4.095427435387674,
+      "control_step_start": 151,
+      "control_step_end": 205,
+      "policy_step_start": 30,
+      "policy_step_end": 41,
+      "phrase": "pushing drawer fully closed",
+      "description": "pushing drawer fully closed"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/closer_drawer/ep040/labels.npz b/closer_drawer/ep040/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..7e4f3a87f8a9fea67e18ca9c2402b5a37c62e6e0
--- /dev/null
+++ b/closer_drawer/ep040/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a18ffbef3d61ece6fbbb58e804e4def6cc1633aafb59c79c4b0ce0d897cdfaf4
+size 3030
diff --git a/closer_drawer/ep041/example.md b/closer_drawer/ep041/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..a3fbe7cf59d7a89fd7e7597372f14dbb503fa626
--- /dev/null
+++ b/closer_drawer/ep041/example.md
@@ -0,0 +1,21 @@
+# t006__ep041__closer_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Close the drawer
+- **success:** True
+- **control steps:** 332  (6.7s at 50 Hz)
+- **policy inferences:** 57  (replan_steps=6)
+- **video:** `closer_drawer/ep041/rollout.mp4`
+- **features:** `57` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–148 | 0–24 | reaching for the open drawer |
+| 3.0–6.0 | 149–297 | 24–49 | pushing the drawer closed |
+| 6.0–6.7 | 298–331 | 49–55 | idle near closed drawer |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/closer_drawer/ep041/labels.json b/closer_drawer/ep041/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..cd5205d6683b6d4700d66750285d0726625cacf1
--- /dev/null
+++ b/closer_drawer/ep041/labels.json
@@ -0,0 +1,68 @@
+{
+  "rollout_id": "t006__ep041__closer_drawer",
+  "scene_variant": "",
+  "task_id": 6,
+  "task_file": "",
+  "instruction": "Close the drawer",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:32:22+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.991967871485944,
+      "control_step_start": 0,
+      "control_step_end": 148,
+      "policy_step_start": 0,
+      "policy_step_end": 24,
+      "phrase": "reaching for the open drawer",
+      "description": "reaching for the open drawer"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.991967871485944,
+      "t_end_sec": 5.983935742971888,
+      "control_step_start": 149,
+      "control_step_end": 297,
+      "policy_step_start": 24,
+      "policy_step_end": 49,
+      "phrase": "pushing the drawer closed",
+      "description": "pushing the drawer closed"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.983935742971888,
+      "t_end_sec": 6.666666666666667,
+      "control_step_start": 298,
+      "control_step_end": 331,
+      "policy_step_start": 49,
+      "policy_step_end": 55,
+      "phrase": "idle near closed drawer",
+      "description": "idle near closed drawer"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/closer_drawer/ep041/labels.npz b/closer_drawer/ep041/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..380c525badef94bc1e33f113dde72516a9f9ce9a
--- /dev/null
+++ b/closer_drawer/ep041/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:df40bdd11278383681b79515d32de8fde19929ca69d9b1092b9c291628060e72
+size 3081
diff --git a/closer_drawer/ep042/example.md b/closer_drawer/ep042/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..df3f2497a69f3cefde8d58265da7789ff4037f82
--- /dev/null
+++ b/closer_drawer/ep042/example.md
@@ -0,0 +1,32 @@
+# t006__ep042__closer_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Close the drawer
+- **success:** False
+- **control steps:** 299  (5.9s at 51 Hz)
+- **policy inferences:** 52  (replan_steps=6)
+- **video:** `closer_drawer/ep042/rollout.mp4`
+- **features:** `52` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–152 | 0–25 | approaching open drawer handle |
+| 3.0–5.9 | 153–298 | 25–49 | slipping off handle and hovering |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 2.9666011787819255 s (00:03.0)
+- **frame / control_step:** 151 (policy_step=25, chunk_index=1)
+- **coarse onset:** 3.0 s (refined=True)
+- **mode:** open_close_failure
+- **reason:** The robot approached the open drawer handle but slipped off during contact, failing to apply the inward force needed to close the drawer.
+  Summary: slipping off drawer handle; drawer remains open
+- **recovery:** Re-align the gripper with the front face or handle of the open drawer and push straight back until the drawer is fully closed.
+  Summary: re-align gripper with drawer; push drawer closed
+- **prevention:** Position the gripper securely against the front of the drawer handle and maintain steady backward pushing force to slide the drawer into the cabinet.
+  Summary: align with drawer handle; apply firm inward force
+- **justification:** At around 3.0 seconds, the robot arm makes contact with the open drawer handle but slips off without applying horizontal inward pressure. As a result, the drawer stays completely open for the remainder of the execution.
diff --git a/closer_drawer/ep042/labels.json b/closer_drawer/ep042/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..0685c0bbe26743a2e57db692bdadaf43e77ad6b9
--- /dev/null
+++ b/closer_drawer/ep042/labels.json
@@ -0,0 +1,87 @@
+{
+  "rollout_id": "t006__ep042__closer_drawer",
+  "scene_variant": "",
+  "task_id": 6,
+  "task_file": "",
+  "instruction": "Close the drawer",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:32:51+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.005893909626719,
+      "control_step_start": 0,
+      "control_step_end": 152,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "approaching open drawer handle",
+      "description": "approaching open drawer handle"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.005893909626719,
+      "t_end_sec": 5.87426326129666,
+      "control_step_start": 153,
+      "control_step_end": 298,
+      "policy_step_start": 25,
+      "policy_step_end": 49,
+      "phrase": "slipping off handle and hovering",
+      "description": "slipping off handle and hovering"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "open_close_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 2.9666011787819255,
+    "vlm_failure_onset_timestamp": "00:03.0",
+    "vlm_failure_onset_frame": 151,
+    "vlm_failure_onset_step": 151,
+    "vlm_coarse_failure_onset_seconds": 3.0,
+    "vlm_failure_window_start_seconds": 2.5,
+    "vlm_failure_window_end_seconds": 3.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot approached the open drawer handle but slipped off during contact, failing to apply the inward force needed to close the drawer.",
+    "vlm_failure_reason_summary": "slipping off drawer handle; drawer remains open",
+    "vlm_recovery_action": "Re-align the gripper with the front face or handle of the open drawer and push straight back until the drawer is fully closed.",
+    "vlm_recovery_action_summary": "re-align gripper with drawer; push drawer closed",
+    "vlm_prevention_action": "Position the gripper securely against the front of the drawer handle and maintain steady backward pushing force to slide the drawer into the cabinet.",
+    "vlm_prevention_action_summary": "align with drawer handle; apply firm inward force",
+    "vlm_justification": "At around 3.0 seconds, the robot arm makes contact with the open drawer handle but slips off without applying horizontal inward pressure. As a result, the drawer stays completely open for the remainder of the execution.",
+    "vlm_temporal_justification": "At clip time 02:31 (second 151), the robot gripper makes contact with the drawer handle but slips off the upper lip instead of exerting inward force to close the drawer.",
+    "vlm_video_fps": 50.9,
+    "vlm_video_num_frames": 299,
+    "vlm_video_duration_seconds": 5.854616895874264,
+    "vlm_usage": {
+      "input_tokens": 21052,
+      "output_tokens": 363,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"open_close_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 3.0,\n \"failure_window_start_seconds\": 2.5,\n \"failure_window_end_seconds\": 3.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot approached the open drawer handle but slipped off during contact, failing to apply the inward force needed to close the drawer.\",\n \"failure_reason_summary\": \"slipping off drawer handle; drawer remains open\",\n \"recovery_action\": \"Re-align the gripper with the front face or handle of the open drawer and push straight back until the drawer is fully closed.\",\n \"recovery_action_summary\": \"re-align gripper with drawer; push drawer closed\",\n \"prevention_action\": \"Position the gripper securely against the front of the drawer handle and maintain steady backward pushing force to slide the drawer into the cabinet.\",\n \"prevention_action_summary\": \"align with drawer handle; apply firm inward force\",\n \"justification\": \"At around 3.0 seconds, the robot arm makes contact with the open drawer handle but slips off without applying horizontal inward pressure. As a result, the drawer stays completely open for the remainder of the execution.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 151,
+    "failure_sim_step": 151,
+    "failure_policy_step": 25,
+    "failure_chunk_index": 1,
+    "first_post_failure_policy_step": 26,
+    "failure_type": "open_close_failure",
+    "correction_action": "Re-align the gripper with the front face or handle of the open drawer and push straight back until the drawer is fully closed."
+  }
+}
\ No newline at end of file
diff --git a/closer_drawer/ep042/labels.npz b/closer_drawer/ep042/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..ad727fd24e2b2ea89faf6af7fa6cd81b1875bd67
--- /dev/null
+++ b/closer_drawer/ep042/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:397a0fd21bcca8bb5a201a5ea33cfdbf06965e4bfc539baf7814cf18193ab24a
+size 3757
diff --git a/closer_drawer/ep043/example.md b/closer_drawer/ep043/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..fa52642d45124bd154bd25fb3a97af0a66217889
--- /dev/null
+++ b/closer_drawer/ep043/example.md
@@ -0,0 +1,20 @@
+# t006__ep043__closer_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Close the drawer
+- **success:** True
+- **control steps:** 279  (5.5s at 50 Hz)
+- **policy inferences:** 47  (replan_steps=6)
+- **video:** `closer_drawer/ep043/rollout.mp4`
+- **features:** `47` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | reaching for the open drawer |
+| 3.0–5.5 | 151–278 | 25–46 | pushing the drawer closed |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/closer_drawer/ep043/labels.json b/closer_drawer/ep043/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..c013a2cd47bde9e31fc3c83d2824519a7feff0fd
--- /dev/null
+++ b/closer_drawer/ep043/labels.json
@@ -0,0 +1,57 @@
+{
+  "rollout_id": "t006__ep043__closer_drawer",
+  "scene_variant": "",
+  "task_id": 6,
+  "task_file": "",
+  "instruction": "Close the drawer",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:32:59+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.996031746031746,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for the open drawer",
+      "description": "reaching for the open drawer"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.996031746031746,
+      "t_end_sec": 5.535714285714286,
+      "control_step_start": 151,
+      "control_step_end": 278,
+      "policy_step_start": 25,
+      "policy_step_end": 46,
+      "phrase": "pushing the drawer closed",
+      "description": "pushing the drawer closed"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/closer_drawer/ep043/labels.npz b/closer_drawer/ep043/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..1be02ba4875129f2afd150cfe5ff0bc00d608dc8
--- /dev/null
+++ b/closer_drawer/ep043/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5bf6c0da51c143a51459c49c7133462417b36c68c66921bfdf803d16018399eb
+size 3043
diff --git a/closer_drawer/ep044/example.md b/closer_drawer/ep044/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..1a00c721cb029808b93f49196fe8375d6a97110d
--- /dev/null
+++ b/closer_drawer/ep044/example.md
@@ -0,0 +1,20 @@
+# t006__ep044__closer_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Close the drawer
+- **success:** True
+- **control steps:** 209  (4.1s at 51 Hz)
+- **policy inferences:** 37  (replan_steps=5)
+- **video:** `closer_drawer/ep044/rollout.mp4`
+- **features:** `37` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–152 | 0–30 | pushing drawer closed |
+| 3.0–4.1 | 153–208 | 30–41 | idle near closed drawer |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/closer_drawer/ep044/labels.json b/closer_drawer/ep044/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..05718860cefd45bdc3c890e209bad8de867c5bb8
--- /dev/null
+++ b/closer_drawer/ep044/labels.json
@@ -0,0 +1,57 @@
+{
+  "rollout_id": "t006__ep044__closer_drawer",
+  "scene_variant": "",
+  "task_id": 6,
+  "task_file": "",
+  "instruction": "Close the drawer",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:33:06+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.005893909626719,
+      "control_step_start": 0,
+      "control_step_end": 152,
+      "policy_step_start": 0,
+      "policy_step_end": 30,
+      "phrase": "pushing drawer closed",
+      "description": "pushing drawer closed"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.005893909626719,
+      "t_end_sec": 4.106090373280943,
+      "control_step_start": 153,
+      "control_step_end": 208,
+      "policy_step_start": 30,
+      "policy_step_end": 41,
+      "phrase": "idle near closed drawer",
+      "description": "idle near closed drawer"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/closer_drawer/ep044/labels.npz b/closer_drawer/ep044/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..aaf1c63304eab1260624a9f01656495d0648da71
--- /dev/null
+++ b/closer_drawer/ep044/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:66f17065fb76ebc9ba98cbdc15be8f30e15b24cd501fe49f6354c79afbaef63c
+size 3027
diff --git a/closer_drawer/ep045/example.md b/closer_drawer/ep045/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..0dba5f9001b4867bb336c1654b00038b1fe39e8c
--- /dev/null
+++ b/closer_drawer/ep045/example.md
@@ -0,0 +1,32 @@
+# t006__ep045__closer_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Close the drawer
+- **success:** False
+- **control steps:** 267  (5.2s at 51 Hz)
+- **policy inferences:** 46  (replan_steps=6)
+- **video:** `closer_drawer/ep045/rollout.mp4`
+- **features:** `46` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–152 | 0–25 | approaching open drawer |
+| 3.0–5.2 | 153–266 | 25–44 | hovering near open drawer |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 3.732809430255403 s (00:03.7)
+- **frame / control_step:** 190 (policy_step=31, chunk_index=4)
+- **coarse onset:** 3.0 s (refined=True)
+- **mode:** open_close_failure
+- **reason:** The robot reached for the open drawer but failed to make contact with the front panel or handle, leaving the drawer fully open.
+  Summary: missed drawer front; failed to push drawer closed
+- **recovery:** Reposition the gripper directly in front of the open drawer and push straight forward until the drawer shuts completely.
+  Summary: realign gripper; push drawer closed
+- **prevention:** Align the gripper precisely with the front face of the open drawer before moving forward to ensure proper contact and closure.
+  Summary: align with drawer handle; execute straight push
+- **justification:** At around 3.0 seconds, the end-effector approaches the open drawer but fails to contact the front panel to push it shut. The robot hovers/moves without pushing the drawer closed for the rest of the rollout.
diff --git a/closer_drawer/ep045/labels.json b/closer_drawer/ep045/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..b634a653a7ab17823bd04ceda576b0662b99e3b1
--- /dev/null
+++ b/closer_drawer/ep045/labels.json
@@ -0,0 +1,87 @@
+{
+  "rollout_id": "t006__ep045__closer_drawer",
+  "scene_variant": "",
+  "task_id": 6,
+  "task_file": "",
+  "instruction": "Close the drawer",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:33:37+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.005893909626719,
+      "control_step_start": 0,
+      "control_step_end": 152,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "approaching open drawer",
+      "description": "approaching open drawer"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.005893909626719,
+      "t_end_sec": 5.245579567779961,
+      "control_step_start": 153,
+      "control_step_end": 266,
+      "policy_step_start": 25,
+      "policy_step_end": 44,
+      "phrase": "hovering near open drawer",
+      "description": "hovering near open drawer"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "open_close_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 3.732809430255403,
+    "vlm_failure_onset_timestamp": "00:03.7",
+    "vlm_failure_onset_frame": 190,
+    "vlm_failure_onset_step": 190,
+    "vlm_coarse_failure_onset_seconds": 3.0,
+    "vlm_failure_window_start_seconds": 2.5,
+    "vlm_failure_window_end_seconds": 5.23,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot reached for the open drawer but failed to make contact with the front panel or handle, leaving the drawer fully open.",
+    "vlm_failure_reason_summary": "missed drawer front; failed to push drawer closed",
+    "vlm_recovery_action": "Reposition the gripper directly in front of the open drawer and push straight forward until the drawer shuts completely.",
+    "vlm_recovery_action_summary": "realign gripper; push drawer closed",
+    "vlm_prevention_action": "Align the gripper precisely with the front face of the open drawer before moving forward to ensure proper contact and closure.",
+    "vlm_prevention_action_summary": "align with drawer handle; execute straight push",
+    "vlm_justification": "At around 3.0 seconds, the end-effector approaches the open drawer but fails to contact the front panel to push it shut. The robot hovers/moves without pushing the drawer closed for the rest of the rollout.",
+    "vlm_temporal_justification": "At second 190 (frame 190), the robot closes its gripper in open air just in front of the drawer handle without contacting or grasping the drawer, making the open/close failure clearly visible.",
+    "vlm_video_fps": 50.9,
+    "vlm_video_num_frames": 267,
+    "vlm_video_duration_seconds": 5.225933202357564,
+    "vlm_usage": {
+      "input_tokens": 18973,
+      "output_tokens": 359,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"open_close_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 3.0,\n \"failure_window_start_seconds\": 2.5,\n \"failure_window_end_seconds\": 5.23,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot reached for the open drawer but failed to make contact with the front panel or handle, leaving the drawer fully open.\",\n \"failure_reason_summary\": \"missed drawer front; failed to push drawer closed\",\n \"recovery_action\": \"Reposition the gripper directly in front of the open drawer and push straight forward until the drawer shuts completely.\",\n \"recovery_action_summary\": \"realign gripper; push drawer closed\",\n \"prevention_action\": \"Align the gripper precisely with the front face of the open drawer before moving forward to ensure proper contact and closure.\",\n \"prevention_action_summary\": \"align with drawer handle; execute straight push\",\n \"justification\": \"At around 3.0 seconds, the end-effector approaches the open drawer but fails to contact the front panel to push it shut. The robot hovers/moves without pushing the drawer closed for the rest of the rollout.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 190,
+    "failure_sim_step": 190,
+    "failure_policy_step": 31,
+    "failure_chunk_index": 4,
+    "first_post_failure_policy_step": 32,
+    "failure_type": "open_close_failure",
+    "correction_action": "Reposition the gripper directly in front of the open drawer and push straight forward until the drawer shuts completely."
+  }
+}
\ No newline at end of file
diff --git a/closer_drawer/ep045/labels.npz b/closer_drawer/ep045/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..6762cb658b7fa27b3a5c4aac40530468cbed0b14
--- /dev/null
+++ b/closer_drawer/ep045/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c8c74561fdca4176ae0182a60d7fdbaad43b039b324faa6da8e2ddf7c71b8546
+size 3700
diff --git a/closer_drawer/ep046/example.md b/closer_drawer/ep046/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..1b76533852b88a7345bc185d95285b33f2aa023b
--- /dev/null
+++ b/closer_drawer/ep046/example.md
@@ -0,0 +1,33 @@
+# t006__ep046__closer_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Close the drawer
+- **success:** False
+- **control steps:** 304  (6.1s at 50 Hz)
+- **policy inferences:** 52  (replan_steps=6)
+- **video:** `closer_drawer/ep046/rollout.mp4`
+- **features:** `52` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–149 | 0–24 | reaching toward open drawer |
+| 3.0–6.0 | 150–299 | 25–49 | attempting to push drawer shut |
+| 6.0–6.1 | 300–303 | 50–50 | hovering near drawer handle |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 3.992015968063872 s (00:04.0)
+- **frame / control_step:** 200 (policy_step=33, chunk_index=2)
+- **coarse onset:** 4.0 s (refined=False)
+- **mode:** open_close_failure
+- **reason:** The robot descended to the open drawer handle but failed to push or pull it shut, slipping off the edge and leaving the drawer open.
+  Summary: failed to push drawer shut; drawer remains open
+- **recovery:** Re-align the end effector with the front panel of the open drawer and push horizontally until it slides completely shut.
+  Summary: re-align gripper; push drawer closed
+- **prevention:** Approach the drawer at table height and apply direct horizontal pushing force against the front face to fully close it.
+  Summary: approach drawer straight; push firmly inward
+- **justification:** At around 4.0 seconds, the robot attempts to contact the drawer lip, but its contact trajectory fails to push the drawer inside the casing. The drawer remains fully open throughout the rest of the rollout.
diff --git a/closer_drawer/ep046/labels.json b/closer_drawer/ep046/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..52a0b5c48e11a6565d141ea58a8c2cd058b1696a
--- /dev/null
+++ b/closer_drawer/ep046/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t006__ep046__closer_drawer",
+  "scene_variant": "",
+  "task_id": 6,
+  "task_file": "",
+  "instruction": "Close the drawer",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:34:05+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9940119760479043,
+      "control_step_start": 0,
+      "control_step_end": 149,
+      "policy_step_start": 0,
+      "policy_step_end": 24,
+      "phrase": "reaching toward open drawer",
+      "description": "reaching toward open drawer"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9940119760479043,
+      "t_end_sec": 5.9880239520958085,
+      "control_step_start": 150,
+      "control_step_end": 299,
+      "policy_step_start": 25,
+      "policy_step_end": 49,
+      "phrase": "attempting to push drawer shut",
+      "description": "attempting to push drawer shut"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.9880239520958085,
+      "t_end_sec": 6.067864271457085,
+      "control_step_start": 300,
+      "control_step_end": 303,
+      "policy_step_start": 50,
+      "policy_step_end": 50,
+      "phrase": "hovering near drawer handle",
+      "description": "hovering near drawer handle"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "open_close_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 3.992015968063872,
+    "vlm_failure_onset_timestamp": "00:04.0",
+    "vlm_failure_onset_frame": 200,
+    "vlm_failure_onset_step": 200,
+    "vlm_coarse_failure_onset_seconds": 4.0,
+    "vlm_failure_window_start_seconds": 3.5,
+    "vlm_failure_window_end_seconds": 4.5,
+    "vlm_temporal_refined": false,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot descended to the open drawer handle but failed to push or pull it shut, slipping off the edge and leaving the drawer open.",
+    "vlm_failure_reason_summary": "failed to push drawer shut; drawer remains open",
+    "vlm_recovery_action": "Re-align the end effector with the front panel of the open drawer and push horizontally until it slides completely shut.",
+    "vlm_recovery_action_summary": "re-align gripper; push drawer closed",
+    "vlm_prevention_action": "Approach the drawer at table height and apply direct horizontal pushing force against the front face to fully close it.",
+    "vlm_prevention_action_summary": "approach drawer straight; push firmly inward",
+    "vlm_justification": "At around 4.0 seconds, the robot attempts to contact the drawer lip, but its contact trajectory fails to push the drawer inside the casing. The drawer remains fully open throughout the rest of the rollout.",
+    "vlm_temporal_justification": "The robot gripper contacts the open drawer's edge but fails to apply the horizontal force required to push it closed, slipping past the edge and missing the drawer handle entirely.",
+    "vlm_video_fps": 50.1,
+    "vlm_video_num_frames": 304,
+    "vlm_video_duration_seconds": 6.047904191616766,
+    "vlm_usage": {
+      "input_tokens": 18222,
+      "output_tokens": 348,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"open_close_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 4.0,\n \"failure_window_start_seconds\": 3.5,\n \"failure_window_end_seconds\": 4.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot descended to the open drawer handle but failed to push or pull it shut, slipping off the edge and leaving the drawer open.\",\n \"failure_reason_summary\": \"failed to push drawer shut; drawer remains open\",\n \"recovery_action\": \"Re-align the end effector with the front panel of the open drawer and push horizontally until it slides completely shut.\",\n \"recovery_action_summary\": \"re-align gripper; push drawer closed\",\n \"prevention_action\": \"Approach the drawer at table height and apply direct horizontal pushing force against the front face to fully close it.\",\n \"prevention_action_summary\": \"approach drawer straight; push firmly inward\",\n \"justification\": \"At around 4.0 seconds, the robot attempts to contact the drawer lip, but its contact trajectory fails to push the drawer inside the casing. The drawer remains fully open throughout the rest of the rollout.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 200,
+    "failure_sim_step": 200,
+    "failure_policy_step": 33,
+    "failure_chunk_index": 2,
+    "first_post_failure_policy_step": 34,
+    "failure_type": "open_close_failure",
+    "correction_action": "Re-align the end effector with the front panel of the open drawer and push horizontally until it slides completely shut."
+  }
+}
\ No newline at end of file
diff --git a/closer_drawer/ep046/labels.npz b/closer_drawer/ep046/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..6c60767dd4f7cc7b23f9e74a534b9d691cffc999
--- /dev/null
+++ b/closer_drawer/ep046/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e7648b0966adebb14e6882c883e8c62f55722885b17a6b1f8ee66597060cb7bd
+size 3772
diff --git a/closer_drawer/ep047/example.md b/closer_drawer/ep047/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..bdf63d3c9341eb83ed4926e4982f6bd2743d8db8
--- /dev/null
+++ b/closer_drawer/ep047/example.md
@@ -0,0 +1,33 @@
+# t006__ep047__closer_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Close the drawer
+- **success:** False
+- **control steps:** 311  (6.1s at 51 Hz)
+- **policy inferences:** 54  (replan_steps=6)
+- **video:** `closer_drawer/ep047/rollout.mp4`
+- **features:** `54` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–152 | 0–25 | approaching open drawer |
+| 3.0–6.0 | 153–305 | 25–50 | pressing against drawer front |
+| 6.0–6.1 | 306–310 | 51–51 | idle near drawer |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 3.43811394891945 s (00:03.4)
+- **frame / control_step:** 175 (policy_step=29, chunk_index=1)
+- **coarse onset:** 3.0 s (refined=True)
+- **mode:** open_close_failure
+- **reason:** The robot reaches the open drawer but fails to apply forward force to push the drawer shut, leaving it open.
+  Summary: failed to push open drawer shut
+- **recovery:** Reposition the gripper against the front face or handle of the open drawer and push inward until the drawer is fully closed.
+  Summary: reposition gripper; push drawer fully closed
+- **prevention:** Align the end effector horizontally with the front face of the open drawer and push straight back to close it smoothly.
+  Summary: align with drawer front; push horizontally to close
+- **justification:** At around 3.0 seconds, the robot reaches the open drawer but misaligns its push effort, failing to move the drawer into a closed position.
diff --git a/closer_drawer/ep047/labels.json b/closer_drawer/ep047/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..e187781c7ec2eb89d6980aa199ee5854e5537ffe
--- /dev/null
+++ b/closer_drawer/ep047/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t006__ep047__closer_drawer",
+  "scene_variant": "",
+  "task_id": 6,
+  "task_file": "",
+  "instruction": "Close the drawer",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:34:35+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.005893909626719,
+      "control_step_start": 0,
+      "control_step_end": 152,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "approaching open drawer",
+      "description": "approaching open drawer"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.005893909626719,
+      "t_end_sec": 6.011787819253438,
+      "control_step_start": 153,
+      "control_step_end": 305,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "pressing against drawer front",
+      "description": "pressing against drawer front"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.011787819253438,
+      "t_end_sec": 6.110019646365423,
+      "control_step_start": 306,
+      "control_step_end": 310,
+      "policy_step_start": 51,
+      "policy_step_end": 51,
+      "phrase": "idle near drawer",
+      "description": "idle near drawer"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "open_close_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 3.43811394891945,
+    "vlm_failure_onset_timestamp": "00:03.4",
+    "vlm_failure_onset_frame": 175,
+    "vlm_failure_onset_step": 175,
+    "vlm_coarse_failure_onset_seconds": 3.0,
+    "vlm_failure_window_start_seconds": 2.5,
+    "vlm_failure_window_end_seconds": 3.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot reaches the open drawer but fails to apply forward force to push the drawer shut, leaving it open.",
+    "vlm_failure_reason_summary": "failed to push open drawer shut",
+    "vlm_recovery_action": "Reposition the gripper against the front face or handle of the open drawer and push inward until the drawer is fully closed.",
+    "vlm_recovery_action_summary": "reposition gripper; push drawer fully closed",
+    "vlm_prevention_action": "Align the end effector horizontally with the front face of the open drawer and push straight back to close it smoothly.",
+    "vlm_prevention_action_summary": "align with drawer front; push horizontally to close",
+    "vlm_justification": "At around 3.0 seconds, the robot reaches the open drawer but misaligns its push effort, failing to move the drawer into a closed position.",
+    "vlm_temporal_justification": "At 02:55 (frame 175), the robotic gripper completes its movement toward the open drawer's front lip and comes to a complete halt without exerting forward force to push the drawer shut. It remains stationary in this position for the rest of the clip.",
+    "vlm_video_fps": 50.9,
+    "vlm_video_num_frames": 311,
+    "vlm_video_duration_seconds": 6.090373280943026,
+    "vlm_usage": {
+      "input_tokens": 21553,
+      "output_tokens": 355,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"open_close_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 3.0,\n \"failure_window_start_seconds\": 2.5,\n \"failure_window_end_seconds\": 3.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot reaches the open drawer but fails to apply forward force to push the drawer shut, leaving it open.\",\n \"failure_reason_summary\": \"failed to push open drawer shut\",\n \"recovery_action\": \"Reposition the gripper against the front face or handle of the open drawer and push inward until the drawer is fully closed.\",\n \"recovery_action_summary\": \"reposition gripper; push drawer fully closed\",\n \"prevention_action\": \"Align the end effector horizontally with the front face of the open drawer and push straight back to close it smoothly.\",\n \"prevention_action_summary\": \"align with drawer front; push horizontally to close\",\n \"justification\": \"At around 3.0 seconds, the robot reaches the open drawer but misaligns its push effort, failing to move the drawer into a closed position.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 175,
+    "failure_sim_step": 175,
+    "failure_policy_step": 29,
+    "failure_chunk_index": 1,
+    "first_post_failure_policy_step": 30,
+    "failure_type": "open_close_failure",
+    "correction_action": "Reposition the gripper against the front face or handle of the open drawer and push inward until the drawer is fully closed."
+  }
+}
\ No newline at end of file
diff --git a/closer_drawer/ep047/labels.npz b/closer_drawer/ep047/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..b92baa47b508c4d69253f0e754cda1cae41284c5
--- /dev/null
+++ b/closer_drawer/ep047/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ec2a925e227b0eeb266029c841bb84fc36a541f98b0c8b0ee56837f2a85113a2
+size 3736
diff --git a/closer_drawer/ep048/example.md b/closer_drawer/ep048/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..4e66060e640c013e794711a677710cd554aafa38
--- /dev/null
+++ b/closer_drawer/ep048/example.md
@@ -0,0 +1,32 @@
+# t006__ep048__closer_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Close the drawer
+- **success:** False
+- **control steps:** 286  (5.7s at 50 Hz)
+- **policy inferences:** 49  (replan_steps=6)
+- **video:** `closer_drawer/ep048/rollout.mp4`
+- **features:** `49` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–149 | 0–24 | moving arm toward open drawer |
+| 3.0–5.7 | 150–285 | 25–47 | slipping off drawer handle |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 3.852295409181637 s (00:03.9)
+- **frame / control_step:** 193 (policy_step=32, chunk_index=1)
+- **coarse onset:** 4.0 s (refined=True)
+- **mode:** open_close_failure
+- **reason:** The robot end-effector attempted to push the drawer shut, but slipped off the front face/handle without exerting inward force, leaving the drawer completely open.
+  Summary: failed to push drawer shut; slipped off handle
+- **recovery:** Reposition the gripper directly in front of the drawer handle, make firm contact, and push the drawer smoothly into its housing until closed.
+  Summary: re-align with handle; push drawer closed
+- **prevention:** Align the gripper squarely with the drawer handle before contact and apply steady inward force to push the drawer completely shut.
+  Summary: align centered with handle; maintain inward pressure
+- **justification:** At around 4.0 seconds, the gripper reaches the drawer but fails to grab or push the handle correctly, slipping to the side and leaving the drawer open at episode termination.
diff --git a/closer_drawer/ep048/labels.json b/closer_drawer/ep048/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..50c4e289f2e8efa0d5a66f4333486ada34becb6f
--- /dev/null
+++ b/closer_drawer/ep048/labels.json
@@ -0,0 +1,87 @@
+{
+  "rollout_id": "t006__ep048__closer_drawer",
+  "scene_variant": "",
+  "task_id": 6,
+  "task_file": "",
+  "instruction": "Close the drawer",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:35:04+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9940119760479043,
+      "control_step_start": 0,
+      "control_step_end": 149,
+      "policy_step_start": 0,
+      "policy_step_end": 24,
+      "phrase": "moving arm toward open drawer",
+      "description": "moving arm toward open drawer"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9940119760479043,
+      "t_end_sec": 5.708582834331337,
+      "control_step_start": 150,
+      "control_step_end": 285,
+      "policy_step_start": 25,
+      "policy_step_end": 47,
+      "phrase": "slipping off drawer handle",
+      "description": "slipping off drawer handle"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "open_close_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 3.852295409181637,
+    "vlm_failure_onset_timestamp": "00:03.9",
+    "vlm_failure_onset_frame": 193,
+    "vlm_failure_onset_step": 193,
+    "vlm_coarse_failure_onset_seconds": 4.0,
+    "vlm_failure_window_start_seconds": 3.5,
+    "vlm_failure_window_end_seconds": 5.69,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot end-effector attempted to push the drawer shut, but slipped off the front face/handle without exerting inward force, leaving the drawer completely open.",
+    "vlm_failure_reason_summary": "failed to push drawer shut; slipped off handle",
+    "vlm_recovery_action": "Reposition the gripper directly in front of the drawer handle, make firm contact, and push the drawer smoothly into its housing until closed.",
+    "vlm_recovery_action_summary": "re-align with handle; push drawer closed",
+    "vlm_prevention_action": "Align the gripper squarely with the drawer handle before contact and apply steady inward force to push the drawer completely shut.",
+    "vlm_prevention_action_summary": "align centered with handle; maintain inward pressure",
+    "vlm_justification": "At around 4.0 seconds, the gripper reaches the drawer but fails to grab or push the handle correctly, slipping to the side and leaving the drawer open at episode termination.",
+    "vlm_temporal_justification": "At timestamp 02:23, the robot end-effector attempts to make contact with the drawer to push it closed, but instead slips off the edge/face without pushing the drawer in.",
+    "vlm_video_fps": 50.1,
+    "vlm_video_num_frames": 286,
+    "vlm_video_duration_seconds": 5.688622754491018,
+    "vlm_usage": {
+      "input_tokens": 17092,
+      "output_tokens": 358,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"open_close_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 4.0,\n \"failure_window_start_seconds\": 3.5,\n \"failure_window_end_seconds\": 5.69,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot end-effector attempted to push the drawer shut, but slipped off the front face/handle without exerting inward force, leaving the drawer completely open.\",\n \"failure_reason_summary\": \"failed to push drawer shut; slipped off handle\",\n \"recovery_action\": \"Reposition the gripper directly in front of the drawer handle, make firm contact, and push the drawer smoothly into its housing until closed.\",\n \"recovery_action_summary\": \"re-align with handle; push drawer closed\",\n \"prevention_action\": \"Align the gripper squarely with the drawer handle before contact and apply steady inward force to push the drawer completely shut.\",\n \"prevention_action_summary\": \"align centered with handle; maintain inward pressure\",\n \"justification\": \"At around 4.0 seconds, the gripper reaches the drawer but fails to grab or push the handle correctly, slipping to the side and leaving the drawer open at episode termination.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 193,
+    "failure_sim_step": 193,
+    "failure_policy_step": 32,
+    "failure_chunk_index": 1,
+    "first_post_failure_policy_step": 33,
+    "failure_type": "open_close_failure",
+    "correction_action": "Reposition the gripper directly in front of the drawer handle, make firm contact, and push the drawer smoothly into its housing until closed."
+  }
+}
\ No newline at end of file
diff --git a/closer_drawer/ep048/labels.npz b/closer_drawer/ep048/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..3959ad2a9cdb43ddc0a0eef33836311194e782d5
--- /dev/null
+++ b/closer_drawer/ep048/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2dd0ffed8f67cf1ea1e6074b6695c660b621986a46bbaf10a40e44376350abe6
+size 3783
diff --git a/closer_drawer/ep049/example.md b/closer_drawer/ep049/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..f4a2cb9c62e8cd5edfb9d2eca6de09676da6c7a5
--- /dev/null
+++ b/closer_drawer/ep049/example.md
@@ -0,0 +1,21 @@
+# t006__ep049__closer_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Close the drawer
+- **success:** True
+- **control steps:** 318  (6.3s at 51 Hz)
+- **policy inferences:** 55  (replan_steps=6)
+- **video:** `closer_drawer/ep049/rollout.mp4`
+- **features:** `55` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching towards open drawer |
+| 3.0–6.0 | 152–303 | 25–50 | pushing the drawer closed |
+| 6.0–6.3 | 304–317 | 50–52 | holding position near drawer |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/closer_drawer/ep049/labels.json b/closer_drawer/ep049/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..f5dc0bb6ab2251a85cd3d0370840ab0e9f9f4484
--- /dev/null
+++ b/closer_drawer/ep049/labels.json
@@ -0,0 +1,68 @@
+{
+  "rollout_id": "t006__ep049__closer_drawer",
+  "scene_variant": "",
+  "task_id": 6,
+  "task_file": "",
+  "instruction": "Close the drawer",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:35:12+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0039525691699605,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching towards open drawer",
+      "description": "reaching towards open drawer"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0039525691699605,
+      "t_end_sec": 6.007905138339921,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "pushing the drawer closed",
+      "description": "pushing the drawer closed"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.007905138339921,
+      "t_end_sec": 6.284584980237154,
+      "control_step_start": 304,
+      "control_step_end": 317,
+      "policy_step_start": 50,
+      "policy_step_end": 52,
+      "phrase": "holding position near drawer",
+      "description": "holding position near drawer"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/closer_drawer/ep049/labels.npz b/closer_drawer/ep049/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..e519634feace72340b1fa15a11183d87a750d19a
--- /dev/null
+++ b/closer_drawer/ep049/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3535283315c3f3cc134a7138d0b039e0069b4a767e2b0d6f8f18caea8110eb63
+size 3096
diff --git a/eggplant/ep000/example.md b/eggplant/ep000/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..3a67c71d34a13718f5ad27e0070c37e3649c0a9c
--- /dev/null
+++ b/eggplant/ep000/example.md
@@ -0,0 +1,33 @@
+# t004__ep000__eggplant
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the eggplant off the plate
+- **success:** False
+- **control steps:** 411  (8.2s at 50 Hz)
+- **policy inferences:** 72  (replan_steps=5)
+- **video:** `eggplant/ep000/rollout.mp4`
+- **features:** `72` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–30 | lowering gripper toward eggplant |
+| 3.0–6.0 | 151–301 | 30–60 | closing gripper and lifting empty |
+| 6.0–8.2 | 302–410 | 60–82 | attempting second grasp and raising |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 3.585657370517928 s (00:03.6)
+- **frame / control_step:** 180 (policy_step=36, chunk_index=0)
+- **coarse onset:** 3.0 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot positions its end-effector over the eggplant and closes its gripper, but fails to achieve a firm grip, lifting up empty.
+  Summary: improper gripper alignment; missed grasp on eggplant
+- **recovery:** Realign the gripper directly over the center of the eggplant, lower the end-effector, and close the gripper securely before lifting.
+  Summary: realign gripper over eggplant; execute secure grasp; lift object
+- **prevention:** Position the gripper precisely around the thickest part of the eggplant and ensure full closure before initiating the lift motion.
+  Summary: center gripper on eggplant; close fingers securely
+- **justification:** None
diff --git a/eggplant/ep000/labels.json b/eggplant/ep000/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..71e420e4bc5ce395c7a22afb5e56439652860042
--- /dev/null
+++ b/eggplant/ep000/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t004__ep000__eggplant",
+  "scene_variant": "",
+  "task_id": 4,
+  "task_file": "",
+  "instruction": "Take the eggplant off the plate",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:35:48+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00796812749004,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 30,
+      "phrase": "lowering gripper toward eggplant",
+      "description": "lowering gripper toward eggplant"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00796812749004,
+      "t_end_sec": 6.01593625498008,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 30,
+      "policy_step_end": 60,
+      "phrase": "closing gripper and lifting empty",
+      "description": "closing gripper and lifting empty"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01593625498008,
+      "t_end_sec": 8.187250996015935,
+      "control_step_start": 302,
+      "control_step_end": 410,
+      "policy_step_start": 60,
+      "policy_step_end": 82,
+      "phrase": "attempting second grasp and raising",
+      "description": "attempting second grasp and raising"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 3.585657370517928,
+    "vlm_failure_onset_timestamp": "00:03.6",
+    "vlm_failure_onset_frame": 180,
+    "vlm_failure_onset_step": 180,
+    "vlm_coarse_failure_onset_seconds": 3.0,
+    "vlm_failure_window_start_seconds": 2.5,
+    "vlm_failure_window_end_seconds": 3.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot positions its end-effector over the eggplant and closes its gripper, but fails to achieve a firm grip, lifting up empty.",
+    "vlm_failure_reason_summary": "improper gripper alignment; missed grasp on eggplant",
+    "vlm_recovery_action": "Realign the gripper directly over the center of the eggplant, lower the end-effector, and close the gripper securely before lifting.",
+    "vlm_recovery_action_summary": "realign gripper over eggplant; execute secure grasp; lift object",
+    "vlm_prevention_action": "Position the gripper precisely around the thickest part of the eggplant and ensure full closure before initiating the lift motion.",
+    "vlm_prevention_action_summary": "center gripper on eggplant; close fingers securely",
+    "vlm_justification": null,
+    "vlm_temporal_justification": "At timestamp 03:00 (corresponding to frame 180 of the original rollout), the robot closes its end-effector fingers completely together without grabbing the eggplant, visibly failing the grasp attempt.",
+    "vlm_video_fps": 50.2,
+    "vlm_video_num_frames": 411,
+    "vlm_video_duration_seconds": 8.167330677290837,
+    "vlm_usage": {
+      "input_tokens": 21442,
+      "output_tokens": 313,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 3.0,\n \"failure_window_start_seconds\": 2.5,\n \"failure_window_end_seconds\": 3.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot positions its end-effector over the eggplant and closes its gripper, but fails to achieve a firm grip, lifting up empty.\",\n \"failure_reason_summary\": \"improper gripper alignment; missed grasp on eggplant\",\n \"recovery_action\": \"Realign the gripper directly over the center of the eggplant, lower the end-effector, and close the gripper securely before lifting.\",\n \"recovery_action_summary\": \"realign gripper over eggplant; execute secure grasp; lift object\",\n \"prevention_action\": \"Position the gripper precisely around the thickest part of the eggplant and ensure full closure before initiating the lift motion.\",\n \"prevention_action_summary\": \"center gripper on eggplant; close fingers securely\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 180,
+    "failure_sim_step": 180,
+    "failure_policy_step": 36,
+    "failure_chunk_index": 0,
+    "first_post_failure_policy_step": 37,
+    "failure_type": "grasp_failure",
+    "correction_action": "Realign the gripper directly over the center of the eggplant, lower the end-effector, and close the gripper securely before lifting."
+  }
+}
\ No newline at end of file
diff --git a/eggplant/ep000/labels.npz b/eggplant/ep000/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..f187f4dcaf336f931b768debe83e997d17c66ae0
--- /dev/null
+++ b/eggplant/ep000/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b8491eff2db5a9272f90457dc39b3c87e1b6abc7819557e08b1b57422ccb0b64
+size 3812
diff --git a/eggplant/ep001/example.md b/eggplant/ep001/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..b2593eab7728740de230628d4855b2070ea9745c
--- /dev/null
+++ b/eggplant/ep001/example.md
@@ -0,0 +1,21 @@
+# t004__ep001__eggplant
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the eggplant off the plate
+- **success:** True
+- **control steps:** 379  (7.5s at 50 Hz)
+- **policy inferences:** 66  (replan_steps=6)
+- **video:** `eggplant/ep001/rollout.mp4`
+- **features:** `66` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | reaching and grasping eggplant |
+| 3.0–6.0 | 151–301 | 25–50 | lifting and placing eggplant on table |
+| 6.0–7.5 | 302–378 | 50–63 | retracting arm away |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/eggplant/ep001/labels.json b/eggplant/ep001/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..af307dcf6b6a74a3c2d438626db3c34177be4a0b
--- /dev/null
+++ b/eggplant/ep001/labels.json
@@ -0,0 +1,68 @@
+{
+  "rollout_id": "t004__ep001__eggplant",
+  "scene_variant": "",
+  "task_id": 4,
+  "task_file": "",
+  "instruction": "Take the eggplant off the plate",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:35:58+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00796812749004,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching and grasping eggplant",
+      "description": "reaching and grasping eggplant"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00796812749004,
+      "t_end_sec": 6.01593625498008,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "lifting and placing eggplant on table",
+      "description": "lifting and placing eggplant on table"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01593625498008,
+      "t_end_sec": 7.5498007968127485,
+      "control_step_start": 302,
+      "control_step_end": 378,
+      "policy_step_start": 50,
+      "policy_step_end": 63,
+      "phrase": "retracting arm away",
+      "description": "retracting arm away"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/eggplant/ep001/labels.npz b/eggplant/ep001/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..de9ba68d3da2b878792e705641d00bbe0d62e8c5
--- /dev/null
+++ b/eggplant/ep001/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1a5a7e64931b2a453f33759631fd8f3fe2152259c7fadc03648c6bdc336fa8d0
+size 3089
diff --git a/eggplant/ep002/example.md b/eggplant/ep002/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..86194fb60bcab92ce0638199df170f768bacddc6
--- /dev/null
+++ b/eggplant/ep002/example.md
@@ -0,0 +1,21 @@
+# t004__ep002__eggplant
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the eggplant off the plate
+- **success:** True
+- **control steps:** 423  (8.3s at 51 Hz)
+- **policy inferences:** 72  (replan_steps=6)
+- **video:** `eggplant/ep002/rollout.mp4`
+- **features:** `72` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–152 | 0–25 | reaching for and grasping eggplant |
+| 3.0–6.0 | 153–305 | 25–50 | lifting and moving eggplant |
+| 6.0–8.3 | 306–422 | 51–70 | releasing eggplant and retracting arm |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/eggplant/ep002/labels.json b/eggplant/ep002/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..7fdf1b07c089f20658e8175ec79bf6b8479d8780
--- /dev/null
+++ b/eggplant/ep002/labels.json
@@ -0,0 +1,68 @@
+{
+  "rollout_id": "t004__ep002__eggplant",
+  "scene_variant": "",
+  "task_id": 4,
+  "task_file": "",
+  "instruction": "Take the eggplant off the plate",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:36:07+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0,
+      "control_step_start": 0,
+      "control_step_end": 152,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for and grasping eggplant",
+      "description": "reaching for and grasping eggplant"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0,
+      "t_end_sec": 6.0,
+      "control_step_start": 153,
+      "control_step_end": 305,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "lifting and moving eggplant",
+      "description": "lifting and moving eggplant"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.0,
+      "t_end_sec": 8.294117647058824,
+      "control_step_start": 306,
+      "control_step_end": 422,
+      "policy_step_start": 51,
+      "policy_step_end": 70,
+      "phrase": "releasing eggplant and retracting arm",
+      "description": "releasing eggplant and retracting arm"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/eggplant/ep002/labels.npz b/eggplant/ep002/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..08891d9d8a5ac5e0e1a5cccbfdd14edb464d9e0f
--- /dev/null
+++ b/eggplant/ep002/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1a433929299331e387e36f5cbf7379a7073d173930be33020c598645dd4a6b37
+size 3092
diff --git a/eggplant/ep003/example.md b/eggplant/ep003/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..8bc3e23013baf497d729bf3bb6614c43e7dbf1e8
--- /dev/null
+++ b/eggplant/ep003/example.md
@@ -0,0 +1,33 @@
+# t004__ep003__eggplant
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the eggplant off the plate
+- **success:** False
+- **control steps:** 320  (6.3s at 51 Hz)
+- **policy inferences:** 55  (replan_steps=6)
+- **video:** `eggplant/ep003/rollout.mp4`
+- **features:** `55` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–152 | 0–25 | lowering gripper toward eggplant |
+| 3.0–6.0 | 153–305 | 25–50 | closing gripper and lifting arm |
+| 6.0–6.3 | 306–319 | 51–53 | hovering above plate |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 3.542074363992172 s (00:03.5)
+- **frame / control_step:** 181 (policy_step=30, chunk_index=1)
+- **coarse onset:** 3.2 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot lowered its end-effector to grasp the eggplant, but failed to secure a firm grip, closing its gripper completely without holding the object before lifting away.
+  Summary: failed grasp on eggplant; gripper closed empty
+- **recovery:** Re-align the gripper directly over the eggplant on the plate, lower the arm, close the fingers firmly around the eggplant, and lift it clear of the plate.
+  Summary: re-align over eggplant; close gripper securely; lift eggplant
+- **prevention:** Position the gripper properly around the middle section of the eggplant and ensure a secure pinch grip before initiating the lift.
+  Summary: center gripper over eggplant; grasp firmly before lifting
+- **justification:** Around 3.2 seconds into the video, the robot closes its gripper on the eggplant, but the object slips out as the fingers close fully empty. The robot then lifts up at 4.0s without the eggplant.
diff --git a/eggplant/ep003/labels.json b/eggplant/ep003/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..bd1f966342c80312163ca7a46b77a2f99eae9df7
--- /dev/null
+++ b/eggplant/ep003/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t004__ep003__eggplant",
+  "scene_variant": "",
+  "task_id": 4,
+  "task_file": "",
+  "instruction": "Take the eggplant off the plate",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:36:41+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.99412915851272,
+      "control_step_start": 0,
+      "control_step_end": 152,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "lowering gripper toward eggplant",
+      "description": "lowering gripper toward eggplant"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.99412915851272,
+      "t_end_sec": 5.98825831702544,
+      "control_step_start": 153,
+      "control_step_end": 305,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "closing gripper and lifting arm",
+      "description": "closing gripper and lifting arm"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.98825831702544,
+      "t_end_sec": 6.262230919765166,
+      "control_step_start": 306,
+      "control_step_end": 319,
+      "policy_step_start": 51,
+      "policy_step_end": 53,
+      "phrase": "hovering above plate",
+      "description": "hovering above plate"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 3.542074363992172,
+    "vlm_failure_onset_timestamp": "00:03.5",
+    "vlm_failure_onset_frame": 181,
+    "vlm_failure_onset_step": 181,
+    "vlm_coarse_failure_onset_seconds": 3.2,
+    "vlm_failure_window_start_seconds": 2.5,
+    "vlm_failure_window_end_seconds": 4.0,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot lowered its end-effector to grasp the eggplant, but failed to secure a firm grip, closing its gripper completely without holding the object before lifting away.",
+    "vlm_failure_reason_summary": "failed grasp on eggplant; gripper closed empty",
+    "vlm_recovery_action": "Re-align the gripper directly over the eggplant on the plate, lower the arm, close the fingers firmly around the eggplant, and lift it clear of the plate.",
+    "vlm_recovery_action_summary": "re-align over eggplant; close gripper securely; lift eggplant",
+    "vlm_prevention_action": "Position the gripper properly around the middle section of the eggplant and ensure a secure pinch grip before initiating the lift.",
+    "vlm_prevention_action_summary": "center gripper over eggplant; grasp firmly before lifting",
+    "vlm_justification": "Around 3.2 seconds into the video, the robot closes its gripper on the eggplant, but the object slips out as the fingers close fully empty. The robot then lifts up at 4.0s without the eggplant.",
+    "vlm_temporal_justification": "At timestamp 02:50 in the slowed clip, the robot's gripper fingers close completely together without holding the eggplant, making the grasp failure visibly apparent as the robot fails to secure the object.",
+    "vlm_video_fps": 51.1,
+    "vlm_video_num_frames": 320,
+    "vlm_video_duration_seconds": 6.2426614481409,
+    "vlm_usage": {
+      "input_tokens": 21576,
+      "output_tokens": 377,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 3.2,\n \"failure_window_start_seconds\": 2.5,\n \"failure_window_end_seconds\": 4.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot lowered its end-effector to grasp the eggplant, but failed to secure a firm grip, closing its gripper completely without holding the object before lifting away.\",\n \"failure_reason_summary\": \"failed grasp on eggplant; gripper closed empty\",\n \"recovery_action\": \"Re-align the gripper directly over the eggplant on the plate, lower the arm, close the fingers firmly around the eggplant, and lift it clear of the plate.\",\n \"recovery_action_summary\": \"re-align over eggplant; close gripper securely; lift eggplant\",\n \"prevention_action\": \"Position the gripper properly around the middle section of the eggplant and ensure a secure pinch grip before initiating the lift.\",\n \"prevention_action_summary\": \"center gripper over eggplant; grasp firmly before lifting\",\n \"justification\": \"Around 3.2 seconds into the video, the robot closes its gripper on the eggplant, but the object slips out as the fingers close fully empty. The robot then lifts up at 4.0s without the eggplant.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 181,
+    "failure_sim_step": 181,
+    "failure_policy_step": 30,
+    "failure_chunk_index": 1,
+    "first_post_failure_policy_step": 31,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-align the gripper directly over the eggplant on the plate, lower the arm, close the fingers firmly around the eggplant, and lift it clear of the plate."
+  }
+}
\ No newline at end of file
diff --git a/eggplant/ep003/labels.npz b/eggplant/ep003/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..83ce08423aed2bd4b30cd50476dbc837ce9ac074
--- /dev/null
+++ b/eggplant/ep003/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6ee0371d7ed5652a3ee1bdafede6ed8f34777002a377a6d775e8df43aaaa8dc5
+size 3838
diff --git a/eggplant/ep004/example.md b/eggplant/ep004/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..02c100e0ae3bbb93ff722c925ac9817b71248453
--- /dev/null
+++ b/eggplant/ep004/example.md
@@ -0,0 +1,21 @@
+# t004__ep004__eggplant
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the eggplant off the plate
+- **success:** True
+- **control steps:** 420  (8.2s at 51 Hz)
+- **policy inferences:** 72  (replan_steps=6)
+- **video:** `eggplant/ep004/rollout.mp4`
+- **features:** `72` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–153 | 0–25 | reaching and grasping eggplant |
+| 3.0–6.0 | 154–307 | 25–51 | lifting and placing eggplant on table |
+| 6.0–8.2 | 308–419 | 51–69 | releasing eggplant and retracting |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/eggplant/ep004/labels.json b/eggplant/ep004/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..61749b6575d6645116ab89bb12c82410c24dffe8
--- /dev/null
+++ b/eggplant/ep004/labels.json
@@ -0,0 +1,68 @@
+{
+  "rollout_id": "t004__ep004__eggplant",
+  "scene_variant": "",
+  "task_id": 4,
+  "task_file": "",
+  "instruction": "Take the eggplant off the plate",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:36:50+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0078125,
+      "control_step_start": 0,
+      "control_step_end": 153,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching and grasping eggplant",
+      "description": "reaching and grasping eggplant"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0078125,
+      "t_end_sec": 6.015625,
+      "control_step_start": 154,
+      "control_step_end": 307,
+      "policy_step_start": 25,
+      "policy_step_end": 51,
+      "phrase": "lifting and placing eggplant on table",
+      "description": "lifting and placing eggplant on table"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.015625,
+      "t_end_sec": 8.203125,
+      "control_step_start": 308,
+      "control_step_end": 419,
+      "policy_step_start": 51,
+      "policy_step_end": 69,
+      "phrase": "releasing eggplant and retracting",
+      "description": "releasing eggplant and retracting"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/eggplant/ep004/labels.npz b/eggplant/ep004/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..742dd0db16cfbcec9581187b2395ff0e2913a1e0
--- /dev/null
+++ b/eggplant/ep004/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:74c7366fc7319a611f9265e79243d8e4a6df9f4cb30f71d2000a0f55ea603eef
+size 3089
diff --git a/eggplant/ep005/example.md b/eggplant/ep005/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..f14236387bc9e2ebceac4f31cb7c00fd2f66dbf2
--- /dev/null
+++ b/eggplant/ep005/example.md
@@ -0,0 +1,21 @@
+# t004__ep005__eggplant
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the eggplant off the plate
+- **success:** True
+- **control steps:** 347  (6.9s at 50 Hz)
+- **policy inferences:** 61  (replan_steps=6)
+- **video:** `eggplant/ep005/rollout.mp4`
+- **features:** `61` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | reaching for and grasping eggplant |
+| 3.0–6.0 | 151–301 | 25–50 | lifting and moving eggplant off plate |
+| 6.0–6.9 | 302–346 | 50–57 | releasing eggplant and retracting arm |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/eggplant/ep005/labels.json b/eggplant/ep005/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..cdaf758fc85be14075cd6cfcc7f8e4114e4f99f4
--- /dev/null
+++ b/eggplant/ep005/labels.json
@@ -0,0 +1,68 @@
+{
+  "rollout_id": "t004__ep005__eggplant",
+  "scene_variant": "",
+  "task_id": 4,
+  "task_file": "",
+  "instruction": "Take the eggplant off the plate",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:36:59+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.996031746031746,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for and grasping eggplant",
+      "description": "reaching for and grasping eggplant"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.996031746031746,
+      "t_end_sec": 5.992063492063492,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "lifting and moving eggplant off plate",
+      "description": "lifting and moving eggplant off plate"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.992063492063492,
+      "t_end_sec": 6.884920634920635,
+      "control_step_start": 302,
+      "control_step_end": 346,
+      "policy_step_start": 50,
+      "policy_step_end": 57,
+      "phrase": "releasing eggplant and retracting arm",
+      "description": "releasing eggplant and retracting arm"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/eggplant/ep005/labels.npz b/eggplant/ep005/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..018531ff040de457e403f2b5cb24f88759814c10
--- /dev/null
+++ b/eggplant/ep005/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5bb68646512ad546f304b71a288c07d1dbcb0066cc176280234f7ee908d5b179
+size 3103
diff --git a/eggplant/ep006/example.md b/eggplant/ep006/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..cd4a58d08aeaeb3343dcdaeb7887a62432f28223
--- /dev/null
+++ b/eggplant/ep006/example.md
@@ -0,0 +1,21 @@
+# t004__ep006__eggplant
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the eggplant off the plate
+- **success:** True
+- **control steps:** 400  (7.9s at 50 Hz)
+- **policy inferences:** 69  (replan_steps=6)
+- **video:** `eggplant/ep006/rollout.mp4`
+- **features:** `69` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | reaching for and grasping eggplant |
+| 3.0–6.0 | 151–301 | 25–50 | lifting and moving eggplant |
+| 6.0–7.9 | 302–399 | 50–66 | releasing eggplant and retracting |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/eggplant/ep006/labels.json b/eggplant/ep006/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..392a3411e83a8cbf69cf12bd1ac8118b41725dfa
--- /dev/null
+++ b/eggplant/ep006/labels.json
@@ -0,0 +1,68 @@
+{
+  "rollout_id": "t004__ep006__eggplant",
+  "scene_variant": "",
+  "task_id": 4,
+  "task_file": "",
+  "instruction": "Take the eggplant off the plate",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:37:09+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.996031746031746,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for and grasping eggplant",
+      "description": "reaching for and grasping eggplant"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.996031746031746,
+      "t_end_sec": 5.992063492063492,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "lifting and moving eggplant",
+      "description": "lifting and moving eggplant"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.992063492063492,
+      "t_end_sec": 7.936507936507937,
+      "control_step_start": 302,
+      "control_step_end": 399,
+      "policy_step_start": 50,
+      "policy_step_end": 66,
+      "phrase": "releasing eggplant and retracting",
+      "description": "releasing eggplant and retracting"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/eggplant/ep006/labels.npz b/eggplant/ep006/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..bdde355c4761965890572cfd06e8c1e9c824005b
--- /dev/null
+++ b/eggplant/ep006/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a7f9a1cbce6b70aad9c2233e2a9cf294ed458391684f8f76924f65f346d91437
+size 3086
diff --git a/eggplant/ep007/example.md b/eggplant/ep007/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..042ddcf816b4f5d925617fc9475228170e4650c5
--- /dev/null
+++ b/eggplant/ep007/example.md
@@ -0,0 +1,33 @@
+# t004__ep007__eggplant
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the eggplant off the plate
+- **success:** False
+- **control steps:** 340  (6.7s at 50 Hz)
+- **policy inferences:** 60  (replan_steps=6)
+- **video:** `eggplant/ep007/rollout.mp4`
+- **features:** `60` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | approaching eggplant on plate |
+| 3.0–6.0 | 151–301 | 25–50 | attempting grasp; lifting empty gripper |
+| 6.0–6.7 | 302–339 | 50–56 | hovering empty gripper above table |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 4.642857142857143 s (00:04.6)
+- **frame / control_step:** 234 (policy_step=39, chunk_index=0)
+- **coarse onset:** 3.8 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the eggplant on the plate, but the gripper misaligned with the object, causing the eggplant to slip out and flip over without being securely grasped.
+  Summary: failed grasp attempt; eggplant slips from gripper
+- **recovery:** Re-align the end-effector over the eggplant, lower the open gripper onto the main body of the eggplant, and close the fingers securely before lifting it clear of the plate.
+  Summary: re-align gripper; grasp eggplant securely; lift off plate
+- **prevention:** Center and align the gripper precisely over the eggplant's main body before closing the fingers to ensure a stable grasp.
+  Summary: align gripper with eggplant; execute secure grasp
+- **justification:** Around 3.8 seconds into the video, the robot closes its gripper on the eggplant, but fails to achieve a stable hold, causing the eggplant to slip out and remain on the plate while the robot lifts away empty.
diff --git a/eggplant/ep007/labels.json b/eggplant/ep007/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..84447734607871bb4519104b0fc35a9badf59f1c
--- /dev/null
+++ b/eggplant/ep007/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t004__ep007__eggplant",
+  "scene_variant": "",
+  "task_id": 4,
+  "task_file": "",
+  "instruction": "Take the eggplant off the plate",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:37:40+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.996031746031746,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "approaching eggplant on plate",
+      "description": "approaching eggplant on plate"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.996031746031746,
+      "t_end_sec": 5.992063492063492,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "attempting grasp; lifting empty gripper",
+      "description": "attempting grasp; lifting empty gripper"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.992063492063492,
+      "t_end_sec": 6.746031746031746,
+      "control_step_start": 302,
+      "control_step_end": 339,
+      "policy_step_start": 50,
+      "policy_step_end": 56,
+      "phrase": "hovering empty gripper above table",
+      "description": "hovering empty gripper above table"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 4.642857142857143,
+    "vlm_failure_onset_timestamp": "00:04.6",
+    "vlm_failure_onset_frame": 234,
+    "vlm_failure_onset_step": 234,
+    "vlm_coarse_failure_onset_seconds": 3.8,
+    "vlm_failure_window_start_seconds": 3.3,
+    "vlm_failure_window_end_seconds": 4.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the eggplant on the plate, but the gripper misaligned with the object, causing the eggplant to slip out and flip over without being securely grasped.",
+    "vlm_failure_reason_summary": "failed grasp attempt; eggplant slips from gripper",
+    "vlm_recovery_action": "Re-align the end-effector over the eggplant, lower the open gripper onto the main body of the eggplant, and close the fingers securely before lifting it clear of the plate.",
+    "vlm_recovery_action_summary": "re-align gripper; grasp eggplant securely; lift off plate",
+    "vlm_prevention_action": "Center and align the gripper precisely over the eggplant's main body before closing the fingers to ensure a stable grasp.",
+    "vlm_prevention_action_summary": "align gripper with eggplant; execute secure grasp",
+    "vlm_justification": "Around 3.8 seconds into the video, the robot closes its gripper on the eggplant, but fails to achieve a stable hold, causing the eggplant to slip out and remain on the plate while the robot lifts away empty.",
+    "vlm_temporal_justification": "At timestamp 03:13, as the robot gripper closes around the eggplant, the fingers make misaligned contact, causing the eggplant to tilt and flip over instead of being successfully grasped.",
+    "vlm_video_fps": 50.4,
+    "vlm_video_num_frames": 340,
+    "vlm_video_duration_seconds": 6.726190476190476,
+    "vlm_usage": {
+      "input_tokens": 21137,
+      "output_tokens": 378,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 3.8,\n \"failure_window_start_seconds\": 3.3,\n \"failure_window_end_seconds\": 4.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the eggplant on the plate, but the gripper misaligned with the object, causing the eggplant to slip out and flip over without being securely grasped.\",\n \"failure_reason_summary\": \"failed grasp attempt; eggplant slips from gripper\",\n \"recovery_action\": \"Re-align the end-effector over the eggplant, lower the open gripper onto the main body of the eggplant, and close the fingers securely before lifting it clear of the plate.\",\n \"recovery_action_summary\": \"re-align gripper; grasp eggplant securely; lift off plate\",\n \"prevention_action\": \"Center and align the gripper precisely over the eggplant's main body before closing the fingers to ensure a stable grasp.\",\n \"prevention_action_summary\": \"align gripper with eggplant; execute secure grasp\",\n \"justification\": \"Around 3.8 seconds into the video, the robot closes its gripper on the eggplant, but fails to achieve a stable hold, causing the eggplant to slip out and remain on the plate while the robot lifts away empty.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 234,
+    "failure_sim_step": 234,
+    "failure_policy_step": 39,
+    "failure_chunk_index": 0,
+    "first_post_failure_policy_step": 40,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-align the end-effector over the eggplant, lower the open gripper onto the main body of the eggplant, and close the fingers securely before lifting it clear of the plate."
+  }
+}
\ No newline at end of file
diff --git a/eggplant/ep007/labels.npz b/eggplant/ep007/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..eeba8a4d9713567dc5e8ebb04a67af374f49d101
--- /dev/null
+++ b/eggplant/ep007/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:51a86c0afedcb226b29db4026497e1de4deb6f09c59af70994d027dadd28afca
+size 3856
diff --git a/eggplant/ep008/example.md b/eggplant/ep008/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..a196103f57312fc3fa2dc029a6b459d95eafa776
--- /dev/null
+++ b/eggplant/ep008/example.md
@@ -0,0 +1,21 @@
+# t004__ep008__eggplant
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the eggplant off the plate
+- **success:** True
+- **control steps:** 388  (7.8s at 50 Hz)
+- **policy inferences:** 67  (replan_steps=6)
+- **video:** `eggplant/ep008/rollout.mp4`
+- **features:** `67` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–149 | 0–24 | reaching and grasping eggplant |
+| 3.0–6.0 | 150–299 | 25–49 | moving eggplant off plate |
+| 6.0–7.8 | 300–387 | 50–64 | retracting robot arm |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/eggplant/ep008/labels.json b/eggplant/ep008/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..f2a2220e6d2ce3ce5ddf098269139a0d592d3405
--- /dev/null
+++ b/eggplant/ep008/labels.json
@@ -0,0 +1,68 @@
+{
+  "rollout_id": "t004__ep008__eggplant",
+  "scene_variant": "",
+  "task_id": 4,
+  "task_file": "",
+  "instruction": "Take the eggplant off the plate",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:37:49+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0,
+      "control_step_start": 0,
+      "control_step_end": 149,
+      "policy_step_start": 0,
+      "policy_step_end": 24,
+      "phrase": "reaching and grasping eggplant",
+      "description": "reaching and grasping eggplant"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0,
+      "t_end_sec": 6.0,
+      "control_step_start": 150,
+      "control_step_end": 299,
+      "policy_step_start": 25,
+      "policy_step_end": 49,
+      "phrase": "moving eggplant off plate",
+      "description": "moving eggplant off plate"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.0,
+      "t_end_sec": 7.76,
+      "control_step_start": 300,
+      "control_step_end": 387,
+      "policy_step_start": 50,
+      "policy_step_end": 64,
+      "phrase": "retracting robot arm",
+      "description": "retracting robot arm"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/eggplant/ep008/labels.npz b/eggplant/ep008/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..d3f94a5a1733bda0d22a6b56d2c7c60c3a3de834
--- /dev/null
+++ b/eggplant/ep008/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:055cc2d72b826c63cdf7da75bae78f6cd777aa7cb0d7f1aadb46abacef3f5907
+size 3075
diff --git a/eggplant/ep009/example.md b/eggplant/ep009/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..be5086a1903228a22b252fdbe80eca3856995906
--- /dev/null
+++ b/eggplant/ep009/example.md
@@ -0,0 +1,22 @@
+# t004__ep009__eggplant
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the eggplant off the plate
+- **success:** True
+- **control steps:** 463  (9.2s at 51 Hz)
+- **policy inferences:** 80  (replan_steps=6)
+- **video:** `eggplant/ep009/rollout.mp4`
+- **features:** `80` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching for eggplant |
+| 3.0–6.0 | 152–303 | 25–50 | lifting and moving eggplant |
+| 6.0–9.0 | 304–455 | 50–75 | placing eggplant and retracting |
+| 9.0–9.2 | 456–462 | 76–77 | idle / remaining still |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/eggplant/ep009/labels.json b/eggplant/ep009/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..3af46a786a5e65c6e2643270e6985212244e6b80
--- /dev/null
+++ b/eggplant/ep009/labels.json
@@ -0,0 +1,79 @@
+{
+  "rollout_id": "t004__ep009__eggplant",
+  "scene_variant": "",
+  "task_id": 4,
+  "task_file": "",
+  "instruction": "Take the eggplant off the plate",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:37:59+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0039525691699605,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for eggplant",
+      "description": "reaching for eggplant"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0039525691699605,
+      "t_end_sec": 6.007905138339921,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "lifting and moving eggplant",
+      "description": "lifting and moving eggplant"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.007905138339921,
+      "t_end_sec": 9.011857707509881,
+      "control_step_start": 304,
+      "control_step_end": 455,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "placing eggplant and retracting",
+      "description": "placing eggplant and retracting"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.011857707509881,
+      "t_end_sec": 9.150197628458498,
+      "control_step_start": 456,
+      "control_step_end": 462,
+      "policy_step_start": 76,
+      "policy_step_end": 77,
+      "phrase": "idle / remaining still",
+      "description": "idle / remaining still"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/eggplant/ep009/labels.npz b/eggplant/ep009/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..bbf71329dfc41a95d12086987a9c80f48c815d34
--- /dev/null
+++ b/eggplant/ep009/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fa3a332e88b13c70932ea0df9ad66415b0870b4f8148e24e64568d5d173efa6f
+size 3117
diff --git a/eggplant/ep010/example.md b/eggplant/ep010/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..d35e33941672ff9db6b082b52fa80223d881a5de
--- /dev/null
+++ b/eggplant/ep010/example.md
@@ -0,0 +1,21 @@
+# t004__ep010__eggplant
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the eggplant off the plate
+- **success:** True
+- **control steps:** 423  (8.3s at 51 Hz)
+- **policy inferences:** 73  (replan_steps=6)
+- **video:** `eggplant/ep010/rollout.mp4`
+- **features:** `73` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching for eggplant |
+| 3.0–6.0 | 152–303 | 25–50 | grasping and moving eggplant |
+| 6.0–8.3 | 304–422 | 50–70 | releasing eggplant and retracting |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/eggplant/ep010/labels.json b/eggplant/ep010/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..b59a0ea2f1c26186f76cc28d3574f0d4023b5013
--- /dev/null
+++ b/eggplant/ep010/labels.json
@@ -0,0 +1,68 @@
+{
+  "rollout_id": "t004__ep010__eggplant",
+  "scene_variant": "",
+  "task_id": 4,
+  "task_file": "",
+  "instruction": "Take the eggplant off the plate",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:38:07+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9980276134122286,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for eggplant",
+      "description": "reaching for eggplant"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9980276134122286,
+      "t_end_sec": 5.996055226824457,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "grasping and moving eggplant",
+      "description": "grasping and moving eggplant"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.996055226824457,
+      "t_end_sec": 8.34319526627219,
+      "control_step_start": 304,
+      "control_step_end": 422,
+      "policy_step_start": 50,
+      "policy_step_end": 70,
+      "phrase": "releasing eggplant and retracting",
+      "description": "releasing eggplant and retracting"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/eggplant/ep010/labels.npz b/eggplant/ep010/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..72de243e80f17567fd343a5ef3b7324266add45b
--- /dev/null
+++ b/eggplant/ep010/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fd9ec51e55c212099d5798bc462587b8be25cf3f8338e654a2b8d27ce056d074
+size 3080
diff --git a/eggplant/ep011/example.md b/eggplant/ep011/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..394554ae6307f30a13d755f65e2b6c4be2a430be
--- /dev/null
+++ b/eggplant/ep011/example.md
@@ -0,0 +1,21 @@
+# t004__ep011__eggplant
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the eggplant off the plate
+- **success:** True
+- **control steps:** 411  (8.1s at 51 Hz)
+- **policy inferences:** 71  (replan_steps=6)
+- **video:** `eggplant/ep011/rollout.mp4`
+- **features:** `71` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching for the eggplant |
+| 3.0–6.0 | 152–303 | 25–50 | lifting eggplant off plate |
+| 6.0–8.1 | 304–410 | 50–68 | releasing eggplant and retracting |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/eggplant/ep011/labels.json b/eggplant/ep011/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..f7d49bf6477bafc5d0ba9527efa8b5383d08e621
--- /dev/null
+++ b/eggplant/ep011/labels.json
@@ -0,0 +1,68 @@
+{
+  "rollout_id": "t004__ep011__eggplant",
+  "scene_variant": "",
+  "task_id": 4,
+  "task_file": "",
+  "instruction": "Take the eggplant off the plate",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:38:14+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0039525691699605,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for the eggplant",
+      "description": "reaching for the eggplant"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0039525691699605,
+      "t_end_sec": 6.007905138339921,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "lifting eggplant off plate",
+      "description": "lifting eggplant off plate"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.007905138339921,
+      "t_end_sec": 8.122529644268775,
+      "control_step_start": 304,
+      "control_step_end": 410,
+      "policy_step_start": 50,
+      "policy_step_end": 68,
+      "phrase": "releasing eggplant and retracting",
+      "description": "releasing eggplant and retracting"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/eggplant/ep011/labels.npz b/eggplant/ep011/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..df81ff8bb607581f846e8ec1c0fc15b2041613e1
--- /dev/null
+++ b/eggplant/ep011/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6406b6fe94475730e627e44c2d8caca1ba0339c3e64acf0e86063ce6d2526f91
+size 3080
diff --git a/eggplant/ep012/example.md b/eggplant/ep012/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..95b7d67575d9c606f1d367842e0a310bbce2e6f9
--- /dev/null
+++ b/eggplant/ep012/example.md
@@ -0,0 +1,21 @@
+# t004__ep012__eggplant
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the eggplant off the plate
+- **success:** True
+- **control steps:** 410  (8.1s at 51 Hz)
+- **policy inferences:** 70  (replan_steps=6)
+- **video:** `eggplant/ep012/rollout.mp4`
+- **features:** `70` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–152 | 0–25 | reaching for eggplant on plate |
+| 3.0–6.0 | 153–305 | 25–50 | lifting and moving eggplant |
+| 6.0–8.1 | 306–409 | 51–68 | placing eggplant and retracting |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/eggplant/ep012/labels.json b/eggplant/ep012/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..3df89100e86d9fc2a6a7582142685cf25fb9ff85
--- /dev/null
+++ b/eggplant/ep012/labels.json
@@ -0,0 +1,68 @@
+{
+  "rollout_id": "t004__ep012__eggplant",
+  "scene_variant": "",
+  "task_id": 4,
+  "task_file": "",
+  "instruction": "Take the eggplant off the plate",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:38:23+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.005893909626719,
+      "control_step_start": 0,
+      "control_step_end": 152,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for eggplant on plate",
+      "description": "reaching for eggplant on plate"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.005893909626719,
+      "t_end_sec": 6.011787819253438,
+      "control_step_start": 153,
+      "control_step_end": 305,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "lifting and moving eggplant",
+      "description": "lifting and moving eggplant"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.011787819253438,
+      "t_end_sec": 8.055009823182711,
+      "control_step_start": 306,
+      "control_step_end": 409,
+      "policy_step_start": 51,
+      "policy_step_end": 68,
+      "phrase": "placing eggplant and retracting",
+      "description": "placing eggplant and retracting"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/eggplant/ep012/labels.npz b/eggplant/ep012/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..b766d3087425d0af861fe8d16aed34eabcbe3c27
--- /dev/null
+++ b/eggplant/ep012/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a8f1123974be8dc5e0404dacbd2c8d0ecd25c9bdb46d42a95d529b5a7033212c
+size 3083
diff --git a/eggplant/ep013/example.md b/eggplant/ep013/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..520c34cf68a45c4cf322a847ddb6973c98bdfbf2
--- /dev/null
+++ b/eggplant/ep013/example.md
@@ -0,0 +1,21 @@
+# t004__ep013__eggplant
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the eggplant off the plate
+- **success:** True
+- **control steps:** 366  (7.2s at 51 Hz)
+- **policy inferences:** 62  (replan_steps=6)
+- **video:** `eggplant/ep013/rollout.mp4`
+- **features:** `62` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–152 | 0–25 | approaching and grasping eggplant |
+| 3.0–6.0 | 153–305 | 25–50 | moving eggplant off plate |
+| 6.0–7.2 | 306–365 | 51–60 | retracting robotic arm |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/eggplant/ep013/labels.json b/eggplant/ep013/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..eb0c31327af66b622b6f3b66720859729ce23e7f
--- /dev/null
+++ b/eggplant/ep013/labels.json
@@ -0,0 +1,68 @@
+{
+  "rollout_id": "t004__ep013__eggplant",
+  "scene_variant": "",
+  "task_id": 4,
+  "task_file": "",
+  "instruction": "Take the eggplant off the plate",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:38:31+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0,
+      "control_step_start": 0,
+      "control_step_end": 152,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "approaching and grasping eggplant",
+      "description": "approaching and grasping eggplant"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0,
+      "t_end_sec": 6.0,
+      "control_step_start": 153,
+      "control_step_end": 305,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "moving eggplant off plate",
+      "description": "moving eggplant off plate"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.0,
+      "t_end_sec": 7.176470588235294,
+      "control_step_start": 306,
+      "control_step_end": 365,
+      "policy_step_start": 51,
+      "policy_step_end": 60,
+      "phrase": "retracting robotic arm",
+      "description": "retracting robotic arm"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/eggplant/ep013/labels.npz b/eggplant/ep013/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..ff672c9a8b17a027167170a7410dbc40f1b5a59e
--- /dev/null
+++ b/eggplant/ep013/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:defc5825cc5c82f7b0454747be699c45bb17a28746c3278deb8e3d605f0491c2
+size 3082
diff --git a/eggplant/ep014/example.md b/eggplant/ep014/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..2eb31b2d6a612afeda832bc717aa438d7feacc58
--- /dev/null
+++ b/eggplant/ep014/example.md
@@ -0,0 +1,21 @@
+# t004__ep014__eggplant
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the eggplant off the plate
+- **success:** True
+- **control steps:** 418  (8.4s at 50 Hz)
+- **policy inferences:** 72  (replan_steps=6)
+- **video:** `eggplant/ep014/rollout.mp4`
+- **features:** `72` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–149 | 0–24 | reaching for and grasping eggplant |
+| 3.0–6.0 | 150–299 | 25–49 | lifting and moving eggplant |
+| 6.0–8.4 | 300–417 | 50–69 | releasing eggplant and retracting |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/eggplant/ep014/labels.json b/eggplant/ep014/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..71d511fbfe00b1dc3c518f70e5a474bdba09955f
--- /dev/null
+++ b/eggplant/ep014/labels.json
@@ -0,0 +1,68 @@
+{
+  "rollout_id": "t004__ep014__eggplant",
+  "scene_variant": "",
+  "task_id": 4,
+  "task_file": "",
+  "instruction": "Take the eggplant off the plate",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:38:38+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0,
+      "control_step_start": 0,
+      "control_step_end": 149,
+      "policy_step_start": 0,
+      "policy_step_end": 24,
+      "phrase": "reaching for and grasping eggplant",
+      "description": "reaching for and grasping eggplant"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0,
+      "t_end_sec": 6.0,
+      "control_step_start": 150,
+      "control_step_end": 299,
+      "policy_step_start": 25,
+      "policy_step_end": 49,
+      "phrase": "lifting and moving eggplant",
+      "description": "lifting and moving eggplant"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.0,
+      "t_end_sec": 8.36,
+      "control_step_start": 300,
+      "control_step_end": 417,
+      "policy_step_start": 50,
+      "policy_step_end": 69,
+      "phrase": "releasing eggplant and retracting",
+      "description": "releasing eggplant and retracting"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/eggplant/ep014/labels.npz b/eggplant/ep014/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..00f26f746909753caeda30c62f36ff26768451ad
--- /dev/null
+++ b/eggplant/ep014/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c0de5faadc10ba9c7286c3041cdf53c215515b6d04408a16f9be76bf9d206571
+size 3083
diff --git a/eggplant/ep015/example.md b/eggplant/ep015/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..00075b11aa99e909bb0724b6995f0996b9c13322
--- /dev/null
+++ b/eggplant/ep015/example.md
@@ -0,0 +1,21 @@
+# t004__ep015__eggplant
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the eggplant off the plate
+- **success:** True
+- **control steps:** 400  (8.0s at 50 Hz)
+- **policy inferences:** 70  (replan_steps=6)
+- **video:** `eggplant/ep015/rollout.mp4`
+- **features:** `70` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–149 | 0–24 | reaching for and grasping eggplant |
+| 3.0–6.0 | 150–299 | 25–49 | lifting and placing eggplant |
+| 6.0–8.0 | 300–399 | 50–66 | retracting arm |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/eggplant/ep015/labels.json b/eggplant/ep015/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..2eef2dd3b5acc1898b13458aa7120e433db7d9dd
--- /dev/null
+++ b/eggplant/ep015/labels.json
@@ -0,0 +1,68 @@
+{
+  "rollout_id": "t004__ep015__eggplant",
+  "scene_variant": "",
+  "task_id": 4,
+  "task_file": "",
+  "instruction": "Take the eggplant off the plate",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:38:46+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9940119760479043,
+      "control_step_start": 0,
+      "control_step_end": 149,
+      "policy_step_start": 0,
+      "policy_step_end": 24,
+      "phrase": "reaching for and grasping eggplant",
+      "description": "reaching for and grasping eggplant"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9940119760479043,
+      "t_end_sec": 5.9880239520958085,
+      "control_step_start": 150,
+      "control_step_end": 299,
+      "policy_step_start": 25,
+      "policy_step_end": 49,
+      "phrase": "lifting and placing eggplant",
+      "description": "lifting and placing eggplant"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.9880239520958085,
+      "t_end_sec": 7.984031936127744,
+      "control_step_start": 300,
+      "control_step_end": 399,
+      "policy_step_start": 50,
+      "policy_step_end": 66,
+      "phrase": "retracting arm",
+      "description": "retracting arm"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/eggplant/ep015/labels.npz b/eggplant/ep015/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..b1d044c6d12c928f98e7a43316f8160d1762fcc9
--- /dev/null
+++ b/eggplant/ep015/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:75678930fd2da66293e749a9ce8d8eb9aa458f54ca6249acd7b43c764d388832
+size 3076
diff --git a/eggplant/ep016/example.md b/eggplant/ep016/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..83b24978fc7212bb81e598a38ed7441c02e52d85
--- /dev/null
+++ b/eggplant/ep016/example.md
@@ -0,0 +1,21 @@
+# t004__ep016__eggplant
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the eggplant off the plate
+- **success:** True
+- **control steps:** 380  (7.5s at 51 Hz)
+- **policy inferences:** 66  (replan_steps=6)
+- **video:** `eggplant/ep016/rollout.mp4`
+- **features:** `66` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching for and grasping eggplant |
+| 3.0–6.0 | 152–303 | 25–50 | lifting and moving eggplant |
+| 6.0–7.5 | 304–379 | 50–63 | releasing eggplant on table |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/eggplant/ep016/labels.json b/eggplant/ep016/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..1008095be2804ed0ed17dc49d625280377a3d5ae
--- /dev/null
+++ b/eggplant/ep016/labels.json
@@ -0,0 +1,68 @@
+{
+  "rollout_id": "t004__ep016__eggplant",
+  "scene_variant": "",
+  "task_id": 4,
+  "task_file": "",
+  "instruction": "Take the eggplant off the plate",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:38:53+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9980276134122286,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for and grasping eggplant",
+      "description": "reaching for and grasping eggplant"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9980276134122286,
+      "t_end_sec": 5.996055226824457,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "lifting and moving eggplant",
+      "description": "lifting and moving eggplant"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.996055226824457,
+      "t_end_sec": 7.495069033530571,
+      "control_step_start": 304,
+      "control_step_end": 379,
+      "policy_step_start": 50,
+      "policy_step_end": 63,
+      "phrase": "releasing eggplant on table",
+      "description": "releasing eggplant on table"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/eggplant/ep016/labels.npz b/eggplant/ep016/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..bf206d4c2aa9074c9105a9ca3866521cf332c065
--- /dev/null
+++ b/eggplant/ep016/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b1dd7b3198cc1cc1c13ebe2588dceea3ae9070ad7ebd375dd6d7036c7f76de1e
+size 3083
diff --git a/eggplant/ep017/example.md b/eggplant/ep017/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..80b589c79cb906d60eac79304acb69a781848bbc
--- /dev/null
+++ b/eggplant/ep017/example.md
@@ -0,0 +1,32 @@
+# t004__ep017__eggplant
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the eggplant off the plate
+- **success:** False
+- **control steps:** 271  (5.4s at 50 Hz)
+- **policy inferences:** 47  (replan_steps=5)
+- **video:** `eggplant/ep017/rollout.mp4`
+- **features:** `47` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–30 | reaching for and grasping eggplant |
+| 3.0–5.4 | 151–270 | 30–54 | lifting empty gripper away |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 3.326693227091633 s (00:03.3)
+- **frame / control_step:** 167 (policy_step=33, chunk_index=2)
+- **coarse onset:** 3.0 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot closed its gripper over the eggplant on the plate but failed to establish a firm grip, lifting away empty while leaving the eggplant behind.
+  Summary: failed grasp attempt; gripper lifted empty
+- **recovery:** Lower the end-effector back down over the eggplant, reposition the gripper fingers around its center, close securely, and lift it off the plate.
+  Summary: re-align gripper; grasp eggplant securely; lift off plate
+- **prevention:** Align the gripper precisely around the middle of the eggplant before closing to ensure a secure grip prior to lifting.
+  Summary: center gripper on eggplant; secure grasp before lifting
+- **justification:** Around 3.0 seconds into the video, the robot attempts to close its gripper around the eggplant but fails to secure it. As a result, the gripper lifts up empty while the eggplant stays on the plate.
diff --git a/eggplant/ep017/labels.json b/eggplant/ep017/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..7ed6327f0dc597020266644fd85cba4f59cbaa7e
--- /dev/null
+++ b/eggplant/ep017/labels.json
@@ -0,0 +1,87 @@
+{
+  "rollout_id": "t004__ep017__eggplant",
+  "scene_variant": "",
+  "task_id": 4,
+  "task_file": "",
+  "instruction": "Take the eggplant off the plate",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:39:24+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00796812749004,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 30,
+      "phrase": "reaching for and grasping eggplant",
+      "description": "reaching for and grasping eggplant"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00796812749004,
+      "t_end_sec": 5.398406374501992,
+      "control_step_start": 151,
+      "control_step_end": 270,
+      "policy_step_start": 30,
+      "policy_step_end": 54,
+      "phrase": "lifting empty gripper away",
+      "description": "lifting empty gripper away"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 3.326693227091633,
+    "vlm_failure_onset_timestamp": "00:03.3",
+    "vlm_failure_onset_frame": 167,
+    "vlm_failure_onset_step": 167,
+    "vlm_coarse_failure_onset_seconds": 3.0,
+    "vlm_failure_window_start_seconds": 2.5,
+    "vlm_failure_window_end_seconds": 3.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot closed its gripper over the eggplant on the plate but failed to establish a firm grip, lifting away empty while leaving the eggplant behind.",
+    "vlm_failure_reason_summary": "failed grasp attempt; gripper lifted empty",
+    "vlm_recovery_action": "Lower the end-effector back down over the eggplant, reposition the gripper fingers around its center, close securely, and lift it off the plate.",
+    "vlm_recovery_action_summary": "re-align gripper; grasp eggplant securely; lift off plate",
+    "vlm_prevention_action": "Align the gripper precisely around the middle of the eggplant before closing to ensure a secure grip prior to lifting.",
+    "vlm_prevention_action_summary": "center gripper on eggplant; secure grasp before lifting",
+    "vlm_justification": "Around 3.0 seconds into the video, the robot attempts to close its gripper around the eggplant but fails to secure it. As a result, the gripper lifts up empty while the eggplant stays on the plate.",
+    "vlm_temporal_justification": "At second 167 (02:47), the gripper fingers finish closing completely together without clamping the eggplant, making the grasp failure visibly apparent as the robot fails to secure the object.",
+    "vlm_video_fps": 50.2,
+    "vlm_video_num_frames": 271,
+    "vlm_video_duration_seconds": 5.378486055776892,
+    "vlm_usage": {
+      "input_tokens": 19238,
+      "output_tokens": 362,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 3.0,\n \"failure_window_start_seconds\": 2.5,\n \"failure_window_end_seconds\": 3.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot closed its gripper over the eggplant on the plate but failed to establish a firm grip, lifting away empty while leaving the eggplant behind.\",\n \"failure_reason_summary\": \"failed grasp attempt; gripper lifted empty\",\n \"recovery_action\": \"Lower the end-effector back down over the eggplant, reposition the gripper fingers around its center, close securely, and lift it off the plate.\",\n \"recovery_action_summary\": \"re-align gripper; grasp eggplant securely; lift off plate\",\n \"prevention_action\": \"Align the gripper precisely around the middle of the eggplant before closing to ensure a secure grip prior to lifting.\",\n \"prevention_action_summary\": \"center gripper on eggplant; secure grasp before lifting\",\n \"justification\": \"Around 3.0 seconds into the video, the robot attempts to close its gripper around the eggplant but fails to secure it. As a result, the gripper lifts up empty while the eggplant stays on the plate.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 167,
+    "failure_sim_step": 167,
+    "failure_policy_step": 33,
+    "failure_chunk_index": 2,
+    "first_post_failure_policy_step": 34,
+    "failure_type": "grasp_failure",
+    "correction_action": "Lower the end-effector back down over the eggplant, reposition the gripper fingers around its center, close securely, and lift it off the plate."
+  }
+}
\ No newline at end of file
diff --git a/eggplant/ep017/labels.npz b/eggplant/ep017/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..6bbb4d8285b0e61833f973f1c03f84f08d2e5b0f
--- /dev/null
+++ b/eggplant/ep017/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:515d192573b90578e55e6eb0a4f4903f793a272c60f72452c07af7f2b7206594
+size 3773
diff --git a/eggplant/ep018/example.md b/eggplant/ep018/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..8824e5c7eaae3f0cf0cd827317d80f6441c3c195
--- /dev/null
+++ b/eggplant/ep018/example.md
@@ -0,0 +1,21 @@
+# t004__ep018__eggplant
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the eggplant off the plate
+- **success:** True
+- **control steps:** 360  (7.2s at 50 Hz)
+- **policy inferences:** 62  (replan_steps=6)
+- **video:** `eggplant/ep018/rollout.mp4`
+- **features:** `62` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–148 | 0–24 | reaching for and grasping eggplant |
+| 3.0–6.0 | 149–297 | 24–49 | moving eggplant off plate |
+| 6.0–7.2 | 298–359 | 49–59 | releasing eggplant and retracting |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/eggplant/ep018/labels.json b/eggplant/ep018/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..8c5afadb337427824743363d06931b0b47ccff72
--- /dev/null
+++ b/eggplant/ep018/labels.json
@@ -0,0 +1,68 @@
+{
+  "rollout_id": "t004__ep018__eggplant",
+  "scene_variant": "",
+  "task_id": 4,
+  "task_file": "",
+  "instruction": "Take the eggplant off the plate",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:39:31+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9979879275653922,
+      "control_step_start": 0,
+      "control_step_end": 148,
+      "policy_step_start": 0,
+      "policy_step_end": 24,
+      "phrase": "reaching for and grasping eggplant",
+      "description": "reaching for and grasping eggplant"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9979879275653922,
+      "t_end_sec": 5.9959758551307845,
+      "control_step_start": 149,
+      "control_step_end": 297,
+      "policy_step_start": 24,
+      "policy_step_end": 49,
+      "phrase": "moving eggplant off plate",
+      "description": "moving eggplant off plate"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.9959758551307845,
+      "t_end_sec": 7.243460764587525,
+      "control_step_start": 298,
+      "control_step_end": 359,
+      "policy_step_start": 49,
+      "policy_step_end": 59,
+      "phrase": "releasing eggplant and retracting",
+      "description": "releasing eggplant and retracting"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/eggplant/ep018/labels.npz b/eggplant/ep018/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..3f005ef02cc4555d8d7d904ca519bc73831f66f7
--- /dev/null
+++ b/eggplant/ep018/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6a4c2a191ddc3661ed15025b5bef4fbaa5314ce76db8923a8f7a65bd7d4af29e
+size 3082
diff --git a/eggplant/ep019/example.md b/eggplant/ep019/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..e3a097caaf8c8e6788bf382627a1fb5a2e0c4f28
--- /dev/null
+++ b/eggplant/ep019/example.md
@@ -0,0 +1,33 @@
+# t004__ep019__eggplant
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the eggplant off the plate
+- **success:** False
+- **control steps:** 345  (6.8s at 51 Hz)
+- **policy inferences:** 59  (replan_steps=6)
+- **video:** `eggplant/ep019/rollout.mp4`
+- **features:** `59` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–152 | 0–25 | approaching and grasping eggplant |
+| 3.0–6.0 | 153–305 | 25–50 | lifting eggplant and dropping it |
+| 6.0–6.8 | 306–344 | 51–57 | moving arm away empty |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 4.294117647058823 s (00:04.3)
+- **frame / control_step:** 219 (policy_step=36, chunk_index=3)
+- **coarse onset:** 3.5 s (refined=True)
+- **mode:** object_displacement
+- **reason:** The robot grasped the eggplant but lost its hold while lifting, dropping the eggplant back onto the plate.
+  Summary: dropping object during lift; lost grip on eggplant
+- **recovery:** Re-align the gripper over the eggplant on the plate, establish a firm grasp, and lift it completely off the plate.
+  Summary: re-grasp eggplant; lift off plate
+- **prevention:** Ensure a firm and secure hold on the eggplant before initiating the lift so it does not slip from the gripper.
+  Summary: firm grasp on eggplant; lift cleanly
+- **justification:** At around 3.5 seconds, the robot lifts the eggplant off the plate, but the grip is insecure and the eggplant slips out, falling back onto the plate. The robot then moves away without holding the object.
diff --git a/eggplant/ep019/labels.json b/eggplant/ep019/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..04ca6a7721bf253099168dd1c149901916b12a82
--- /dev/null
+++ b/eggplant/ep019/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t004__ep019__eggplant",
+  "scene_variant": "",
+  "task_id": 4,
+  "task_file": "",
+  "instruction": "Take the eggplant off the plate",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:40:00+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0,
+      "control_step_start": 0,
+      "control_step_end": 152,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "approaching and grasping eggplant",
+      "description": "approaching and grasping eggplant"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0,
+      "t_end_sec": 6.0,
+      "control_step_start": 153,
+      "control_step_end": 305,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "lifting eggplant and dropping it",
+      "description": "lifting eggplant and dropping it"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.0,
+      "t_end_sec": 6.764705882352941,
+      "control_step_start": 306,
+      "control_step_end": 344,
+      "policy_step_start": 51,
+      "policy_step_end": 57,
+      "phrase": "moving arm away empty",
+      "description": "moving arm away empty"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "object_displacement",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 4.294117647058823,
+    "vlm_failure_onset_timestamp": "00:04.3",
+    "vlm_failure_onset_frame": 219,
+    "vlm_failure_onset_step": 219,
+    "vlm_coarse_failure_onset_seconds": 3.5,
+    "vlm_failure_window_start_seconds": 3.0,
+    "vlm_failure_window_end_seconds": 4.0,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot grasped the eggplant but lost its hold while lifting, dropping the eggplant back onto the plate.",
+    "vlm_failure_reason_summary": "dropping object during lift; lost grip on eggplant",
+    "vlm_recovery_action": "Re-align the gripper over the eggplant on the plate, establish a firm grasp, and lift it completely off the plate.",
+    "vlm_recovery_action_summary": "re-grasp eggplant; lift off plate",
+    "vlm_prevention_action": "Ensure a firm and secure hold on the eggplant before initiating the lift so it does not slip from the gripper.",
+    "vlm_prevention_action_summary": "firm grasp on eggplant; lift cleanly",
+    "vlm_justification": "At around 3.5 seconds, the robot lifts the eggplant off the plate, but the grip is insecure and the eggplant slips out, falling back onto the plate. The robot then moves away without holding the object.",
+    "vlm_temporal_justification": "The robot attempts to lift the eggplant after grasping it, but at 03:14 (194.0s in clip), the eggplant slips from the gripper fingers and drops back onto the plate as the arm ascends.",
+    "vlm_video_fps": 51.0,
+    "vlm_video_num_frames": 345,
+    "vlm_video_duration_seconds": 6.745098039215686,
+    "vlm_usage": {
+      "input_tokens": 21627,
+      "output_tokens": 357,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"object_displacement\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 3.5,\n \"failure_window_start_seconds\": 3.0,\n \"failure_window_end_seconds\": 4.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot grasped the eggplant but lost its hold while lifting, dropping the eggplant back onto the plate.\",\n \"failure_reason_summary\": \"dropping object during lift; lost grip on eggplant\",\n \"recovery_action\": \"Re-align the gripper over the eggplant on the plate, establish a firm grasp, and lift it completely off the plate.\",\n \"recovery_action_summary\": \"re-grasp eggplant; lift off plate\",\n \"prevention_action\": \"Ensure a firm and secure hold on the eggplant before initiating the lift so it does not slip from the gripper.\",\n \"prevention_action_summary\": \"firm grasp on eggplant; lift cleanly\",\n \"justification\": \"At around 3.5 seconds, the robot lifts the eggplant off the plate, but the grip is insecure and the eggplant slips out, falling back onto the plate. The robot then moves away without holding the object.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 219,
+    "failure_sim_step": 219,
+    "failure_policy_step": 36,
+    "failure_chunk_index": 3,
+    "first_post_failure_policy_step": 37,
+    "failure_type": "object_displacement",
+    "correction_action": "Re-align the gripper over the eggplant on the plate, establish a firm grasp, and lift it completely off the plate."
+  }
+}
\ No newline at end of file
diff --git a/eggplant/ep019/labels.npz b/eggplant/ep019/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..d2babd4d88e62ffe499cca6484c8444cc2d7d9ef
--- /dev/null
+++ b/eggplant/ep019/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:65064a67be2debb4075e21d202f554baccefc75b707a69ece585b55e3f291457
+size 3698
diff --git a/eggplant/ep020/example.md b/eggplant/ep020/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..918ee8278160843087455ad92aa200bbe0c06d3f
--- /dev/null
+++ b/eggplant/ep020/example.md
@@ -0,0 +1,21 @@
+# t004__ep020__eggplant
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the eggplant off the plate
+- **success:** True
+- **control steps:** 356  (7.0s at 51 Hz)
+- **policy inferences:** 62  (replan_steps=6)
+- **video:** `eggplant/ep020/rollout.mp4`
+- **features:** `62` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching for and grasping eggplant |
+| 3.0–6.0 | 152–303 | 25–50 | moving and placing eggplant on table |
+| 6.0–7.0 | 304–355 | 50–59 | retracting arm and staying idle |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/eggplant/ep020/labels.json b/eggplant/ep020/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..e17459387c672db60fa71c659d48f83957729ba1
--- /dev/null
+++ b/eggplant/ep020/labels.json
@@ -0,0 +1,68 @@
+{
+  "rollout_id": "t004__ep020__eggplant",
+  "scene_variant": "",
+  "task_id": 4,
+  "task_file": "",
+  "instruction": "Take the eggplant off the plate",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:40:07+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9921259842519685,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for and grasping eggplant",
+      "description": "reaching for and grasping eggplant"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9921259842519685,
+      "t_end_sec": 5.984251968503937,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "moving and placing eggplant on table",
+      "description": "moving and placing eggplant on table"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.984251968503937,
+      "t_end_sec": 7.0078740157480315,
+      "control_step_start": 304,
+      "control_step_end": 355,
+      "policy_step_start": 50,
+      "policy_step_end": 59,
+      "phrase": "retracting arm and staying idle",
+      "description": "retracting arm and staying idle"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/eggplant/ep020/labels.npz b/eggplant/ep020/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..c4686c09c0db7154a168b4ac6a4806611a8783a9
--- /dev/null
+++ b/eggplant/ep020/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:70207aa19b654b534426ed451c5e0b840a554998fd36187eda20d97c8bd97195
+size 3100
diff --git a/eggplant/ep021/example.md b/eggplant/ep021/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..227c61cb53f3f6d67b26a8946cd49ac5f2dcf582
--- /dev/null
+++ b/eggplant/ep021/example.md
@@ -0,0 +1,33 @@
+# t004__ep021__eggplant
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the eggplant off the plate
+- **success:** False
+- **control steps:** 437  (8.7s at 50 Hz)
+- **policy inferences:** 76  (replan_steps=6)
+- **video:** `eggplant/ep021/rollout.mp4`
+- **features:** `76` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–149 | 0–24 | lowering gripper toward eggplant |
+| 3.0–6.0 | 150–299 | 25–49 | grasping and dropping eggplant |
+| 6.0–8.7 | 300–436 | 50–72 | raising empty gripper |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 5.209580838323353 s (00:05.2)
+- **frame / control_step:** 261 (policy_step=43, chunk_index=3)
+- **coarse onset:** 5.5 s (refined=True)
+- **mode:** object_displacement
+- **reason:** The robot grasped the eggplant and attempted to lift it off the plate, but lost its hold, causing the eggplant to slip out of the gripper and fall back down.
+  Summary: dropping eggplant during lift; unstable grip
+- **recovery:** Re-align the gripper over the eggplant, establish a firm grasp, and lift it completely off the plate.
+  Summary: re-grasp eggplant; lift clear of plate
+- **prevention:** Ensure a secure, centered grip on the eggplant before lifting up to prevent it from slipping.
+  Summary: grasp eggplant securely; lift straight up
+- **justification:** At around 5.5 seconds, as the arm lifts upward, the eggplant slips out from between the gripper fingers and falls back onto the plate, leaving the arm holding nothing.
diff --git a/eggplant/ep021/labels.json b/eggplant/ep021/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..a2a4f3b0ec9191edc95d4e1eb4e9e76aec06e379
--- /dev/null
+++ b/eggplant/ep021/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t004__ep021__eggplant",
+  "scene_variant": "",
+  "task_id": 4,
+  "task_file": "",
+  "instruction": "Take the eggplant off the plate",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:40:40+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9940119760479043,
+      "control_step_start": 0,
+      "control_step_end": 149,
+      "policy_step_start": 0,
+      "policy_step_end": 24,
+      "phrase": "lowering gripper toward eggplant",
+      "description": "lowering gripper toward eggplant"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9940119760479043,
+      "t_end_sec": 5.9880239520958085,
+      "control_step_start": 150,
+      "control_step_end": 299,
+      "policy_step_start": 25,
+      "policy_step_end": 49,
+      "phrase": "grasping and dropping eggplant",
+      "description": "grasping and dropping eggplant"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.9880239520958085,
+      "t_end_sec": 8.72255489021956,
+      "control_step_start": 300,
+      "control_step_end": 436,
+      "policy_step_start": 50,
+      "policy_step_end": 72,
+      "phrase": "raising empty gripper",
+      "description": "raising empty gripper"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "object_displacement",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 5.209580838323353,
+    "vlm_failure_onset_timestamp": "00:05.2",
+    "vlm_failure_onset_frame": 261,
+    "vlm_failure_onset_step": 261,
+    "vlm_coarse_failure_onset_seconds": 5.5,
+    "vlm_failure_window_start_seconds": 5.0,
+    "vlm_failure_window_end_seconds": 6.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot grasped the eggplant and attempted to lift it off the plate, but lost its hold, causing the eggplant to slip out of the gripper and fall back down.",
+    "vlm_failure_reason_summary": "dropping eggplant during lift; unstable grip",
+    "vlm_recovery_action": "Re-align the gripper over the eggplant, establish a firm grasp, and lift it completely off the plate.",
+    "vlm_recovery_action_summary": "re-grasp eggplant; lift clear of plate",
+    "vlm_prevention_action": "Ensure a secure, centered grip on the eggplant before lifting up to prevent it from slipping.",
+    "vlm_prevention_action_summary": "grasp eggplant securely; lift straight up",
+    "vlm_justification": "At around 5.5 seconds, as the arm lifts upward, the eggplant slips out from between the gripper fingers and falls back onto the plate, leaving the arm holding nothing.",
+    "vlm_temporal_justification": "At 02:15 (corresponding to frame 261 of the original rollout), the robotic arm begins lifting up, but the gripper fails to maintain a secure hold on the eggplant, causing it to slip out of the fingers and fall back down onto the plate.",
+    "vlm_video_fps": 50.1,
+    "vlm_video_num_frames": 437,
+    "vlm_video_duration_seconds": 8.70259481037924,
+    "vlm_usage": {
+      "input_tokens": 21391,
+      "output_tokens": 361,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"object_displacement\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 5.5,\n \"failure_window_start_seconds\": 5.0,\n \"failure_window_end_seconds\": 6.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot grasped the eggplant and attempted to lift it off the plate, but lost its hold, causing the eggplant to slip out of the gripper and fall back down.\",\n \"failure_reason_summary\": \"dropping eggplant during lift; unstable grip\",\n \"recovery_action\": \"Re-align the gripper over the eggplant, establish a firm grasp, and lift it completely off the plate.\",\n \"recovery_action_summary\": \"re-grasp eggplant; lift clear of plate\",\n \"prevention_action\": \"Ensure a secure, centered grip on the eggplant before lifting up to prevent it from slipping.\",\n \"prevention_action_summary\": \"grasp eggplant securely; lift straight up\",\n \"justification\": \"At around 5.5 seconds, as the arm lifts upward, the eggplant slips out from between the gripper fingers and falls back onto the plate, leaving the arm holding nothing.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 261,
+    "failure_sim_step": 261,
+    "failure_policy_step": 43,
+    "failure_chunk_index": 3,
+    "first_post_failure_policy_step": 44,
+    "failure_type": "object_displacement",
+    "correction_action": "Re-align the gripper over the eggplant, establish a firm grasp, and lift it completely off the plate."
+  }
+}
\ No newline at end of file
diff --git a/eggplant/ep021/labels.npz b/eggplant/ep021/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..01947b150b29fb7caaae050e7d82cdb7a5c03b13
--- /dev/null
+++ b/eggplant/ep021/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c76496c6d847e1e766c940429a8272d018d42c1c870bb031d6599110818328d5
+size 3734
diff --git a/eggplant/ep022/example.md b/eggplant/ep022/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..1dbb99a69dd64266bc06df0b16fc914a4352529c
--- /dev/null
+++ b/eggplant/ep022/example.md
@@ -0,0 +1,32 @@
+# t004__ep022__eggplant
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the eggplant off the plate
+- **success:** False
+- **control steps:** 300  (6.0s at 50 Hz)
+- **policy inferences:** 53  (replan_steps=5)
+- **video:** `eggplant/ep022/rollout.mp4`
+- **features:** `53` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–30 | lowering gripper to grasp eggplant |
+| 3.0–6.0 | 151–299 | 30–59 | attempting to lift eggplant from plate |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 2.8031809145129225 s (00:02.8)
+- **frame / control_step:** 141 (policy_step=28, chunk_index=1)
+- **coarse onset:** 2.8 s (refined=False)
+- **mode:** grasp_failure
+- **reason:** The robot end-effector attempted to grasp the eggplant on the plate but failed to secure a firm grip, leaving the eggplant resting on the plate.
+  Summary: failed grasp on eggplant; gripper slips
+- **recovery:** Re-open the gripper, center the fingers over the eggplant, and re-attempt to grasp and lift it off the plate.
+  Summary: re-open gripper; align fingers; re-grasp eggplant
+- **prevention:** Lower the end-effector further and align the gripper jaw carefully with the body of the eggplant prior to closing the fingers.
+  Summary: position gripper lower; align around eggplant; secure firm grasp
+- **justification:** At around 2.8 seconds, the robot closes its gripper over the eggplant but fails to establish a solid hold, making it impossible to complete the task of removing the eggplant from the plate.
diff --git a/eggplant/ep022/labels.json b/eggplant/ep022/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..aa9045ec93408342e6979ac40933ac2cb5174d62
--- /dev/null
+++ b/eggplant/ep022/labels.json
@@ -0,0 +1,87 @@
+{
+  "rollout_id": "t004__ep022__eggplant",
+  "scene_variant": "",
+  "task_id": 4,
+  "task_file": "",
+  "instruction": "Take the eggplant off the plate",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:41:11+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.001988071570577,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 30,
+      "phrase": "lowering gripper to grasp eggplant",
+      "description": "lowering gripper to grasp eggplant"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.001988071570577,
+      "t_end_sec": 5.964214711729623,
+      "control_step_start": 151,
+      "control_step_end": 299,
+      "policy_step_start": 30,
+      "policy_step_end": 59,
+      "phrase": "attempting to lift eggplant from plate",
+      "description": "attempting to lift eggplant from plate"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 2.8031809145129225,
+    "vlm_failure_onset_timestamp": "00:02.8",
+    "vlm_failure_onset_frame": 141,
+    "vlm_failure_onset_step": 141,
+    "vlm_coarse_failure_onset_seconds": 2.8,
+    "vlm_failure_window_start_seconds": 2.3,
+    "vlm_failure_window_end_seconds": 3.3,
+    "vlm_temporal_refined": false,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot end-effector attempted to grasp the eggplant on the plate but failed to secure a firm grip, leaving the eggplant resting on the plate.",
+    "vlm_failure_reason_summary": "failed grasp on eggplant; gripper slips",
+    "vlm_recovery_action": "Re-open the gripper, center the fingers over the eggplant, and re-attempt to grasp and lift it off the plate.",
+    "vlm_recovery_action_summary": "re-open gripper; align fingers; re-grasp eggplant",
+    "vlm_prevention_action": "Lower the end-effector further and align the gripper jaw carefully with the body of the eggplant prior to closing the fingers.",
+    "vlm_prevention_action_summary": "position gripper lower; align around eggplant; secure firm grasp",
+    "vlm_justification": "At around 2.8 seconds, the robot closes its gripper over the eggplant but fails to establish a solid hold, making it impossible to complete the task of removing the eggplant from the plate.",
+    "vlm_temporal_justification": "The robot gripper fully closes its fingers without securing the eggplant, leaving it resting on the plate as the grasp fails.",
+    "vlm_video_fps": 50.3,
+    "vlm_video_num_frames": 300,
+    "vlm_video_duration_seconds": 5.9443339960238575,
+    "vlm_usage": {
+      "input_tokens": 20688,
+      "output_tokens": 346,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 2.8,\n \"failure_window_start_seconds\": 2.3,\n \"failure_window_end_seconds\": 3.3,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot end-effector attempted to grasp the eggplant on the plate but failed to secure a firm grip, leaving the eggplant resting on the plate.\",\n \"failure_reason_summary\": \"failed grasp on eggplant; gripper slips\",\n \"recovery_action\": \"Re-open the gripper, center the fingers over the eggplant, and re-attempt to grasp and lift it off the plate.\",\n \"recovery_action_summary\": \"re-open gripper; align fingers; re-grasp eggplant\",\n \"prevention_action\": \"Lower the end-effector further and align the gripper jaw carefully with the body of the eggplant prior to closing the fingers.\",\n \"prevention_action_summary\": \"position gripper lower; align around eggplant; secure firm grasp\",\n \"justification\": \"At around 2.8 seconds, the robot closes its gripper over the eggplant but fails to establish a solid hold, making it impossible to complete the task of removing the eggplant from the plate.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 141,
+    "failure_sim_step": 141,
+    "failure_policy_step": 28,
+    "failure_chunk_index": 1,
+    "first_post_failure_policy_step": 29,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-open the gripper, center the fingers over the eggplant, and re-attempt to grasp and lift it off the plate."
+  }
+}
\ No newline at end of file
diff --git a/eggplant/ep022/labels.npz b/eggplant/ep022/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..b7aeaca8c66aad2ac5ad3b0e56d6395894cddf4a
--- /dev/null
+++ b/eggplant/ep022/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f66994b21196e05e3c3b91a6a3e853eb16cb721cc48edb0ac0c7f90c9ff8bc6c
+size 3728
diff --git a/eggplant/ep023/example.md b/eggplant/ep023/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..12767b2d01472d5601855990034a51f592a6812c
--- /dev/null
+++ b/eggplant/ep023/example.md
@@ -0,0 +1,35 @@
+# t004__ep023__eggplant
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the eggplant off the plate
+- **success:** False
+- **control steps:** 674  (13.3s at 51 Hz)
+- **policy inferences:** 117  (replan_steps=6)
+- **video:** `eggplant/ep023/rollout.mp4`
+- **features:** `117` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | lowering gripper toward eggplant |
+| 3.0–6.0 | 152–303 | 25–50 | grasping eggplant and losing grip |
+| 6.0–9.0 | 304–455 | 50–75 | repositioning gripper above plate |
+| 9.0–12.0 | 456–607 | 76–101 | hovering near eggplant |
+| 12.0–13.3 | 608–673 | 101–112 | remaining idle above plate |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 4.970414201183432 s (00:05.0)
+- **frame / control_step:** 252 (policy_step=42, chunk_index=0)
+- **coarse onset:** 3.8 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the eggplant on the plate, but the gripper failed to establish a secure grip, causing the eggplant to slip out and fall back onto the plate.
+  Summary: failed grasp attempt; object slipped from gripper
+- **recovery:** Re-align the end-effector over the lying eggplant, close the gripper securely around its body, and lift it clear of the plate.
+  Summary: realign gripper; securely grasp eggplant; lift off plate
+- **prevention:** Position the gripper centered around the middle of the eggplant with proper orientation and force before attempting to lift.
+  Summary: center gripper on eggplant; secure firm grip
+- **justification:** At around 3.8 seconds, the gripper attempts to close around the eggplant but loses its hold almost immediately, letting the eggplant slip back onto the plate. The robot fails to lift or remove the eggplant for the remainder of the video.
diff --git a/eggplant/ep023/labels.json b/eggplant/ep023/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..d3939824ee409dbb335850ff02259cc08d556e9a
--- /dev/null
+++ b/eggplant/ep023/labels.json
@@ -0,0 +1,120 @@
+{
+  "rollout_id": "t004__ep023__eggplant",
+  "scene_variant": "",
+  "task_id": 4,
+  "task_file": "",
+  "instruction": "Take the eggplant off the plate",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:41:47+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9980276134122286,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "lowering gripper toward eggplant",
+      "description": "lowering gripper toward eggplant"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9980276134122286,
+      "t_end_sec": 5.996055226824457,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "grasping eggplant and losing grip",
+      "description": "grasping eggplant and losing grip"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.996055226824457,
+      "t_end_sec": 8.994082840236686,
+      "control_step_start": 304,
+      "control_step_end": 455,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "repositioning gripper above plate",
+      "description": "repositioning gripper above plate"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.994082840236686,
+      "t_end_sec": 11.992110453648914,
+      "control_step_start": 456,
+      "control_step_end": 607,
+      "policy_step_start": 76,
+      "policy_step_end": 101,
+      "phrase": "hovering near eggplant",
+      "description": "hovering near eggplant"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 11.992110453648914,
+      "t_end_sec": 13.293885601577909,
+      "control_step_start": 608,
+      "control_step_end": 673,
+      "policy_step_start": 101,
+      "policy_step_end": 112,
+      "phrase": "remaining idle above plate",
+      "description": "remaining idle above plate"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 4.970414201183432,
+    "vlm_failure_onset_timestamp": "00:05.0",
+    "vlm_failure_onset_frame": 252,
+    "vlm_failure_onset_step": 252,
+    "vlm_coarse_failure_onset_seconds": 3.8,
+    "vlm_failure_window_start_seconds": 3.3,
+    "vlm_failure_window_end_seconds": 4.8,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the eggplant on the plate, but the gripper failed to establish a secure grip, causing the eggplant to slip out and fall back onto the plate.",
+    "vlm_failure_reason_summary": "failed grasp attempt; object slipped from gripper",
+    "vlm_recovery_action": "Re-align the end-effector over the lying eggplant, close the gripper securely around its body, and lift it clear of the plate.",
+    "vlm_recovery_action_summary": "realign gripper; securely grasp eggplant; lift off plate",
+    "vlm_prevention_action": "Position the gripper centered around the middle of the eggplant with proper orientation and force before attempting to lift.",
+    "vlm_prevention_action_summary": "center gripper on eggplant; secure firm grip",
+    "vlm_justification": "At around 3.8 seconds, the gripper attempts to close around the eggplant but loses its hold almost immediately, letting the eggplant slip back onto the plate. The robot fails to lift or remove the eggplant for the remainder of the video.",
+    "vlm_temporal_justification": "At 03:31 (clip second 211), the robot arm begins lifting upward, but the gripper fails to maintain a secure hold on the eggplant, causing it to slip out and drop back onto the plate.",
+    "vlm_video_fps": 50.7,
+    "vlm_video_num_frames": 674,
+    "vlm_video_duration_seconds": 13.274161735700197,
+    "vlm_usage": {
+      "input_tokens": 21895,
+      "output_tokens": 377,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 3.8,\n \"failure_window_start_seconds\": 3.3,\n \"failure_window_end_seconds\": 4.8,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the eggplant on the plate, but the gripper failed to establish a secure grip, causing the eggplant to slip out and fall back onto the plate.\",\n \"failure_reason_summary\": \"failed grasp attempt; object slipped from gripper\",\n \"recovery_action\": \"Re-align the end-effector over the lying eggplant, close the gripper securely around its body, and lift it clear of the plate.\",\n \"recovery_action_summary\": \"realign gripper; securely grasp eggplant; lift off plate\",\n \"prevention_action\": \"Position the gripper centered around the middle of the eggplant with proper orientation and force before attempting to lift.\",\n \"prevention_action_summary\": \"center gripper on eggplant; secure firm grip\",\n \"justification\": \"At around 3.8 seconds, the gripper attempts to close around the eggplant but loses its hold almost immediately, letting the eggplant slip back onto the plate. The robot fails to lift or remove the eggplant for the remainder of the video.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 252,
+    "failure_sim_step": 252,
+    "failure_policy_step": 42,
+    "failure_chunk_index": 0,
+    "first_post_failure_policy_step": 43,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-align the end-effector over the lying eggplant, close the gripper securely around its body, and lift it clear of the plate."
+  }
+}
\ No newline at end of file
diff --git a/eggplant/ep023/labels.npz b/eggplant/ep023/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..288044af9ba3ba428baaed2f9f5b81c24aed8b49
--- /dev/null
+++ b/eggplant/ep023/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d74f34e4ccba189c9148ce93af78155c92ffcfb13ffaed8cd849830f128f3864
+size 3875
diff --git a/eggplant/ep024/example.md b/eggplant/ep024/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..2375ec651978b967748cd9e9014041db8f671ae2
--- /dev/null
+++ b/eggplant/ep024/example.md
@@ -0,0 +1,34 @@
+# t004__ep024__eggplant
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the eggplant off the plate
+- **success:** False
+- **control steps:** 501  (9.9s at 51 Hz)
+- **policy inferences:** 87  (replan_steps=6)
+- **video:** `eggplant/ep024/rollout.mp4`
+- **features:** `87` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching for eggplant; attempting grasp |
+| 3.0–6.0 | 152–303 | 25–50 | lifting up; attempting second grasp |
+| 6.0–9.0 | 304–455 | 50–75 | closing gripper; missing eggplant |
+| 9.0–9.9 | 456–500 | 76–83 | hovering above plate with empty gripper |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 3.3070866141732287 s (00:03.3)
+- **frame / control_step:** 168 (policy_step=28, chunk_index=0)
+- **coarse onset:** 2.5 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the eggplant but failed to close its gripper securely around it, lifting away empty while the eggplant remained on the plate.
+  Summary: failed grasp attempt; eggplant left on plate
+- **recovery:** Reposition the gripper directly over the center of the eggplant, lower the arm, close the fingers firmly around the object, and lift it off the plate.
+  Summary: reposition gripper; grasp eggplant securely; lift off plate
+- **prevention:** Align the gripper precisely with the body of the eggplant and ensure adequate grasp depth before closing the fingers.
+  Summary: align gripper with eggplant body; ensure proper grasp depth
+- **justification:** At 00:02.5, the robot reaches down to grasp the eggplant but fails to secure it, lifting up empty by 00:03. Subsequent grasp attempts also fail, making this initial failed grasp the point where the failure begins.
diff --git a/eggplant/ep024/labels.json b/eggplant/ep024/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..6edd29befe4e65a2b3e996005ed21f55db868bfa
--- /dev/null
+++ b/eggplant/ep024/labels.json
@@ -0,0 +1,109 @@
+{
+  "rollout_id": "t004__ep024__eggplant",
+  "scene_variant": "",
+  "task_id": 4,
+  "task_file": "",
+  "instruction": "Take the eggplant off the plate",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:42:23+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9921259842519685,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for eggplant; attempting grasp",
+      "description": "reaching for eggplant; attempting grasp"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9921259842519685,
+      "t_end_sec": 5.984251968503937,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "lifting up; attempting second grasp",
+      "description": "lifting up; attempting second grasp"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.984251968503937,
+      "t_end_sec": 8.976377952755906,
+      "control_step_start": 304,
+      "control_step_end": 455,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "closing gripper; missing eggplant",
+      "description": "closing gripper; missing eggplant"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.976377952755906,
+      "t_end_sec": 9.86220472440945,
+      "control_step_start": 456,
+      "control_step_end": 500,
+      "policy_step_start": 76,
+      "policy_step_end": 83,
+      "phrase": "hovering above plate with empty gripper",
+      "description": "hovering above plate with empty gripper"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 3.3070866141732287,
+    "vlm_failure_onset_timestamp": "00:03.3",
+    "vlm_failure_onset_frame": 168,
+    "vlm_failure_onset_step": 168,
+    "vlm_coarse_failure_onset_seconds": 2.5,
+    "vlm_failure_window_start_seconds": 2.0,
+    "vlm_failure_window_end_seconds": 3.0,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the eggplant but failed to close its gripper securely around it, lifting away empty while the eggplant remained on the plate.",
+    "vlm_failure_reason_summary": "failed grasp attempt; eggplant left on plate",
+    "vlm_recovery_action": "Reposition the gripper directly over the center of the eggplant, lower the arm, close the fingers firmly around the object, and lift it off the plate.",
+    "vlm_recovery_action_summary": "reposition gripper; grasp eggplant securely; lift off plate",
+    "vlm_prevention_action": "Align the gripper precisely with the body of the eggplant and ensure adequate grasp depth before closing the fingers.",
+    "vlm_prevention_action_summary": "align gripper with eggplant body; ensure proper grasp depth",
+    "vlm_justification": "At 00:02.5, the robot reaches down to grasp the eggplant but fails to secure it, lifting up empty by 00:03. Subsequent grasp attempts also fail, making this initial failed grasp the point where the failure begins.",
+    "vlm_temporal_justification": "The robot positions its gripper around the eggplant and attempts to close it around 02:45. By 02:48 (frame 168), the gripper fingers close completely empty without securing the eggplant, making the grasp failure visibly apparent before the arm begins lifting.",
+    "vlm_video_fps": 50.8,
+    "vlm_video_num_frames": 501,
+    "vlm_video_duration_seconds": 9.84251968503937,
+    "vlm_usage": {
+      "input_tokens": 20120,
+      "output_tokens": 392,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 2.5,\n \"failure_window_start_seconds\": 2.0,\n \"failure_window_end_seconds\": 3.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the eggplant but failed to close its gripper securely around it, lifting away empty while the eggplant remained on the plate.\",\n \"failure_reason_summary\": \"failed grasp attempt; eggplant left on plate\",\n \"recovery_action\": \"Reposition the gripper directly over the center of the eggplant, lower the arm, close the fingers firmly around the object, and lift it off the plate.\",\n \"recovery_action_summary\": \"reposition gripper; grasp eggplant securely; lift off plate\",\n \"prevention_action\": \"Align the gripper precisely with the body of the eggplant and ensure adequate grasp depth before closing the fingers.\",\n \"prevention_action_summary\": \"align gripper with eggplant body; ensure proper grasp depth\",\n \"justification\": \"At 00:02.5, the robot reaches down to grasp the eggplant but fails to secure it, lifting up empty by 00:03. Subsequent grasp attempts also fail, making this initial failed grasp the point where the failure begins.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 168,
+    "failure_sim_step": 168,
+    "failure_policy_step": 28,
+    "failure_chunk_index": 0,
+    "first_post_failure_policy_step": 29,
+    "failure_type": "grasp_failure",
+    "correction_action": "Reposition the gripper directly over the center of the eggplant, lower the arm, close the fingers firmly around the object, and lift it off the plate."
+  }
+}
\ No newline at end of file
diff --git a/eggplant/ep024/labels.npz b/eggplant/ep024/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..93f853c79246341d7cc87d472be82fadda870cfd
--- /dev/null
+++ b/eggplant/ep024/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a7893897e8c6fc34c3d4941128c8bd8d18c9c81f12d64a308135b08364d81599
+size 3895
diff --git a/eggplant/ep025/example.md b/eggplant/ep025/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..b5911919450ff99a41acd91151a015a83189a398
--- /dev/null
+++ b/eggplant/ep025/example.md
@@ -0,0 +1,21 @@
+# t004__ep025__eggplant
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the eggplant off the plate
+- **success:** True
+- **control steps:** 419  (8.3s at 51 Hz)
+- **policy inferences:** 73  (replan_steps=6)
+- **video:** `eggplant/ep025/rollout.mp4`
+- **features:** `73` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching for the eggplant |
+| 3.0–6.0 | 152–303 | 25–50 | grasping and moving the eggplant |
+| 6.0–8.3 | 304–418 | 50–69 | releasing eggplant on table |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/eggplant/ep025/labels.json b/eggplant/ep025/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..fc8efaed77d2eafee605b99561a692f651884d8b
--- /dev/null
+++ b/eggplant/ep025/labels.json
@@ -0,0 +1,68 @@
+{
+  "rollout_id": "t004__ep025__eggplant",
+  "scene_variant": "",
+  "task_id": 4,
+  "task_file": "",
+  "instruction": "Take the eggplant off the plate",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:42:30+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9980276134122286,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for the eggplant",
+      "description": "reaching for the eggplant"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9980276134122286,
+      "t_end_sec": 5.996055226824457,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "grasping and moving the eggplant",
+      "description": "grasping and moving the eggplant"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.996055226824457,
+      "t_end_sec": 8.264299802761341,
+      "control_step_start": 304,
+      "control_step_end": 418,
+      "policy_step_start": 50,
+      "policy_step_end": 69,
+      "phrase": "releasing eggplant on table",
+      "description": "releasing eggplant on table"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/eggplant/ep025/labels.npz b/eggplant/ep025/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..c12304e4303b8c47d72276e938cd415873b0a4a7
--- /dev/null
+++ b/eggplant/ep025/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4b0d1d3068cd926353bd25541aecb3c26c103502af9d63865f9f60b064ddc304
+size 3082
diff --git a/eggplant/ep026/example.md b/eggplant/ep026/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..79349ca4274148b44957cfed439b8a961dc8079c
--- /dev/null
+++ b/eggplant/ep026/example.md
@@ -0,0 +1,22 @@
+# t004__ep026__eggplant
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the eggplant off the plate
+- **success:** True
+- **control steps:** 463  (9.2s at 50 Hz)
+- **policy inferences:** 79  (replan_steps=6)
+- **video:** `eggplant/ep026/rollout.mp4`
+- **features:** `79` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching for the eggplant |
+| 3.0–6.0 | 152–303 | 25–50 | lifting eggplant off plate |
+| 6.0–9.0 | 304–455 | 50–75 | placing eggplant and retracting arm |
+| 9.0–9.2 | 456–462 | 76–77 | idle above table |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/eggplant/ep026/labels.json b/eggplant/ep026/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..7a5a56d14a1f693c829aea911a47897811bb0e19
--- /dev/null
+++ b/eggplant/ep026/labels.json
@@ -0,0 +1,79 @@
+{
+  "rollout_id": "t004__ep026__eggplant",
+  "scene_variant": "",
+  "task_id": 4,
+  "task_file": "",
+  "instruction": "Take the eggplant off the plate",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:42:39+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00990099009901,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for the eggplant",
+      "description": "reaching for the eggplant"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00990099009901,
+      "t_end_sec": 6.01980198019802,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "lifting eggplant off plate",
+      "description": "lifting eggplant off plate"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01980198019802,
+      "t_end_sec": 9.029702970297029,
+      "control_step_start": 304,
+      "control_step_end": 455,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "placing eggplant and retracting arm",
+      "description": "placing eggplant and retracting arm"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.029702970297029,
+      "t_end_sec": 9.168316831683168,
+      "control_step_start": 456,
+      "control_step_end": 462,
+      "policy_step_start": 76,
+      "policy_step_end": 77,
+      "phrase": "idle above table",
+      "description": "idle above table"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/eggplant/ep026/labels.npz b/eggplant/ep026/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..7377788e9cc4c046f0edc4ad9f9191e61689ea72
--- /dev/null
+++ b/eggplant/ep026/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:78ca1562ed6684eaeb435a2e993b4a9ee06f94ae50c320dfa3eca09ce9ae26dc
+size 3123
diff --git a/eggplant/ep027/example.md b/eggplant/ep027/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..faa2576a69872b2aa47cc0d69c2f750e706a5f60
--- /dev/null
+++ b/eggplant/ep027/example.md
@@ -0,0 +1,32 @@
+# t004__ep027__eggplant
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the eggplant off the plate
+- **success:** False
+- **control steps:** 287  (5.7s at 51 Hz)
+- **policy inferences:** 50  (replan_steps=6)
+- **video:** `eggplant/ep027/rollout.mp4`
+- **features:** `50` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching to grasp the eggplant |
+| 3.0–5.7 | 152–286 | 25–47 | lifting empty gripper above plate |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 3.195266272189349 s (00:03.2)
+- **frame / control_step:** 162 (policy_step=27, chunk_index=0)
+- **coarse onset:** 2.7 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot descended towards the eggplant on the plate but failed to establish a secure grip, closing its gripper improperly and lifting away empty.
+  Summary: failed grasp attempt; lifting empty gripper
+- **recovery:** Lower the gripper back down to align with the center of the eggplant, close the gripper firmly around it, and lift it off the plate.
+  Summary: re-align gripper; grasp eggplant; lift off plate
+- **prevention:** Position the gripper directly over the body of the eggplant and close the fingers securely before initiating the upward motion.
+  Summary: align gripper over eggplant; close fingers securely
+- **justification:** At around 2.7 seconds, the robot gripper closes near the eggplant on the plate but misses securing it, subsequently lifting upward without the object and failing the task goal.
diff --git a/eggplant/ep027/labels.json b/eggplant/ep027/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..e3c91acf6d72b5fa2e0db5ce2c0be59dc1911d17
--- /dev/null
+++ b/eggplant/ep027/labels.json
@@ -0,0 +1,87 @@
+{
+  "rollout_id": "t004__ep027__eggplant",
+  "scene_variant": "",
+  "task_id": 4,
+  "task_file": "",
+  "instruction": "Take the eggplant off the plate",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:43:12+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9980276134122286,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching to grasp the eggplant",
+      "description": "reaching to grasp the eggplant"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9980276134122286,
+      "t_end_sec": 5.660749506903353,
+      "control_step_start": 152,
+      "control_step_end": 286,
+      "policy_step_start": 25,
+      "policy_step_end": 47,
+      "phrase": "lifting empty gripper above plate",
+      "description": "lifting empty gripper above plate"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 3.195266272189349,
+    "vlm_failure_onset_timestamp": "00:03.2",
+    "vlm_failure_onset_frame": 162,
+    "vlm_failure_onset_step": 162,
+    "vlm_coarse_failure_onset_seconds": 2.7,
+    "vlm_failure_window_start_seconds": 2.2,
+    "vlm_failure_window_end_seconds": 3.2,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot descended towards the eggplant on the plate but failed to establish a secure grip, closing its gripper improperly and lifting away empty.",
+    "vlm_failure_reason_summary": "failed grasp attempt; lifting empty gripper",
+    "vlm_recovery_action": "Lower the gripper back down to align with the center of the eggplant, close the gripper firmly around it, and lift it off the plate.",
+    "vlm_recovery_action_summary": "re-align gripper; grasp eggplant; lift off plate",
+    "vlm_prevention_action": "Position the gripper directly over the body of the eggplant and close the fingers securely before initiating the upward motion.",
+    "vlm_prevention_action_summary": "align gripper over eggplant; close fingers securely",
+    "vlm_justification": "At around 2.7 seconds, the robot gripper closes near the eggplant on the plate but misses securing it, subsequently lifting upward without the object and failing the task goal.",
+    "vlm_temporal_justification": "At timestamp 02:42 (frame 162), the robot's gripper closes improperly around the eggplant on the plate, failing to establish a secure hold before lifting away empty.",
+    "vlm_video_fps": 50.7,
+    "vlm_video_num_frames": 287,
+    "vlm_video_duration_seconds": 5.6410256410256405,
+    "vlm_usage": {
+      "input_tokens": 20307,
+      "output_tokens": 349,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 2.7,\n \"failure_window_start_seconds\": 2.2,\n \"failure_window_end_seconds\": 3.2,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot descended towards the eggplant on the plate but failed to establish a secure grip, closing its gripper improperly and lifting away empty.\",\n \"failure_reason_summary\": \"failed grasp attempt; lifting empty gripper\",\n \"recovery_action\": \"Lower the gripper back down to align with the center of the eggplant, close the gripper firmly around it, and lift it off the plate.\",\n \"recovery_action_summary\": \"re-align gripper; grasp eggplant; lift off plate\",\n \"prevention_action\": \"Position the gripper directly over the body of the eggplant and close the fingers securely before initiating the upward motion.\",\n \"prevention_action_summary\": \"align gripper over eggplant; close fingers securely\",\n \"justification\": \"At around 2.7 seconds, the robot gripper closes near the eggplant on the plate but misses securing it, subsequently lifting upward without the object and failing the task goal.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 162,
+    "failure_sim_step": 162,
+    "failure_policy_step": 27,
+    "failure_chunk_index": 0,
+    "first_post_failure_policy_step": 28,
+    "failure_type": "grasp_failure",
+    "correction_action": "Lower the gripper back down to align with the center of the eggplant, close the gripper firmly around it, and lift it off the plate."
+  }
+}
\ No newline at end of file
diff --git a/eggplant/ep027/labels.npz b/eggplant/ep027/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..c92cf4e886274040f9eca021e77fa175e9da4ec6
--- /dev/null
+++ b/eggplant/ep027/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8ab21b97694d44411bb7eaa06b0236a9efce34a887825112d0e4245feb97ed6a
+size 3762
diff --git a/eggplant/ep028/example.md b/eggplant/ep028/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..85c5e5aed771f4e29dc6c151530ddda94c621a9e
--- /dev/null
+++ b/eggplant/ep028/example.md
@@ -0,0 +1,33 @@
+# t004__ep028__eggplant
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the eggplant off the plate
+- **success:** False
+- **control steps:** 316  (6.2s at 51 Hz)
+- **policy inferences:** 54  (replan_steps=6)
+- **video:** `eggplant/ep028/rollout.mp4`
+- **features:** `54` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | approaching eggplant on plate |
+| 3.0–6.0 | 152–303 | 25–50 | closing gripper and missing grasp |
+| 6.0–6.2 | 304–315 | 50–52 | hovering above plate |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 4.122287968441815 s (00:04.1)
+- **frame / control_step:** 209 (policy_step=34, chunk_index=5)
+- **coarse onset:** 4.0 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot aligned its gripper poorly over the edge of the eggplant, causing the fingers to close without securing a firm grip and leaving the eggplant on the plate.
+  Summary: improper grasp alignment; failed to secure eggplant
+- **recovery:** Re-center the gripper over the main body of the eggplant, close the fingers firmly around it, and lift it off the plate.
+  Summary: recenter gripper; grasp and lift eggplant
+- **prevention:** Position the end-effector directly over the center of mass of the eggplant before closing the gripper to ensure a successful hold.
+  Summary: align gripper with eggplant center; close fingers securely
+- **justification:** At around 4.0 seconds, the robot attempts to grasp the eggplant, but the fingers close near its edge and slip off without lifting the object, resulting in a grasp failure.
diff --git a/eggplant/ep028/labels.json b/eggplant/ep028/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..1bab3f148c97f30be4ac61d9af3a4326fe93b511
--- /dev/null
+++ b/eggplant/ep028/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t004__ep028__eggplant",
+  "scene_variant": "",
+  "task_id": 4,
+  "task_file": "",
+  "instruction": "Take the eggplant off the plate",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:43:44+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9980276134122286,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "approaching eggplant on plate",
+      "description": "approaching eggplant on plate"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9980276134122286,
+      "t_end_sec": 5.996055226824457,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "closing gripper and missing grasp",
+      "description": "closing gripper and missing grasp"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.996055226824457,
+      "t_end_sec": 6.232741617357002,
+      "control_step_start": 304,
+      "control_step_end": 315,
+      "policy_step_start": 50,
+      "policy_step_end": 52,
+      "phrase": "hovering above plate",
+      "description": "hovering above plate"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 4.122287968441815,
+    "vlm_failure_onset_timestamp": "00:04.1",
+    "vlm_failure_onset_frame": 209,
+    "vlm_failure_onset_step": 209,
+    "vlm_coarse_failure_onset_seconds": 4.0,
+    "vlm_failure_window_start_seconds": 3.5,
+    "vlm_failure_window_end_seconds": 5.0,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot aligned its gripper poorly over the edge of the eggplant, causing the fingers to close without securing a firm grip and leaving the eggplant on the plate.",
+    "vlm_failure_reason_summary": "improper grasp alignment; failed to secure eggplant",
+    "vlm_recovery_action": "Re-center the gripper over the main body of the eggplant, close the fingers firmly around it, and lift it off the plate.",
+    "vlm_recovery_action_summary": "recenter gripper; grasp and lift eggplant",
+    "vlm_prevention_action": "Position the end-effector directly over the center of mass of the eggplant before closing the gripper to ensure a successful hold.",
+    "vlm_prevention_action_summary": "align gripper with eggplant center; close fingers securely",
+    "vlm_justification": "At around 4.0 seconds, the robot attempts to grasp the eggplant, but the fingers close near its edge and slip off without lifting the object, resulting in a grasp failure.",
+    "vlm_temporal_justification": "At timestamp 02:38 (frame 209 of the clip), the robot closes its gripper fingers completely without capturing the eggplant, leaving the object undisturbed on the plate and confirming the grasp failure.",
+    "vlm_video_fps": 50.7,
+    "vlm_video_num_frames": 316,
+    "vlm_video_duration_seconds": 6.213017751479289,
+    "vlm_usage": {
+      "input_tokens": 18929,
+      "output_tokens": 361,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 4.0,\n \"failure_window_start_seconds\": 3.5,\n \"failure_window_end_seconds\": 5.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot aligned its gripper poorly over the edge of the eggplant, causing the fingers to close without securing a firm grip and leaving the eggplant on the plate.\",\n \"failure_reason_summary\": \"improper grasp alignment; failed to secure eggplant\",\n \"recovery_action\": \"Re-center the gripper over the main body of the eggplant, close the fingers firmly around it, and lift it off the plate.\",\n \"recovery_action_summary\": \"recenter gripper; grasp and lift eggplant\",\n \"prevention_action\": \"Position the end-effector directly over the center of mass of the eggplant before closing the gripper to ensure a successful hold.\",\n \"prevention_action_summary\": \"align gripper with eggplant center; close fingers securely\",\n \"justification\": \"At around 4.0 seconds, the robot attempts to grasp the eggplant, but the fingers close near its edge and slip off without lifting the object, resulting in a grasp failure.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 209,
+    "failure_sim_step": 209,
+    "failure_policy_step": 34,
+    "failure_chunk_index": 5,
+    "first_post_failure_policy_step": 35,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-center the gripper over the main body of the eggplant, close the fingers firmly around it, and lift it off the plate."
+  }
+}
\ No newline at end of file
diff --git a/eggplant/ep028/labels.npz b/eggplant/ep028/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..260948d310dff36f9e7cb7086281235604a716da
--- /dev/null
+++ b/eggplant/ep028/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2fb4cb7dd82aae783c788300b15496feac37b31933d604594056cedef2214c00
+size 3796
diff --git a/eggplant/ep029/example.md b/eggplant/ep029/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..507a91270bb980374304f521e25018a0e61a906f
--- /dev/null
+++ b/eggplant/ep029/example.md
@@ -0,0 +1,33 @@
+# t004__ep029__eggplant
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the eggplant off the plate
+- **success:** False
+- **control steps:** 384  (7.6s at 50 Hz)
+- **policy inferences:** 67  (replan_steps=6)
+- **video:** `eggplant/ep029/rollout.mp4`
+- **features:** `67` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | lowering gripper toward eggplant |
+| 3.0–6.0 | 151–301 | 25–50 | grasping eggplant and dropping it |
+| 6.0–7.6 | 302–383 | 50–63 | hovering empty above plate |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 5.447316103379722 s (00:05.4)
+- **frame / control_step:** 274 (policy_step=45, chunk_index=4)
+- **coarse onset:** 4.0 s (refined=True)
+- **mode:** object_displacement
+- **reason:** The robot grasped the eggplant and attempted to lift it, but lost its grip mid-air, dropping the eggplant back onto the plate.
+  Summary: dropping gripped object; unsecure grip during lift
+- **recovery:** Re-align the gripper over the eggplant on the plate, close the fingers firmly around it, and lift it clear of the plate.
+  Summary: re-align gripper; secure eggplant; lift off plate
+- **prevention:** Grasp the eggplant firmly near its center of mass with sufficient force before lifting to prevent it from slipping.
+  Summary: firm center grasp; prevent slippage during lift
+- **justification:** At around 4.0 seconds, the robot lifts the gripper after closing on the eggplant, but the eggplant slips from the gripper and falls back onto the plate by 5.0 seconds.
diff --git a/eggplant/ep029/labels.json b/eggplant/ep029/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..c60db876743fa98694b39a132efab3cf71521c9d
--- /dev/null
+++ b/eggplant/ep029/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t004__ep029__eggplant",
+  "scene_variant": "",
+  "task_id": 4,
+  "task_file": "",
+  "instruction": "Take the eggplant off the plate",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:44:20+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.001988071570577,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "lowering gripper toward eggplant",
+      "description": "lowering gripper toward eggplant"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.001988071570577,
+      "t_end_sec": 6.003976143141154,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "grasping eggplant and dropping it",
+      "description": "grasping eggplant and dropping it"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.003976143141154,
+      "t_end_sec": 7.634194831013917,
+      "control_step_start": 302,
+      "control_step_end": 383,
+      "policy_step_start": 50,
+      "policy_step_end": 63,
+      "phrase": "hovering empty above plate",
+      "description": "hovering empty above plate"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "object_displacement",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 5.447316103379722,
+    "vlm_failure_onset_timestamp": "00:05.4",
+    "vlm_failure_onset_frame": 274,
+    "vlm_failure_onset_step": 274,
+    "vlm_coarse_failure_onset_seconds": 4.0,
+    "vlm_failure_window_start_seconds": 3.5,
+    "vlm_failure_window_end_seconds": 5.0,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot grasped the eggplant and attempted to lift it, but lost its grip mid-air, dropping the eggplant back onto the plate.",
+    "vlm_failure_reason_summary": "dropping gripped object; unsecure grip during lift",
+    "vlm_recovery_action": "Re-align the gripper over the eggplant on the plate, close the fingers firmly around it, and lift it clear of the plate.",
+    "vlm_recovery_action_summary": "re-align gripper; secure eggplant; lift off plate",
+    "vlm_prevention_action": "Grasp the eggplant firmly near its center of mass with sufficient force before lifting to prevent it from slipping.",
+    "vlm_prevention_action_summary": "firm center grasp; prevent slippage during lift",
+    "vlm_justification": "At around 4.0 seconds, the robot lifts the gripper after closing on the eggplant, but the eggplant slips from the gripper and falls back onto the plate by 5.0 seconds.",
+    "vlm_temporal_justification": "The robot attempts to lift the eggplant, but at 03:44 (frame 224 of the clip) the eggplant visibly slips from the gripper's grasp and drops back down onto the plate.",
+    "vlm_video_fps": 50.3,
+    "vlm_video_num_frames": 384,
+    "vlm_video_duration_seconds": 7.614314115308152,
+    "vlm_usage": {
+      "input_tokens": 21445,
+      "output_tokens": 360,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"object_displacement\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 4.0,\n \"failure_window_start_seconds\": 3.5,\n \"failure_window_end_seconds\": 5.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot grasped the eggplant and attempted to lift it, but lost its grip mid-air, dropping the eggplant back onto the plate.\",\n \"failure_reason_summary\": \"dropping gripped object; unsecure grip during lift\",\n \"recovery_action\": \"Re-align the gripper over the eggplant on the plate, close the fingers firmly around it, and lift it clear of the plate.\",\n \"recovery_action_summary\": \"re-align gripper; secure eggplant; lift off plate\",\n \"prevention_action\": \"Grasp the eggplant firmly near its center of mass with sufficient force before lifting to prevent it from slipping.\",\n \"prevention_action_summary\": \"firm center grasp; prevent slippage during lift\",\n \"justification\": \"At around 4.0 seconds, the robot lifts the gripper after closing on the eggplant, but the eggplant slips from the gripper and falls back onto the plate by 5.0 seconds.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 274,
+    "failure_sim_step": 274,
+    "failure_policy_step": 45,
+    "failure_chunk_index": 4,
+    "first_post_failure_policy_step": 46,
+    "failure_type": "object_displacement",
+    "correction_action": "Re-align the gripper over the eggplant on the plate, close the fingers firmly around it, and lift it clear of the plate."
+  }
+}
\ No newline at end of file
diff --git a/eggplant/ep029/labels.npz b/eggplant/ep029/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..bdb9b7d05040b492c8b8a2b8b371ee9226512726
--- /dev/null
+++ b/eggplant/ep029/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3fc6abb41f6119bb218e4e53f544f0579a09a777b2d9b29534b79a075019d83e
+size 3773
diff --git a/eggplant/ep030/example.md b/eggplant/ep030/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..e904d6746b22eb6f24550f149ec9d1a8d3318a60
--- /dev/null
+++ b/eggplant/ep030/example.md
@@ -0,0 +1,22 @@
+# t004__ep030__eggplant
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the eggplant off the plate
+- **success:** True
+- **control steps:** 577  (11.4s at 51 Hz)
+- **policy inferences:** 100  (replan_steps=6)
+- **video:** `eggplant/ep030/rollout.mp4`
+- **features:** `100` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching for the eggplant |
+| 3.0–6.0 | 152–303 | 25–50 | grasping the eggplant |
+| 6.0–9.0 | 304–455 | 50–75 | lifting eggplant off plate |
+| 9.0–11.4 | 456–576 | 76–96 | placing eggplant on table |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/eggplant/ep030/labels.json b/eggplant/ep030/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..6237bca550b0ed259d88b1f90ea7ff9bc00ffb89
--- /dev/null
+++ b/eggplant/ep030/labels.json
@@ -0,0 +1,79 @@
+{
+  "rollout_id": "t004__ep030__eggplant",
+  "scene_variant": "",
+  "task_id": 4,
+  "task_file": "",
+  "instruction": "Take the eggplant off the plate",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:44:29+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9980276134122286,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for the eggplant",
+      "description": "reaching for the eggplant"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9980276134122286,
+      "t_end_sec": 5.996055226824457,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "grasping the eggplant",
+      "description": "grasping the eggplant"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.996055226824457,
+      "t_end_sec": 8.994082840236686,
+      "control_step_start": 304,
+      "control_step_end": 455,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "lifting eggplant off plate",
+      "description": "lifting eggplant off plate"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.994082840236686,
+      "t_end_sec": 11.380670611439841,
+      "control_step_start": 456,
+      "control_step_end": 576,
+      "policy_step_start": 76,
+      "policy_step_end": 96,
+      "phrase": "placing eggplant on table",
+      "description": "placing eggplant on table"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/eggplant/ep030/labels.npz b/eggplant/ep030/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..5eae759b64a46baf471320a62574749e0b51055d
--- /dev/null
+++ b/eggplant/ep030/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7e9a985108e5abd93776f637bedb4c37ae94b43debb948097a68d7de54c534c0
+size 3104
diff --git a/eggplant/ep031/example.md b/eggplant/ep031/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..bee8cabf938a263d79ee1aded505ebe7a85ddcd8
--- /dev/null
+++ b/eggplant/ep031/example.md
@@ -0,0 +1,34 @@
+# t004__ep031__eggplant
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the eggplant off the plate
+- **success:** False
+- **control steps:** 484  (9.6s at 50 Hz)
+- **policy inferences:** 83  (replan_steps=6)
+- **video:** `eggplant/ep031/rollout.mp4`
+- **features:** `83` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | reaching toward eggplant on plate |
+| 3.0–6.0 | 151–301 | 25–50 | closing gripper around eggplant |
+| 6.0–9.0 | 302–452 | 50–75 | holding empty gripper above plate |
+| 9.0–9.6 | 453–483 | 75–80 | remaining stationary near plate |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 3.784860557768924 s (00:03.8)
+- **frame / control_step:** 190 (policy_step=31, chunk_index=4)
+- **coarse onset:** 4.0 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the eggplant on the plate, but the gripper closed improperly and slipped off, failing to pick up the object.
+  Summary: failed grasp on eggplant; object left on plate
+- **recovery:** Open the gripper, re-align the end-effector over the body of the eggplant, and close the gripper securely before lifting.
+  Summary: realign end-effector; grasp eggplant securely
+- **prevention:** Lower the gripper further and align it centrally over the eggplant before closing the fingers to establish a secure hold.
+  Summary: align gripper centrally; establish secure hold
+- **justification:** None
diff --git a/eggplant/ep031/labels.json b/eggplant/ep031/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..e01b9a9f432a2f74f54388d8567e091d5a61ae48
--- /dev/null
+++ b/eggplant/ep031/labels.json
@@ -0,0 +1,109 @@
+{
+  "rollout_id": "t004__ep031__eggplant",
+  "scene_variant": "",
+  "task_id": 4,
+  "task_file": "",
+  "instruction": "Take the eggplant off the plate",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:45:01+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00796812749004,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching toward eggplant on plate",
+      "description": "reaching toward eggplant on plate"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00796812749004,
+      "t_end_sec": 6.01593625498008,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "closing gripper around eggplant",
+      "description": "closing gripper around eggplant"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01593625498008,
+      "t_end_sec": 9.02390438247012,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "holding empty gripper above plate",
+      "description": "holding empty gripper above plate"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.02390438247012,
+      "t_end_sec": 9.641434262948207,
+      "control_step_start": 453,
+      "control_step_end": 483,
+      "policy_step_start": 75,
+      "policy_step_end": 80,
+      "phrase": "remaining stationary near plate",
+      "description": "remaining stationary near plate"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 3.784860557768924,
+    "vlm_failure_onset_timestamp": "00:03.8",
+    "vlm_failure_onset_frame": 190,
+    "vlm_failure_onset_step": 190,
+    "vlm_coarse_failure_onset_seconds": 4.0,
+    "vlm_failure_window_start_seconds": 3.5,
+    "vlm_failure_window_end_seconds": 5.0,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the eggplant on the plate, but the gripper closed improperly and slipped off, failing to pick up the object.",
+    "vlm_failure_reason_summary": "failed grasp on eggplant; object left on plate",
+    "vlm_recovery_action": "Open the gripper, re-align the end-effector over the body of the eggplant, and close the gripper securely before lifting.",
+    "vlm_recovery_action_summary": "realign end-effector; grasp eggplant securely",
+    "vlm_prevention_action": "Lower the gripper further and align it centrally over the eggplant before closing the fingers to establish a secure hold.",
+    "vlm_prevention_action_summary": "align gripper centrally; establish secure hold",
+    "vlm_justification": null,
+    "vlm_temporal_justification": "At second 02:20 in the clip, the robot's gripper finishes closing around the eggplant but slips off, failing to grasp and secure the object on the plate.",
+    "vlm_video_fps": 50.2,
+    "vlm_video_num_frames": 484,
+    "vlm_video_duration_seconds": 9.621513944223107,
+    "vlm_usage": {
+      "input_tokens": 21571,
+      "output_tokens": 303,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 4.0,\n \"failure_window_start_seconds\": 3.5,\n \"failure_window_end_seconds\": 5.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the eggplant on the plate, but the gripper closed improperly and slipped off, failing to pick up the object.\",\n \"failure_reason_summary\": \"failed grasp on eggplant; object left on plate\",\n \"recovery_action\": \"Open the gripper, re-align the end-effector over the body of the eggplant, and close the gripper securely before lifting.\",\n \"recovery_action_summary\": \"realign end-effector; grasp eggplant securely\",\n \"prevention_action\": \"Lower the gripper further and align it centrally over the eggplant before closing the fingers to establish a secure hold.\",\n \"prevention_action_summary\": \"align gripper centrally; establish secure hold\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 190,
+    "failure_sim_step": 190,
+    "failure_policy_step": 31,
+    "failure_chunk_index": 4,
+    "first_post_failure_policy_step": 32,
+    "failure_type": "grasp_failure",
+    "correction_action": "Open the gripper, re-align the end-effector over the body of the eggplant, and close the gripper securely before lifting."
+  }
+}
\ No newline at end of file
diff --git a/eggplant/ep031/labels.npz b/eggplant/ep031/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..d35a9e72b05ba44dc09eb029d74fd01a562107de
--- /dev/null
+++ b/eggplant/ep031/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4e7c54ec8c3e9730a0c2d004c215dfcdc90f5e1bc6a6cfcf8a18c87cedfbee43
+size 3834
diff --git a/eggplant/ep032/example.md b/eggplant/ep032/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..782bf0e65121391b830cee1dc9ba233ae57432da
--- /dev/null
+++ b/eggplant/ep032/example.md
@@ -0,0 +1,33 @@
+# t004__ep032__eggplant
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the eggplant off the plate
+- **success:** False
+- **control steps:** 367  (7.3s at 50 Hz)
+- **policy inferences:** 63  (replan_steps=6)
+- **video:** `eggplant/ep032/rollout.mp4`
+- **features:** `63` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | descending toward eggplant |
+| 3.0–6.0 | 152–303 | 25–50 | attempting grasp and knocking eggplant |
+| 6.0–7.3 | 304–366 | 50–61 | hovering above plate |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 3.2475247524752477 s (00:03.2)
+- **frame / control_step:** 164 (policy_step=27, chunk_index=2)
+- **coarse onset:** 3.0 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot descended onto the eggplant with an incorrect grasp angle and height, pressing down on it and tipping it over onto the plate instead of closing the gripper securely around it.
+  Summary: failed grasp attempt; tipping object over
+- **recovery:** Re-align the end effector above the eggplant lying on the plate, lower the open gripper around its body, close the fingers firmly, and lift it off the plate.
+  Summary: re-align gripper; grasp eggplant firmly; lift off plate
+- **prevention:** Align the gripper properly with the upright eggplant before descending, ensuring the fingers enclose the object securely to achieve a successful pick.
+  Summary: align gripper with eggplant; secure grasp before lifting
+- **justification:** At 3.0 seconds into the rollout, the robot attempts to grasp the eggplant but miscalculates the grasp height and angle, pushing the eggplant over onto the plate and ascending empty at 4.0 seconds.
diff --git a/eggplant/ep032/labels.json b/eggplant/ep032/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..c84a3cafbc6a72f5e63f7f55b479a075960c449f
--- /dev/null
+++ b/eggplant/ep032/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t004__ep032__eggplant",
+  "scene_variant": "",
+  "task_id": 4,
+  "task_file": "",
+  "instruction": "Take the eggplant off the plate",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:45:40+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00990099009901,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "descending toward eggplant",
+      "description": "descending toward eggplant"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00990099009901,
+      "t_end_sec": 6.01980198019802,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "attempting grasp and knocking eggplant",
+      "description": "attempting grasp and knocking eggplant"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01980198019802,
+      "t_end_sec": 7.267326732673268,
+      "control_step_start": 304,
+      "control_step_end": 366,
+      "policy_step_start": 50,
+      "policy_step_end": 61,
+      "phrase": "hovering above plate",
+      "description": "hovering above plate"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 3.2475247524752477,
+    "vlm_failure_onset_timestamp": "00:03.2",
+    "vlm_failure_onset_frame": 164,
+    "vlm_failure_onset_step": 164,
+    "vlm_coarse_failure_onset_seconds": 3.0,
+    "vlm_failure_window_start_seconds": 2.5,
+    "vlm_failure_window_end_seconds": 3.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot descended onto the eggplant with an incorrect grasp angle and height, pressing down on it and tipping it over onto the plate instead of closing the gripper securely around it.",
+    "vlm_failure_reason_summary": "failed grasp attempt; tipping object over",
+    "vlm_recovery_action": "Re-align the end effector above the eggplant lying on the plate, lower the open gripper around its body, close the fingers firmly, and lift it off the plate.",
+    "vlm_recovery_action_summary": "re-align gripper; grasp eggplant firmly; lift off plate",
+    "vlm_prevention_action": "Align the gripper properly with the upright eggplant before descending, ensuring the fingers enclose the object securely to achieve a successful pick.",
+    "vlm_prevention_action_summary": "align gripper with eggplant; secure grasp before lifting",
+    "vlm_justification": "At 3.0 seconds into the rollout, the robot attempts to grasp the eggplant but miscalculates the grasp height and angle, pushing the eggplant over onto the plate and ascending empty at 4.0 seconds.",
+    "vlm_temporal_justification": "At timestamp 02:44 (164 seconds), the robot gripper descends directly onto the top of the eggplant at an incorrect angle and height instead of enveloping it. As it continues to push down, it presses the eggplant into the plate and tips it over.",
+    "vlm_video_fps": 50.5,
+    "vlm_video_num_frames": 367,
+    "vlm_video_duration_seconds": 7.247524752475248,
+    "vlm_usage": {
+      "input_tokens": 21512,
+      "output_tokens": 392,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 3.0,\n \"failure_window_start_seconds\": 2.5,\n \"failure_window_end_seconds\": 3.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot descended onto the eggplant with an incorrect grasp angle and height, pressing down on it and tipping it over onto the plate instead of closing the gripper securely around it.\",\n \"failure_reason_summary\": \"failed grasp attempt; tipping object over\",\n \"recovery_action\": \"Re-align the end effector above the eggplant lying on the plate, lower the open gripper around its body, close the fingers firmly, and lift it off the plate.\",\n \"recovery_action_summary\": \"re-align gripper; grasp eggplant firmly; lift off plate\",\n \"prevention_action\": \"Align the gripper properly with the upright eggplant before descending, ensuring the fingers enclose the object securely to achieve a successful pick.\",\n \"prevention_action_summary\": \"align gripper with eggplant; secure grasp before lifting\",\n \"justification\": \"At 3.0 seconds into the rollout, the robot attempts to grasp the eggplant but miscalculates the grasp height and angle, pushing the eggplant over onto the plate and ascending empty at 4.0 seconds.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 164,
+    "failure_sim_step": 164,
+    "failure_policy_step": 27,
+    "failure_chunk_index": 2,
+    "first_post_failure_policy_step": 28,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-align the end effector above the eggplant lying on the plate, lower the open gripper around its body, close the fingers firmly, and lift it off the plate."
+  }
+}
\ No newline at end of file
diff --git a/eggplant/ep032/labels.npz b/eggplant/ep032/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..6c2aa6c645545bf712183069eff57652c21ceb78
--- /dev/null
+++ b/eggplant/ep032/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ed4808e14416d63912e26ccc4525d78bb035505386081f2440b2b7a59e01c2b9
+size 3899
diff --git a/eggplant/ep033/example.md b/eggplant/ep033/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..f99e7ea1ccc73286a9a10b6526cf24818d6cf516
--- /dev/null
+++ b/eggplant/ep033/example.md
@@ -0,0 +1,21 @@
+# t004__ep033__eggplant
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the eggplant off the plate
+- **success:** True
+- **control steps:** 407  (8.1s at 50 Hz)
+- **policy inferences:** 71  (replan_steps=6)
+- **video:** `eggplant/ep033/rollout.mp4`
+- **features:** `71` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | reaching for and grasping eggplant |
+| 3.0–6.0 | 151–301 | 25–50 | lifting eggplant off plate |
+| 6.0–8.1 | 302–406 | 50–67 | releasing eggplant and retracting arm |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/eggplant/ep033/labels.json b/eggplant/ep033/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..07c35f29bbdc820fffdc161caafb9acda4954a55
--- /dev/null
+++ b/eggplant/ep033/labels.json
@@ -0,0 +1,68 @@
+{
+  "rollout_id": "t004__ep033__eggplant",
+  "scene_variant": "",
+  "task_id": 4,
+  "task_file": "",
+  "instruction": "Take the eggplant off the plate",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:45:48+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.996031746031746,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for and grasping eggplant",
+      "description": "reaching for and grasping eggplant"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.996031746031746,
+      "t_end_sec": 5.992063492063492,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "lifting eggplant off plate",
+      "description": "lifting eggplant off plate"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.992063492063492,
+      "t_end_sec": 8.075396825396826,
+      "control_step_start": 302,
+      "control_step_end": 406,
+      "policy_step_start": 50,
+      "policy_step_end": 67,
+      "phrase": "releasing eggplant and retracting arm",
+      "description": "releasing eggplant and retracting arm"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/eggplant/ep033/labels.npz b/eggplant/ep033/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..749c128ab0b50262b5da483edf569a2c5e35f339
--- /dev/null
+++ b/eggplant/ep033/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:453fa82d02e6b3cd044378fe6a8bd2a4f2c6cafce05f3e5a405c12dd66a593ad
+size 3093
diff --git a/eggplant/ep034/example.md b/eggplant/ep034/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..32e1529c6621db3a2947cce02b95ed6806e2a34b
--- /dev/null
+++ b/eggplant/ep034/example.md
@@ -0,0 +1,33 @@
+# t004__ep034__eggplant
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the eggplant off the plate
+- **success:** False
+- **control steps:** 406  (7.9s at 52 Hz)
+- **policy inferences:** 70  (replan_steps=6)
+- **video:** `eggplant/ep034/rollout.mp4`
+- **features:** `70` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–153 | 0–25 | reaching toward eggplant on plate |
+| 3.0–6.0 | 154–307 | 25–51 | closing gripper; lifting empty arm |
+| 6.0–7.9 | 308–405 | 51–67 | hovering above plate empty-handed |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 4.0 s (00:04.0)
+- **frame / control_step:** 206 (policy_step=34, chunk_index=2)
+- **coarse onset:** 4.0 s (refined=False)
+- **mode:** grasp_failure
+- **reason:** The robot closed its gripper around the eggplant on the plate but failed to establish a firm grip, lifting its end-effector away without the object.
+  Summary: failed grasp on eggplant; gripper lifted empty
+- **recovery:** Lower the gripper back down over the plate, align the fingers around the body of the eggplant, close the gripper firmly, and lift the eggplant off the plate.
+  Summary: realign gripper; grasp eggplant securely; lift off plate
+- **prevention:** Position the open gripper deeper and closer around the thickest section of the eggplant before closing the fingers to secure a stable grip prior to lifting.
+  Summary: center gripper over eggplant; close tightly before lifting
+- **justification:** At approximately 4.0 seconds, the robot attempts to lift the eggplant off the plate after closing its gripper, but the eggplant slips out and remains on the plate while the gripper moves upward empty.
diff --git a/eggplant/ep034/labels.json b/eggplant/ep034/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..3851ba9f5a291265ccad9bfb78376630b0343f9c
--- /dev/null
+++ b/eggplant/ep034/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t004__ep034__eggplant",
+  "scene_variant": "",
+  "task_id": 4,
+  "task_file": "",
+  "instruction": "Take the eggplant off the plate",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:46:27+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9902912621359223,
+      "control_step_start": 0,
+      "control_step_end": 153,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching toward eggplant on plate",
+      "description": "reaching toward eggplant on plate"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9902912621359223,
+      "t_end_sec": 5.980582524271845,
+      "control_step_start": 154,
+      "control_step_end": 307,
+      "policy_step_start": 25,
+      "policy_step_end": 51,
+      "phrase": "closing gripper; lifting empty arm",
+      "description": "closing gripper; lifting empty arm"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.980582524271845,
+      "t_end_sec": 7.883495145631068,
+      "control_step_start": 308,
+      "control_step_end": 405,
+      "policy_step_start": 51,
+      "policy_step_end": 67,
+      "phrase": "hovering above plate empty-handed",
+      "description": "hovering above plate empty-handed"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 4.0,
+    "vlm_failure_onset_timestamp": "00:04.0",
+    "vlm_failure_onset_frame": 206,
+    "vlm_failure_onset_step": 206,
+    "vlm_coarse_failure_onset_seconds": 4.0,
+    "vlm_failure_window_start_seconds": 3.5,
+    "vlm_failure_window_end_seconds": 4.5,
+    "vlm_temporal_refined": false,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot closed its gripper around the eggplant on the plate but failed to establish a firm grip, lifting its end-effector away without the object.",
+    "vlm_failure_reason_summary": "failed grasp on eggplant; gripper lifted empty",
+    "vlm_recovery_action": "Lower the gripper back down over the plate, align the fingers around the body of the eggplant, close the gripper firmly, and lift the eggplant off the plate.",
+    "vlm_recovery_action_summary": "realign gripper; grasp eggplant securely; lift off plate",
+    "vlm_prevention_action": "Position the open gripper deeper and closer around the thickest section of the eggplant before closing the fingers to secure a stable grip prior to lifting.",
+    "vlm_prevention_action_summary": "center gripper over eggplant; close tightly before lifting",
+    "vlm_justification": "At approximately 4.0 seconds, the robot attempts to lift the eggplant off the plate after closing its gripper, but the eggplant slips out and remains on the plate while the gripper moves upward empty.",
+    "vlm_temporal_justification": "The robot completes closing its gripper around 03:28-03:30 without securing the eggplant, and at 03:48 the end-effector begins lifting upward away from the table, leaving the eggplant behind on the plate.",
+    "vlm_video_fps": 51.5,
+    "vlm_video_num_frames": 406,
+    "vlm_video_duration_seconds": 7.864077669902913,
+    "vlm_usage": {
+      "input_tokens": 21825,
+      "output_tokens": 384,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 4.0,\n \"failure_window_start_seconds\": 3.5,\n \"failure_window_end_seconds\": 4.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot closed its gripper around the eggplant on the plate but failed to establish a firm grip, lifting its end-effector away without the object.\",\n \"failure_reason_summary\": \"failed grasp on eggplant; gripper lifted empty\",\n \"recovery_action\": \"Lower the gripper back down over the plate, align the fingers around the body of the eggplant, close the gripper firmly, and lift the eggplant off the plate.\",\n \"recovery_action_summary\": \"realign gripper; grasp eggplant securely; lift off plate\",\n \"prevention_action\": \"Position the open gripper deeper and closer around the thickest section of the eggplant before closing the fingers to secure a stable grip prior to lifting.\",\n \"prevention_action_summary\": \"center gripper over eggplant; close tightly before lifting\",\n \"justification\": \"At approximately 4.0 seconds, the robot attempts to lift the eggplant off the plate after closing its gripper, but the eggplant slips out and remains on the plate while the gripper moves upward empty.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 206,
+    "failure_sim_step": 206,
+    "failure_policy_step": 34,
+    "failure_chunk_index": 2,
+    "first_post_failure_policy_step": 35,
+    "failure_type": "grasp_failure",
+    "correction_action": "Lower the gripper back down over the plate, align the fingers around the body of the eggplant, close the gripper firmly, and lift the eggplant off the plate."
+  }
+}
\ No newline at end of file
diff --git a/eggplant/ep034/labels.npz b/eggplant/ep034/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..aac8a96e340d08eb58ec167c69d0f7d8f827e387
--- /dev/null
+++ b/eggplant/ep034/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d57229aa45cf14830636bdf212c2b560dff1d329653d1600dcc362412d00ae35
+size 3866
diff --git a/eggplant/ep035/example.md b/eggplant/ep035/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..498b7a413380ec29e28dcc7301ef707e574d3769
--- /dev/null
+++ b/eggplant/ep035/example.md
@@ -0,0 +1,33 @@
+# t004__ep035__eggplant
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the eggplant off the plate
+- **success:** False
+- **control steps:** 321  (6.3s at 51 Hz)
+- **policy inferences:** 55  (replan_steps=6)
+- **video:** `eggplant/ep035/rollout.mp4`
+- **features:** `55` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–152 | 0–25 | reaching toward eggplant on plate |
+| 3.0–6.0 | 153–305 | 25–50 | closing gripper on plate edge |
+| 6.0–6.3 | 306–320 | 51–53 | holding position near plate |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 3.1372549019607843 s (00:03.1)
+- **frame / control_step:** 160 (policy_step=26, chunk_index=4)
+- **coarse onset:** 3.0 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot lowered its gripper onto the edge of the plate rather than centering over the eggplant, causing the plate to tip and preventing the gripper from acquiring the eggplant.
+  Summary: improper grasp alignment; plate tipped during grasp
+- **recovery:** Release the plate edge, raise the end-effector, align the gripper directly over the center of the eggplant, and re-attempt the grasp.
+  Summary: re-align gripper over eggplant; execute clean grasp
+- **prevention:** Position the gripper centered above the eggplant's body at a sufficient clearance height before descending to close the fingers.
+  Summary: center grasp on eggplant body; avoid hitting plate edge
+- **justification:** At approximately 3.0 seconds into the video, the robot lowers its end-effector onto the rim of the plate, causing the plate to tilt upward and preventing a successful grasp on the eggplant.
diff --git a/eggplant/ep035/labels.json b/eggplant/ep035/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..78e9475495b4abf643622c6d57bce5f897ee78fa
--- /dev/null
+++ b/eggplant/ep035/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t004__ep035__eggplant",
+  "scene_variant": "",
+  "task_id": 4,
+  "task_file": "",
+  "instruction": "Take the eggplant off the plate",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:46:57+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0,
+      "control_step_start": 0,
+      "control_step_end": 152,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching toward eggplant on plate",
+      "description": "reaching toward eggplant on plate"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0,
+      "t_end_sec": 6.0,
+      "control_step_start": 153,
+      "control_step_end": 305,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "closing gripper on plate edge",
+      "description": "closing gripper on plate edge"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.0,
+      "t_end_sec": 6.294117647058823,
+      "control_step_start": 306,
+      "control_step_end": 320,
+      "policy_step_start": 51,
+      "policy_step_end": 53,
+      "phrase": "holding position near plate",
+      "description": "holding position near plate"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 3.1372549019607843,
+    "vlm_failure_onset_timestamp": "00:03.1",
+    "vlm_failure_onset_frame": 160,
+    "vlm_failure_onset_step": 160,
+    "vlm_coarse_failure_onset_seconds": 3.0,
+    "vlm_failure_window_start_seconds": 2.5,
+    "vlm_failure_window_end_seconds": 3.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot lowered its gripper onto the edge of the plate rather than centering over the eggplant, causing the plate to tip and preventing the gripper from acquiring the eggplant.",
+    "vlm_failure_reason_summary": "improper grasp alignment; plate tipped during grasp",
+    "vlm_recovery_action": "Release the plate edge, raise the end-effector, align the gripper directly over the center of the eggplant, and re-attempt the grasp.",
+    "vlm_recovery_action_summary": "re-align gripper over eggplant; execute clean grasp",
+    "vlm_prevention_action": "Position the gripper centered above the eggplant's body at a sufficient clearance height before descending to close the fingers.",
+    "vlm_prevention_action_summary": "center grasp on eggplant body; avoid hitting plate edge",
+    "vlm_justification": "At approximately 3.0 seconds into the video, the robot lowers its end-effector onto the rim of the plate, causing the plate to tilt upward and preventing a successful grasp on the eggplant.",
+    "vlm_temporal_justification": "At 02:40 (160 seconds into the slowed clip), the robot lowers its gripper directly over the left rim of the plate rather than centering over the eggplant. At 02:41 (161 seconds), the gripper makes contact with the plate rim, causing the plate to tip and preventing a successful grasp.",
+    "vlm_video_fps": 51.0,
+    "vlm_video_num_frames": 321,
+    "vlm_video_duration_seconds": 6.2745098039215685,
+    "vlm_usage": {
+      "input_tokens": 21573,
+      "output_tokens": 395,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 3.0,\n \"failure_window_start_seconds\": 2.5,\n \"failure_window_end_seconds\": 3.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot lowered its gripper onto the edge of the plate rather than centering over the eggplant, causing the plate to tip and preventing the gripper from acquiring the eggplant.\",\n \"failure_reason_summary\": \"improper grasp alignment; plate tipped during grasp\",\n \"recovery_action\": \"Release the plate edge, raise the end-effector, align the gripper directly over the center of the eggplant, and re-attempt the grasp.\",\n \"recovery_action_summary\": \"re-align gripper over eggplant; execute clean grasp\",\n \"prevention_action\": \"Position the gripper centered above the eggplant's body at a sufficient clearance height before descending to close the fingers.\",\n \"prevention_action_summary\": \"center grasp on eggplant body; avoid hitting plate edge\",\n \"justification\": \"At approximately 3.0 seconds into the video, the robot lowers its end-effector onto the rim of the plate, causing the plate to tilt upward and preventing a successful grasp on the eggplant.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 160,
+    "failure_sim_step": 160,
+    "failure_policy_step": 26,
+    "failure_chunk_index": 4,
+    "first_post_failure_policy_step": 27,
+    "failure_type": "grasp_failure",
+    "correction_action": "Release the plate edge, raise the end-effector, align the gripper directly over the center of the eggplant, and re-attempt the grasp."
+  }
+}
\ No newline at end of file
diff --git a/eggplant/ep035/labels.npz b/eggplant/ep035/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..959f298169b735d4a719ee6c541f1de8b28c3c3a
--- /dev/null
+++ b/eggplant/ep035/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:abb5d9aba59a1b72934e6665effb9c7e344877b0f8dea8d55d718578f7437886
+size 3811
diff --git a/eggplant/ep036/example.md b/eggplant/ep036/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..b0767c31e84e2641bed9e3e3303e9f1586f62b32
--- /dev/null
+++ b/eggplant/ep036/example.md
@@ -0,0 +1,21 @@
+# t004__ep036__eggplant
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the eggplant off the plate
+- **success:** True
+- **control steps:** 377  (7.5s at 50 Hz)
+- **policy inferences:** 65  (replan_steps=6)
+- **video:** `eggplant/ep036/rollout.mp4`
+- **features:** `65` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | reaching for the eggplant |
+| 3.0–6.0 | 151–301 | 25–50 | lifting eggplant off plate |
+| 6.0–7.5 | 302–376 | 50–62 | releasing eggplant and retracting |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/eggplant/ep036/labels.json b/eggplant/ep036/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..5f8fbae29e62b77cc4d2da72d3554eacc7bd446f
--- /dev/null
+++ b/eggplant/ep036/labels.json
@@ -0,0 +1,68 @@
+{
+  "rollout_id": "t004__ep036__eggplant",
+  "scene_variant": "",
+  "task_id": 4,
+  "task_file": "",
+  "instruction": "Take the eggplant off the plate",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:47:05+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.996031746031746,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for the eggplant",
+      "description": "reaching for the eggplant"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.996031746031746,
+      "t_end_sec": 5.992063492063492,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "lifting eggplant off plate",
+      "description": "lifting eggplant off plate"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.992063492063492,
+      "t_end_sec": 7.48015873015873,
+      "control_step_start": 302,
+      "control_step_end": 376,
+      "policy_step_start": 50,
+      "policy_step_end": 62,
+      "phrase": "releasing eggplant and retracting",
+      "description": "releasing eggplant and retracting"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/eggplant/ep036/labels.npz b/eggplant/ep036/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..e5553be44632c7f9554b010d932c230c56ad1baa
--- /dev/null
+++ b/eggplant/ep036/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e7ede3be8f6e2a4b0db5e254303236a0dfe44715299fba1c56c9f262cc034d6d
+size 3080
diff --git a/eggplant/ep037/example.md b/eggplant/ep037/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..95886f2d6bb739cd6232fd11a06ae3f27ddfe730
--- /dev/null
+++ b/eggplant/ep037/example.md
@@ -0,0 +1,21 @@
+# t004__ep037__eggplant
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the eggplant off the plate
+- **success:** True
+- **control steps:** 402  (8.0s at 50 Hz)
+- **policy inferences:** 69  (replan_steps=6)
+- **video:** `eggplant/ep037/rollout.mp4`
+- **features:** `69` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | approaching and grasping eggplant |
+| 3.0–6.0 | 151–301 | 25–50 | lifting and moving eggplant |
+| 6.0–8.0 | 302–401 | 50–66 | releasing eggplant and retracting |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/eggplant/ep037/labels.json b/eggplant/ep037/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..d2ff5de3e8740791bbf0e540e63a9e3a4e54493d
--- /dev/null
+++ b/eggplant/ep037/labels.json
@@ -0,0 +1,68 @@
+{
+  "rollout_id": "t004__ep037__eggplant",
+  "scene_variant": "",
+  "task_id": 4,
+  "task_file": "",
+  "instruction": "Take the eggplant off the plate",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:47:14+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.996031746031746,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "approaching and grasping eggplant",
+      "description": "approaching and grasping eggplant"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.996031746031746,
+      "t_end_sec": 5.992063492063492,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "lifting and moving eggplant",
+      "description": "lifting and moving eggplant"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.992063492063492,
+      "t_end_sec": 7.976190476190476,
+      "control_step_start": 302,
+      "control_step_end": 401,
+      "policy_step_start": 50,
+      "policy_step_end": 66,
+      "phrase": "releasing eggplant and retracting",
+      "description": "releasing eggplant and retracting"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/eggplant/ep037/labels.npz b/eggplant/ep037/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..c596b45bb1dcd287611149db9ed9c53d61385f66
--- /dev/null
+++ b/eggplant/ep037/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0ca43727de4d5d2192958f420afe2d808ed5f0b8d0cd15cffb684b7c33787a08
+size 3084
diff --git a/eggplant/ep038/example.md b/eggplant/ep038/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..db32376888e5d92c3beed828d5fe95c909a6af1c
--- /dev/null
+++ b/eggplant/ep038/example.md
@@ -0,0 +1,22 @@
+# t004__ep038__eggplant
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the eggplant off the plate
+- **success:** True
+- **control steps:** 515  (10.2s at 50 Hz)
+- **policy inferences:** 89  (replan_steps=6)
+- **video:** `eggplant/ep038/rollout.mp4`
+- **features:** `89` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching for the eggplant |
+| 3.0–6.0 | 152–303 | 25–50 | grasping and lifting eggplant |
+| 6.0–9.0 | 304–455 | 50–75 | moving eggplant off plate |
+| 9.0–10.2 | 456–514 | 76–85 | retracting robot arm |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/eggplant/ep038/labels.json b/eggplant/ep038/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..b6e5955d1083fe287e55abf5d444f58afd57b135
--- /dev/null
+++ b/eggplant/ep038/labels.json
@@ -0,0 +1,79 @@
+{
+  "rollout_id": "t004__ep038__eggplant",
+  "scene_variant": "",
+  "task_id": 4,
+  "task_file": "",
+  "instruction": "Take the eggplant off the plate",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:47:23+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00990099009901,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for the eggplant",
+      "description": "reaching for the eggplant"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00990099009901,
+      "t_end_sec": 6.01980198019802,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "grasping and lifting eggplant",
+      "description": "grasping and lifting eggplant"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01980198019802,
+      "t_end_sec": 9.029702970297029,
+      "control_step_start": 304,
+      "control_step_end": 455,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "moving eggplant off plate",
+      "description": "moving eggplant off plate"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.029702970297029,
+      "t_end_sec": 10.198019801980198,
+      "control_step_start": 456,
+      "control_step_end": 514,
+      "policy_step_start": 76,
+      "policy_step_end": 85,
+      "phrase": "retracting robot arm",
+      "description": "retracting robot arm"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/eggplant/ep038/labels.npz b/eggplant/ep038/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..ec6c4d48b9a1ae662f09db572a46a775630d3fc8
--- /dev/null
+++ b/eggplant/ep038/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f83b763cee7b453f08d8533a50a998143f71cb82cdc9c96c22460f4d742736d4
+size 3121
diff --git a/eggplant/ep039/example.md b/eggplant/ep039/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..6f39014475c5c3a11b38ad47f9dbd05d09b3f8f6
--- /dev/null
+++ b/eggplant/ep039/example.md
@@ -0,0 +1,36 @@
+# t004__ep039__eggplant
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the eggplant off the plate
+- **success:** False
+- **control steps:** 837  (16.6s at 50 Hz)
+- **policy inferences:** 144  (replan_steps=6)
+- **video:** `eggplant/ep039/rollout.mp4`
+- **features:** `144` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | approaching and attempting grasp |
+| 3.0–6.0 | 151–301 | 25–50 | opening gripper and repositioning |
+| 6.0–9.0 | 302–452 | 50–75 | attempting to grasp eggplant again |
+| 9.0–12.0 | 453–603 | 75–100 | slipping and raising empty gripper |
+| 12.0–15.0 | 604–754 | 100–125 | lowering gripper toward plate |
+| 15.0–16.6 | 755–836 | 125–139 | hovering above plate without eggplant |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 3.134920634920635 s (00:03.1)
+- **frame / control_step:** 158 (policy_step=26, chunk_index=2)
+- **coarse onset:** 3.0 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the eggplant on the plate multiple times, but failed to secure a firm grip, leaving the eggplant on the plate.
+  Summary: failed to grasp eggplant; object remained on plate
+- **recovery:** Re-align the gripper directly over the center of mass of the eggplant, close the fingers tightly around it, and lift it clear of the plate.
+  Summary: re-align gripper over eggplant; grasp securely; lift off plate
+- **prevention:** Position the end-effector accurately over the thickest part of the eggplant and close the gripper with sufficient force to pick it up in a single attempt.
+  Summary: align gripper properly; grasp eggplant firmly; lift away
+- **justification:** The robot repeatedly closes its gripper around the eggplant starting around 3.0 seconds, but fails to secure a hold or lift it off the plate throughout the episode.
diff --git a/eggplant/ep039/labels.json b/eggplant/ep039/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..a9c90ec3c1fdf348fd28c3011f5c46bb8adcd9a9
--- /dev/null
+++ b/eggplant/ep039/labels.json
@@ -0,0 +1,131 @@
+{
+  "rollout_id": "t004__ep039__eggplant",
+  "scene_variant": "",
+  "task_id": 4,
+  "task_file": "",
+  "instruction": "Take the eggplant off the plate",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:48:01+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.996031746031746,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "approaching and attempting grasp",
+      "description": "approaching and attempting grasp"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.996031746031746,
+      "t_end_sec": 5.992063492063492,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "opening gripper and repositioning",
+      "description": "opening gripper and repositioning"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.992063492063492,
+      "t_end_sec": 8.988095238095239,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "attempting to grasp eggplant again",
+      "description": "attempting to grasp eggplant again"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.988095238095239,
+      "t_end_sec": 11.984126984126984,
+      "control_step_start": 453,
+      "control_step_end": 603,
+      "policy_step_start": 75,
+      "policy_step_end": 100,
+      "phrase": "slipping and raising empty gripper",
+      "description": "slipping and raising empty gripper"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 11.984126984126984,
+      "t_end_sec": 14.980158730158731,
+      "control_step_start": 604,
+      "control_step_end": 754,
+      "policy_step_start": 100,
+      "policy_step_end": 125,
+      "phrase": "lowering gripper toward plate",
+      "description": "lowering gripper toward plate"
+    },
+    {
+      "segment_index": 5,
+      "t_start_sec": 14.980158730158731,
+      "t_end_sec": 16.607142857142858,
+      "control_step_start": 755,
+      "control_step_end": 836,
+      "policy_step_start": 125,
+      "policy_step_end": 139,
+      "phrase": "hovering above plate without eggplant",
+      "description": "hovering above plate without eggplant"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 3.134920634920635,
+    "vlm_failure_onset_timestamp": "00:03.1",
+    "vlm_failure_onset_frame": 158,
+    "vlm_failure_onset_step": 158,
+    "vlm_coarse_failure_onset_seconds": 3.0,
+    "vlm_failure_window_start_seconds": 2.5,
+    "vlm_failure_window_end_seconds": 3.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the eggplant on the plate multiple times, but failed to secure a firm grip, leaving the eggplant on the plate.",
+    "vlm_failure_reason_summary": "failed to grasp eggplant; object remained on plate",
+    "vlm_recovery_action": "Re-align the gripper directly over the center of mass of the eggplant, close the fingers tightly around it, and lift it clear of the plate.",
+    "vlm_recovery_action_summary": "re-align gripper over eggplant; grasp securely; lift off plate",
+    "vlm_prevention_action": "Position the end-effector accurately over the thickest part of the eggplant and close the gripper with sufficient force to pick it up in a single attempt.",
+    "vlm_prevention_action_summary": "align gripper properly; grasp eggplant firmly; lift away",
+    "vlm_justification": "The robot repeatedly closes its gripper around the eggplant starting around 3.0 seconds, but fails to secure a hold or lift it off the plate throughout the episode.",
+    "vlm_temporal_justification": "At 02:38 (frame/second 158), the robot's gripper fingers finish closing without grasping or enclosing the eggplant, leaving it sitting untouched on the plate.",
+    "vlm_video_fps": 50.4,
+    "vlm_video_num_frames": 837,
+    "vlm_video_duration_seconds": 16.58730158730159,
+    "vlm_usage": {
+      "input_tokens": 22013,
+      "output_tokens": 365,
+      "cached_tokens": 1157
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 3.0,\n \"failure_window_start_seconds\": 2.5,\n \"failure_window_end_seconds\": 3.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the eggplant on the plate multiple times, but failed to secure a firm grip, leaving the eggplant on the plate.\",\n \"failure_reason_summary\": \"failed to grasp eggplant; object remained on plate\",\n \"recovery_action\": \"Re-align the gripper directly over the center of mass of the eggplant, close the fingers tightly around it, and lift it clear of the plate.\",\n \"recovery_action_summary\": \"re-align gripper over eggplant; grasp securely; lift off plate\",\n \"prevention_action\": \"Position the end-effector accurately over the thickest part of the eggplant and close the gripper with sufficient force to pick it up in a single attempt.\",\n \"prevention_action_summary\": \"align gripper properly; grasp eggplant firmly; lift away\",\n \"justification\": \"The robot repeatedly closes its gripper around the eggplant starting around 3.0 seconds, but fails to secure a hold or lift it off the plate throughout the episode.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 158,
+    "failure_sim_step": 158,
+    "failure_policy_step": 26,
+    "failure_chunk_index": 2,
+    "first_post_failure_policy_step": 27,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-align the gripper directly over the center of mass of the eggplant, close the fingers tightly around it, and lift it clear of the plate."
+  }
+}
\ No newline at end of file
diff --git a/eggplant/ep039/labels.npz b/eggplant/ep039/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..51374fa674623d441d2cc5febf8b819d49452a37
--- /dev/null
+++ b/eggplant/ep039/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:61ef3f8c8b67795668bf54c840e81c6a4769f899ecc398a70833e4f2ba4ffe68
+size 3979
diff --git a/eggplant/ep040/example.md b/eggplant/ep040/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..67df0b6a84f57a0f003322135b05e570bc6fab6a
--- /dev/null
+++ b/eggplant/ep040/example.md
@@ -0,0 +1,33 @@
+# t004__ep040__eggplant
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the eggplant off the plate
+- **success:** False
+- **control steps:** 392  (7.8s at 50 Hz)
+- **policy inferences:** 67  (replan_steps=6)
+- **video:** `eggplant/ep040/rollout.mp4`
+- **features:** `67` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching and attempting grasp on eggplant |
+| 3.0–6.0 | 152–303 | 25–50 | hovering empty gripper above plate |
+| 6.0–7.8 | 304–391 | 50–65 | idle above plate with open gripper |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 2.9702970297029703 s (00:03.0)
+- **frame / control_step:** 150 (policy_step=25, chunk_index=0)
+- **coarse onset:** 2.5 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot aligned poorly with the eggplant on the plate and closed its gripper without securing a firm hold, failing to pick it up.
+  Summary: failed grasp on eggplant; gripper closed empty
+- **recovery:** Re-align the gripper directly over the body of the eggplant, descend to a lower grasp depth, and close the fingers firmly around the eggplant before lifting.
+  Summary: reposition gripper over eggplant; close grasp securely; lift off plate
+- **prevention:** Ensure accurate depth and centering relative to the eggplant before closing the gripper so the fingers clamp securely around the object.
+  Summary: align gripper depth and position; close fingers around eggplant body
+- **justification:** The robot approaches the plate around 2.0s and attempts to grasp the eggplant at 2.5s, but closes its gripper without securing the object and lifts away empty by 3.0s.
diff --git a/eggplant/ep040/labels.json b/eggplant/ep040/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..af58042273a358d33a38347bb3f5ff6552439b86
--- /dev/null
+++ b/eggplant/ep040/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t004__ep040__eggplant",
+  "scene_variant": "",
+  "task_id": 4,
+  "task_file": "",
+  "instruction": "Take the eggplant off the plate",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:48:30+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00990099009901,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching and attempting grasp on eggplant",
+      "description": "reaching and attempting grasp on eggplant"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00990099009901,
+      "t_end_sec": 6.01980198019802,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "hovering empty gripper above plate",
+      "description": "hovering empty gripper above plate"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01980198019802,
+      "t_end_sec": 7.762376237623762,
+      "control_step_start": 304,
+      "control_step_end": 391,
+      "policy_step_start": 50,
+      "policy_step_end": 65,
+      "phrase": "idle above plate with open gripper",
+      "description": "idle above plate with open gripper"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 2.9702970297029703,
+    "vlm_failure_onset_timestamp": "00:03.0",
+    "vlm_failure_onset_frame": 150,
+    "vlm_failure_onset_step": 150,
+    "vlm_coarse_failure_onset_seconds": 2.5,
+    "vlm_failure_window_start_seconds": 2.0,
+    "vlm_failure_window_end_seconds": 3.0,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot aligned poorly with the eggplant on the plate and closed its gripper without securing a firm hold, failing to pick it up.",
+    "vlm_failure_reason_summary": "failed grasp on eggplant; gripper closed empty",
+    "vlm_recovery_action": "Re-align the gripper directly over the body of the eggplant, descend to a lower grasp depth, and close the fingers firmly around the eggplant before lifting.",
+    "vlm_recovery_action_summary": "reposition gripper over eggplant; close grasp securely; lift off plate",
+    "vlm_prevention_action": "Ensure accurate depth and centering relative to the eggplant before closing the gripper so the fingers clamp securely around the object.",
+    "vlm_prevention_action_summary": "align gripper depth and position; close fingers around eggplant body",
+    "vlm_justification": "The robot approaches the plate around 2.0s and attempts to grasp the eggplant at 2.5s, but closes its gripper without securing the object and lifts away empty by 3.0s.",
+    "vlm_temporal_justification": "At timestamp 02:30 (second 150 of the clip), the robot's gripper fully closes without surrounding or securing the body of the eggplant, resulting in a failed grasp as the arm begins to ascend empty.",
+    "vlm_video_fps": 50.5,
+    "vlm_video_num_frames": 392,
+    "vlm_video_duration_seconds": 7.742574257425742,
+    "vlm_usage": {
+      "input_tokens": 19929,
+      "output_tokens": 375,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 2.5,\n \"failure_window_start_seconds\": 2.0,\n \"failure_window_end_seconds\": 3.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot aligned poorly with the eggplant on the plate and closed its gripper without securing a firm hold, failing to pick it up.\",\n \"failure_reason_summary\": \"failed grasp on eggplant; gripper closed empty\",\n \"recovery_action\": \"Re-align the gripper directly over the body of the eggplant, descend to a lower grasp depth, and close the fingers firmly around the eggplant before lifting.\",\n \"recovery_action_summary\": \"reposition gripper over eggplant; close grasp securely; lift off plate\",\n \"prevention_action\": \"Ensure accurate depth and centering relative to the eggplant before closing the gripper so the fingers clamp securely around the object.\",\n \"prevention_action_summary\": \"align gripper depth and position; close fingers around eggplant body\",\n \"justification\": \"The robot approaches the plate around 2.0s and attempts to grasp the eggplant at 2.5s, but closes its gripper without securing the object and lifts away empty by 3.0s.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 150,
+    "failure_sim_step": 150,
+    "failure_policy_step": 25,
+    "failure_chunk_index": 0,
+    "first_post_failure_policy_step": 26,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-align the gripper directly over the body of the eggplant, descend to a lower grasp depth, and close the fingers firmly around the eggplant before lifting."
+  }
+}
\ No newline at end of file
diff --git a/eggplant/ep040/labels.npz b/eggplant/ep040/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..a2fda9f49bf95562c1e391f7257794530a1637fb
--- /dev/null
+++ b/eggplant/ep040/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a8cec1721cc0c9eddfbdb5594e890303893f27e5385d169dfd94921ec976e2c1
+size 3894
diff --git a/eggplant/ep041/example.md b/eggplant/ep041/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..1a22135bfbd5f3158f3b39e8ec3a9314cfdf028a
--- /dev/null
+++ b/eggplant/ep041/example.md
@@ -0,0 +1,34 @@
+# t004__ep041__eggplant
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the eggplant off the plate
+- **success:** False
+- **control steps:** 576  (11.5s at 50 Hz)
+- **policy inferences:** 101  (replan_steps=6)
+- **video:** `eggplant/ep041/rollout.mp4`
+- **features:** `101` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | approaching plate with eggplant |
+| 3.0–6.0 | 151–301 | 25–50 | positioning gripper over eggplant |
+| 6.0–9.0 | 302–452 | 50–75 | attempting grasp and lifting empty |
+| 9.0–11.5 | 453–575 | 75–95 | retracting arm above table |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 8.568588469184892 s (00:08.6)
+- **frame / control_step:** 431 (policy_step=71, chunk_index=5)
+- **coarse onset:** 7.0 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the eggplant on the plate, but the end effector closed without securing a proper grip, leaving the eggplant on the plate when the arm retracted.
+  Summary: failed grasp attempt; eggplant left on plate
+- **recovery:** Re-align the end effector over the eggplant, lower the gripper around the body of the eggplant, close the gripper securely, and lift it off the plate.
+  Summary: re-align gripper; grasp eggplant securely
+- **prevention:** Lower the gripper lower and more centered over the eggplant before closing the fingers to guarantee a firm hold.
+  Summary: center gripper lower; close fingers firmly
+- **justification:** None
diff --git a/eggplant/ep041/labels.json b/eggplant/ep041/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..efb516b05160c4a50973925e688450e25a766b01
--- /dev/null
+++ b/eggplant/ep041/labels.json
@@ -0,0 +1,109 @@
+{
+  "rollout_id": "t004__ep041__eggplant",
+  "scene_variant": "",
+  "task_id": 4,
+  "task_file": "",
+  "instruction": "Take the eggplant off the plate",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:49:02+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.001988071570577,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "approaching plate with eggplant",
+      "description": "approaching plate with eggplant"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.001988071570577,
+      "t_end_sec": 6.003976143141154,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "positioning gripper over eggplant",
+      "description": "positioning gripper over eggplant"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.003976143141154,
+      "t_end_sec": 9.00596421471173,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "attempting grasp and lifting empty",
+      "description": "attempting grasp and lifting empty"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.00596421471173,
+      "t_end_sec": 11.451292246520875,
+      "control_step_start": 453,
+      "control_step_end": 575,
+      "policy_step_start": 75,
+      "policy_step_end": 95,
+      "phrase": "retracting arm above table",
+      "description": "retracting arm above table"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 8.568588469184892,
+    "vlm_failure_onset_timestamp": "00:08.6",
+    "vlm_failure_onset_frame": 431,
+    "vlm_failure_onset_step": 431,
+    "vlm_coarse_failure_onset_seconds": 7.0,
+    "vlm_failure_window_start_seconds": 6.5,
+    "vlm_failure_window_end_seconds": 8.0,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the eggplant on the plate, but the end effector closed without securing a proper grip, leaving the eggplant on the plate when the arm retracted.",
+    "vlm_failure_reason_summary": "failed grasp attempt; eggplant left on plate",
+    "vlm_recovery_action": "Re-align the end effector over the eggplant, lower the gripper around the body of the eggplant, close the gripper securely, and lift it off the plate.",
+    "vlm_recovery_action_summary": "re-align gripper; grasp eggplant securely",
+    "vlm_prevention_action": "Lower the gripper lower and more centered over the eggplant before closing the fingers to guarantee a firm hold.",
+    "vlm_prevention_action_summary": "center gripper lower; close fingers firmly",
+    "vlm_justification": null,
+    "vlm_temporal_justification": "At 03:50 in the clip (clip second 230, corresponding to original frame 431), the robot's gripper fully closes its fingers together without clamping onto the eggplant, making the grasp failure visibly apparent as the end effector closes empty.",
+    "vlm_video_fps": 50.3,
+    "vlm_video_num_frames": 576,
+    "vlm_video_duration_seconds": 11.43141153081511,
+    "vlm_usage": {
+      "input_tokens": 21645,
+      "output_tokens": 331,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 7.0,\n \"failure_window_start_seconds\": 6.5,\n \"failure_window_end_seconds\": 8.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the eggplant on the plate, but the end effector closed without securing a proper grip, leaving the eggplant on the plate when the arm retracted.\",\n \"failure_reason_summary\": \"failed grasp attempt; eggplant left on plate\",\n \"recovery_action\": \"Re-align the end effector over the eggplant, lower the gripper around the body of the eggplant, close the gripper securely, and lift it off the plate.\",\n \"recovery_action_summary\": \"re-align gripper; grasp eggplant securely\",\n \"prevention_action\": \"Lower the gripper lower and more centered over the eggplant before closing the fingers to guarantee a firm hold.\",\n \"prevention_action_summary\": \"center gripper lower; close fingers firmly\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 431,
+    "failure_sim_step": 431,
+    "failure_policy_step": 71,
+    "failure_chunk_index": 5,
+    "first_post_failure_policy_step": 72,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-align the end effector over the eggplant, lower the gripper around the body of the eggplant, close the gripper securely, and lift it off the plate."
+  }
+}
\ No newline at end of file
diff --git a/eggplant/ep041/labels.npz b/eggplant/ep041/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..30aa8d509f0de5a7a875607341f7eb2745607f4a
--- /dev/null
+++ b/eggplant/ep041/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:83b4fb22661880df52d4f5f6e5d453463315228cca2e5e48e89912b3bd06849b
+size 3820
diff --git a/eggplant/ep042/example.md b/eggplant/ep042/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..8fb7888d51ef1e6ee8907378a83d9612ff625328
--- /dev/null
+++ b/eggplant/ep042/example.md
@@ -0,0 +1,36 @@
+# t004__ep042__eggplant
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the eggplant off the plate
+- **success:** False
+- **control steps:** 861  (17.1s at 50 Hz)
+- **policy inferences:** 149  (replan_steps=6)
+- **video:** `eggplant/ep042/rollout.mp4`
+- **features:** `149` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | approaching eggplant on plate |
+| 3.0–6.0 | 151–301 | 25–50 | closing gripper; tipping eggplant upright |
+| 6.0–9.0 | 302–452 | 50–75 | squeezing upright eggplant against plate |
+| 9.0–12.0 | 453–603 | 75–100 | holding upright eggplant on plate |
+| 12.0–15.0 | 604–754 | 100–125 | holding eggplant vertically on plate |
+| 15.0–17.1 | 755–860 | 125–143 | idle holding eggplant on plate |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 3.253968253968254 s (00:03.3)
+- **frame / control_step:** 164 (policy_step=27, chunk_index=2)
+- **coarse onset:** 5.8 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot misaligned its gripper during the initial grasp attempt, causing the eggplant to flip into a vertical orientation on the plate and preventing the robot from lifting it.
+  Summary: failed grasp alignment; flipping object vertically
+- **recovery:** Release the eggplant, re-align the gripper parallel to the object's horizontal body, and re-grasp it firmly before lifting it off the plate.
+  Summary: release eggplant; regrasp horizontally; lift off plate
+- **prevention:** Position the open gripper directly over the center of the eggplant and close the fingers horizontally to secure a stable grasp before lifting.
+  Summary: align gripper horizontally; grasp eggplant securely
+- **justification:** At around 5.8 seconds, the robot closes its gripper prematurely without proper alignment, knocking the eggplant into an upright position on the plate. The robot then remains stuck squeezing the eggplant against the plate without ever lifting it.
diff --git a/eggplant/ep042/labels.json b/eggplant/ep042/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..8cc8bb95e33b6984f0d74fe7b8f358895ed264d2
--- /dev/null
+++ b/eggplant/ep042/labels.json
@@ -0,0 +1,131 @@
+{
+  "rollout_id": "t004__ep042__eggplant",
+  "scene_variant": "",
+  "task_id": 4,
+  "task_file": "",
+  "instruction": "Take the eggplant off the plate",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:49:37+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.996031746031746,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "approaching eggplant on plate",
+      "description": "approaching eggplant on plate"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.996031746031746,
+      "t_end_sec": 5.992063492063492,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "closing gripper; tipping eggplant upright",
+      "description": "closing gripper; tipping eggplant upright"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.992063492063492,
+      "t_end_sec": 8.988095238095239,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "squeezing upright eggplant against plate",
+      "description": "squeezing upright eggplant against plate"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.988095238095239,
+      "t_end_sec": 11.984126984126984,
+      "control_step_start": 453,
+      "control_step_end": 603,
+      "policy_step_start": 75,
+      "policy_step_end": 100,
+      "phrase": "holding upright eggplant on plate",
+      "description": "holding upright eggplant on plate"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 11.984126984126984,
+      "t_end_sec": 14.980158730158731,
+      "control_step_start": 604,
+      "control_step_end": 754,
+      "policy_step_start": 100,
+      "policy_step_end": 125,
+      "phrase": "holding eggplant vertically on plate",
+      "description": "holding eggplant vertically on plate"
+    },
+    {
+      "segment_index": 5,
+      "t_start_sec": 14.980158730158731,
+      "t_end_sec": 17.083333333333332,
+      "control_step_start": 755,
+      "control_step_end": 860,
+      "policy_step_start": 125,
+      "policy_step_end": 143,
+      "phrase": "idle holding eggplant on plate",
+      "description": "idle holding eggplant on plate"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 3.253968253968254,
+    "vlm_failure_onset_timestamp": "00:03.3",
+    "vlm_failure_onset_frame": 164,
+    "vlm_failure_onset_step": 164,
+    "vlm_coarse_failure_onset_seconds": 5.8,
+    "vlm_failure_window_start_seconds": 5.0,
+    "vlm_failure_window_end_seconds": 6.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot misaligned its gripper during the initial grasp attempt, causing the eggplant to flip into a vertical orientation on the plate and preventing the robot from lifting it.",
+    "vlm_failure_reason_summary": "failed grasp alignment; flipping object vertically",
+    "vlm_recovery_action": "Release the eggplant, re-align the gripper parallel to the object's horizontal body, and re-grasp it firmly before lifting it off the plate.",
+    "vlm_recovery_action_summary": "release eggplant; regrasp horizontally; lift off plate",
+    "vlm_prevention_action": "Position the open gripper directly over the center of the eggplant and close the fingers horizontally to secure a stable grasp before lifting.",
+    "vlm_prevention_action_summary": "align gripper horizontally; grasp eggplant securely",
+    "vlm_justification": "At around 5.8 seconds, the robot closes its gripper prematurely without proper alignment, knocking the eggplant into an upright position on the plate. The robot then remains stuck squeezing the eggplant against the plate without ever lifting it.",
+    "vlm_temporal_justification": "At second 23, the robot's gripper closes on the eggplant with a misaligned grasp, pressing against its edge and causing it to tilt into a vertical orientation rather than securing it flat.",
+    "vlm_video_fps": 50.4,
+    "vlm_video_num_frames": 861,
+    "vlm_video_duration_seconds": 17.063492063492063,
+    "vlm_usage": {
+      "input_tokens": 22023,
+      "output_tokens": 368,
+      "cached_tokens": 1157
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 5.8,\n \"failure_window_start_seconds\": 5.0,\n \"failure_window_end_seconds\": 6.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot misaligned its gripper during the initial grasp attempt, causing the eggplant to flip into a vertical orientation on the plate and preventing the robot from lifting it.\",\n \"failure_reason_summary\": \"failed grasp alignment; flipping object vertically\",\n \"recovery_action\": \"Release the eggplant, re-align the gripper parallel to the object's horizontal body, and re-grasp it firmly before lifting it off the plate.\",\n \"recovery_action_summary\": \"release eggplant; regrasp horizontally; lift off plate\",\n \"prevention_action\": \"Position the open gripper directly over the center of the eggplant and close the fingers horizontally to secure a stable grasp before lifting.\",\n \"prevention_action_summary\": \"align gripper horizontally; grasp eggplant securely\",\n \"justification\": \"At around 5.8 seconds, the robot closes its gripper prematurely without proper alignment, knocking the eggplant into an upright position on the plate. The robot then remains stuck squeezing the eggplant against the plate without ever lifting it.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 164,
+    "failure_sim_step": 164,
+    "failure_policy_step": 27,
+    "failure_chunk_index": 2,
+    "first_post_failure_policy_step": 28,
+    "failure_type": "grasp_failure",
+    "correction_action": "Release the eggplant, re-align the gripper parallel to the object's horizontal body, and re-grasp it firmly before lifting it off the plate."
+  }
+}
\ No newline at end of file
diff --git a/eggplant/ep042/labels.npz b/eggplant/ep042/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..79b37e3c07ccdd6e89dafd6820a60f16e8971e1d
--- /dev/null
+++ b/eggplant/ep042/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e54a87f51036624e8e223a06a941ec3288b3930d38ec79a5b99ae24089d13ad1
+size 3975
diff --git a/eggplant/ep043/example.md b/eggplant/ep043/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..6890a1b0ba829290420b176de374d6d9c8f53bf4
--- /dev/null
+++ b/eggplant/ep043/example.md
@@ -0,0 +1,33 @@
+# t004__ep043__eggplant
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the eggplant off the plate
+- **success:** False
+- **control steps:** 344  (6.9s at 50 Hz)
+- **policy inferences:** 60  (replan_steps=6)
+- **video:** `eggplant/ep043/rollout.mp4`
+- **features:** `60` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | approaching eggplant on plate |
+| 3.0–6.0 | 151–301 | 25–50 | attempting grasp and lifting empty |
+| 6.0–6.9 | 302–343 | 50–57 | hovering above plate |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 3.3466135458167328 s (00:03.3)
+- **frame / control_step:** 168 (policy_step=28, chunk_index=0)
+- **coarse onset:** 3.8 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot descended to grasp the eggplant on the plate, but the gripper closed improperly without securing the object, lifting back up empty.
+  Summary: failed grasp attempt; lifting empty gripper
+- **recovery:** Lower the gripper back down to align around the eggplant, close the fingers firmly around its center, and lift it off the plate.
+  Summary: re-align gripper; grasp eggplant; lift off plate
+- **prevention:** Align the end-effector accurately over the eggplant before closing the gripper to establish a secure hold.
+  Summary: align gripper precisely; execute firm grasp on eggplant
+- **justification:** Around 3.8 seconds, the robot closes its gripper on the eggplant but fails to establish a secure grip, pulling away empty at 5.0 seconds.
diff --git a/eggplant/ep043/labels.json b/eggplant/ep043/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..c790e568c81fcd37c7d9490feb7b7970c4391e80
--- /dev/null
+++ b/eggplant/ep043/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t004__ep043__eggplant",
+  "scene_variant": "",
+  "task_id": 4,
+  "task_file": "",
+  "instruction": "Take the eggplant off the plate",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:50:06+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00796812749004,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "approaching eggplant on plate",
+      "description": "approaching eggplant on plate"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00796812749004,
+      "t_end_sec": 6.01593625498008,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "attempting grasp and lifting empty",
+      "description": "attempting grasp and lifting empty"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01593625498008,
+      "t_end_sec": 6.852589641434262,
+      "control_step_start": 302,
+      "control_step_end": 343,
+      "policy_step_start": 50,
+      "policy_step_end": 57,
+      "phrase": "hovering above plate",
+      "description": "hovering above plate"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 3.3466135458167328,
+    "vlm_failure_onset_timestamp": "00:03.3",
+    "vlm_failure_onset_frame": 168,
+    "vlm_failure_onset_step": 168,
+    "vlm_coarse_failure_onset_seconds": 3.8,
+    "vlm_failure_window_start_seconds": 3.3,
+    "vlm_failure_window_end_seconds": 4.8,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot descended to grasp the eggplant on the plate, but the gripper closed improperly without securing the object, lifting back up empty.",
+    "vlm_failure_reason_summary": "failed grasp attempt; lifting empty gripper",
+    "vlm_recovery_action": "Lower the gripper back down to align around the eggplant, close the fingers firmly around its center, and lift it off the plate.",
+    "vlm_recovery_action_summary": "re-align gripper; grasp eggplant; lift off plate",
+    "vlm_prevention_action": "Align the end-effector accurately over the eggplant before closing the gripper to establish a secure hold.",
+    "vlm_prevention_action_summary": "align gripper precisely; execute firm grasp on eggplant",
+    "vlm_justification": "Around 3.8 seconds, the robot closes its gripper on the eggplant but fails to establish a secure grip, pulling away empty at 5.0 seconds.",
+    "vlm_temporal_justification": "The robot lowers its end-effector to grasp the eggplant, but at timestamp 02:08 (128 seconds into the clip), the gripper fingers close completely against each other without pinching the object, failing to secure it before lifting.",
+    "vlm_video_fps": 50.2,
+    "vlm_video_num_frames": 344,
+    "vlm_video_duration_seconds": 6.832669322709163,
+    "vlm_usage": {
+      "input_tokens": 21381,
+      "output_tokens": 354,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 3.8,\n \"failure_window_start_seconds\": 3.3,\n \"failure_window_end_seconds\": 4.8,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot descended to grasp the eggplant on the plate, but the gripper closed improperly without securing the object, lifting back up empty.\",\n \"failure_reason_summary\": \"failed grasp attempt; lifting empty gripper\",\n \"recovery_action\": \"Lower the gripper back down to align around the eggplant, close the fingers firmly around its center, and lift it off the plate.\",\n \"recovery_action_summary\": \"re-align gripper; grasp eggplant; lift off plate\",\n \"prevention_action\": \"Align the end-effector accurately over the eggplant before closing the gripper to establish a secure hold.\",\n \"prevention_action_summary\": \"align gripper precisely; execute firm grasp on eggplant\",\n \"justification\": \"Around 3.8 seconds, the robot closes its gripper on the eggplant but fails to establish a secure grip, pulling away empty at 5.0 seconds.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 168,
+    "failure_sim_step": 168,
+    "failure_policy_step": 28,
+    "failure_chunk_index": 0,
+    "first_post_failure_policy_step": 29,
+    "failure_type": "grasp_failure",
+    "correction_action": "Lower the gripper back down to align around the eggplant, close the fingers firmly around its center, and lift it off the plate."
+  }
+}
\ No newline at end of file
diff --git a/eggplant/ep043/labels.npz b/eggplant/ep043/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..719b38c7adb61de956824962dc8eaafb526386a6
--- /dev/null
+++ b/eggplant/ep043/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2bfcf3d8613535b29a88c27032021ff318d9784caa4a2d663e7f8bdd637fb31c
+size 3775
diff --git a/eggplant/ep044/example.md b/eggplant/ep044/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..fb392298963fb86f3eeee5a4494948cdc5b0c9fe
--- /dev/null
+++ b/eggplant/ep044/example.md
@@ -0,0 +1,21 @@
+# t004__ep044__eggplant
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the eggplant off the plate
+- **success:** True
+- **control steps:** 436  (8.6s at 50 Hz)
+- **policy inferences:** 75  (replan_steps=6)
+- **video:** `eggplant/ep044/rollout.mp4`
+- **features:** `75` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching for the eggplant |
+| 3.0–6.0 | 152–303 | 25–50 | grasping and lifting the eggplant |
+| 6.0–8.6 | 304–435 | 50–72 | placing eggplant on the table |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/eggplant/ep044/labels.json b/eggplant/ep044/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..b7e05172a0681456397803df967577d57cfe767a
--- /dev/null
+++ b/eggplant/ep044/labels.json
@@ -0,0 +1,68 @@
+{
+  "rollout_id": "t004__ep044__eggplant",
+  "scene_variant": "",
+  "task_id": 4,
+  "task_file": "",
+  "instruction": "Take the eggplant off the plate",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:50:13+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00990099009901,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for the eggplant",
+      "description": "reaching for the eggplant"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00990099009901,
+      "t_end_sec": 6.01980198019802,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "grasping and lifting the eggplant",
+      "description": "grasping and lifting the eggplant"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01980198019802,
+      "t_end_sec": 8.633663366336634,
+      "control_step_start": 304,
+      "control_step_end": 435,
+      "policy_step_start": 50,
+      "policy_step_end": 72,
+      "phrase": "placing eggplant on the table",
+      "description": "placing eggplant on the table"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/eggplant/ep044/labels.npz b/eggplant/ep044/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..05acdb62cba633a58c5872bd44d55ff6f5f29dbe
--- /dev/null
+++ b/eggplant/ep044/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e484a0677e0da3e67a736c3a4e1a222a5fdf1fb13f7802e26a605593bed8a4fe
+size 3084
diff --git a/eggplant/ep045/example.md b/eggplant/ep045/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..2b66b9e44fca96ee8129d3917b744eed75acc190
--- /dev/null
+++ b/eggplant/ep045/example.md
@@ -0,0 +1,34 @@
+# t004__ep045__eggplant
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the eggplant off the plate
+- **success:** False
+- **control steps:** 458  (9.1s at 51 Hz)
+- **policy inferences:** 78  (replan_steps=6)
+- **video:** `eggplant/ep045/rollout.mp4`
+- **features:** `78` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching toward eggplant on plate |
+| 3.0–6.0 | 152–303 | 25–50 | attempting grasp and slipping off |
+| 6.0–9.0 | 304–455 | 50–75 | descending for second grasp attempt |
+| 9.0–9.1 | 456–457 | 76–76 | hovering over eggplant on plate |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 4.8023715415019765 s (00:04.8)
+- **frame / control_step:** 243 (policy_step=40, chunk_index=3)
+- **coarse onset:** 4.8 s (refined=False)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the eggplant on the plate, but the gripper closed improperly and failed to establish a secure hold, leaving the eggplant on the plate when lifting.
+  Summary: failed grasp on eggplant; gripper slips off object
+- **recovery:** Re-align the gripper directly over the central mass of the eggplant, close the fingers firmly around it, and lift it clear of the plate.
+  Summary: re-align gripper; securely grasp and lift eggplant
+- **prevention:** Position the end-effector accurately over the center of the eggplant before closing the fingers to guarantee a solid hold before lifting.
+  Summary: center gripper on eggplant; ensure firm hold before lifting
+- **justification:** At around 4.8 seconds, the robot closes its gripper onto the eggplant but fails to secure it. The arm moves away empty, leaving the task incomplete.
diff --git a/eggplant/ep045/labels.json b/eggplant/ep045/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..43d7ae1f2909ea81d6e0b27d080a37bafbc66d88
--- /dev/null
+++ b/eggplant/ep045/labels.json
@@ -0,0 +1,109 @@
+{
+  "rollout_id": "t004__ep045__eggplant",
+  "scene_variant": "",
+  "task_id": 4,
+  "task_file": "",
+  "instruction": "Take the eggplant off the plate",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:50:49+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0039525691699605,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching toward eggplant on plate",
+      "description": "reaching toward eggplant on plate"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0039525691699605,
+      "t_end_sec": 6.007905138339921,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "attempting grasp and slipping off",
+      "description": "attempting grasp and slipping off"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.007905138339921,
+      "t_end_sec": 9.011857707509881,
+      "control_step_start": 304,
+      "control_step_end": 455,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "descending for second grasp attempt",
+      "description": "descending for second grasp attempt"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.011857707509881,
+      "t_end_sec": 9.051383399209486,
+      "control_step_start": 456,
+      "control_step_end": 457,
+      "policy_step_start": 76,
+      "policy_step_end": 76,
+      "phrase": "hovering over eggplant on plate",
+      "description": "hovering over eggplant on plate"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 4.8023715415019765,
+    "vlm_failure_onset_timestamp": "00:04.8",
+    "vlm_failure_onset_frame": 243,
+    "vlm_failure_onset_step": 243,
+    "vlm_coarse_failure_onset_seconds": 4.8,
+    "vlm_failure_window_start_seconds": 3.8,
+    "vlm_failure_window_end_seconds": 5.8,
+    "vlm_temporal_refined": false,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the eggplant on the plate, but the gripper closed improperly and failed to establish a secure hold, leaving the eggplant on the plate when lifting.",
+    "vlm_failure_reason_summary": "failed grasp on eggplant; gripper slips off object",
+    "vlm_recovery_action": "Re-align the gripper directly over the central mass of the eggplant, close the fingers firmly around it, and lift it clear of the plate.",
+    "vlm_recovery_action_summary": "re-align gripper; securely grasp and lift eggplant",
+    "vlm_prevention_action": "Position the end-effector accurately over the center of the eggplant before closing the fingers to guarantee a solid hold before lifting.",
+    "vlm_prevention_action_summary": "center gripper on eggplant; ensure firm hold before lifting",
+    "vlm_justification": "At around 4.8 seconds, the robot closes its gripper onto the eggplant but fails to secure it. The arm moves away empty, leaving the task incomplete.",
+    "vlm_temporal_justification": "The robot positions its gripper over the eggplant and attempts to close its fingers around 03:18. By 03:22, the fingers close improperly and slip off the object, leaving the eggplant on the plate as the arm prepares to lift up empty.",
+    "vlm_video_fps": 50.6,
+    "vlm_video_num_frames": 458,
+    "vlm_video_duration_seconds": 9.031620553359684,
+    "vlm_usage": {
+      "input_tokens": 21640,
+      "output_tokens": 378,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 4.8,\n \"failure_window_start_seconds\": 3.8,\n \"failure_window_end_seconds\": 5.8,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the eggplant on the plate, but the gripper closed improperly and failed to establish a secure hold, leaving the eggplant on the plate when lifting.\",\n \"failure_reason_summary\": \"failed grasp on eggplant; gripper slips off object\",\n \"recovery_action\": \"Re-align the gripper directly over the central mass of the eggplant, close the fingers firmly around it, and lift it clear of the plate.\",\n \"recovery_action_summary\": \"re-align gripper; securely grasp and lift eggplant\",\n \"prevention_action\": \"Position the end-effector accurately over the center of the eggplant before closing the fingers to guarantee a solid hold before lifting.\",\n \"prevention_action_summary\": \"center gripper on eggplant; ensure firm hold before lifting\",\n \"justification\": \"At around 4.8 seconds, the robot closes its gripper onto the eggplant but fails to secure it. The arm moves away empty, leaving the task incomplete.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 243,
+    "failure_sim_step": 243,
+    "failure_policy_step": 40,
+    "failure_chunk_index": 3,
+    "first_post_failure_policy_step": 41,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-align the gripper directly over the central mass of the eggplant, close the fingers firmly around it, and lift it clear of the plate."
+  }
+}
\ No newline at end of file
diff --git a/eggplant/ep045/labels.npz b/eggplant/ep045/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..461a7521d61f2ee2e94d63dab1c4c9a11ed37777
--- /dev/null
+++ b/eggplant/ep045/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:37f17cd05c1bd15494a31d01c844c5aa3547e60e71f99dabb764c72445c13a00
+size 3888
diff --git a/eggplant/ep046/example.md b/eggplant/ep046/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..1e7d083bfe2122f1d0c18f86ad3f74da1fc13fe2
--- /dev/null
+++ b/eggplant/ep046/example.md
@@ -0,0 +1,21 @@
+# t004__ep046__eggplant
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the eggplant off the plate
+- **success:** True
+- **control steps:** 438  (8.7s at 50 Hz)
+- **policy inferences:** 76  (replan_steps=6)
+- **video:** `eggplant/ep046/rollout.mp4`
+- **features:** `76` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | reaching for the eggplant |
+| 3.0–6.0 | 151–301 | 25–50 | picking eggplant off plate |
+| 6.0–8.7 | 302–437 | 50–72 | placing eggplant and retracting |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/eggplant/ep046/labels.json b/eggplant/ep046/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..856ff814dee0026edf65544795bf0b395ded7443
--- /dev/null
+++ b/eggplant/ep046/labels.json
@@ -0,0 +1,68 @@
+{
+  "rollout_id": "t004__ep046__eggplant",
+  "scene_variant": "",
+  "task_id": 4,
+  "task_file": "",
+  "instruction": "Take the eggplant off the plate",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:50:58+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.996031746031746,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for the eggplant",
+      "description": "reaching for the eggplant"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.996031746031746,
+      "t_end_sec": 5.992063492063492,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "picking eggplant off plate",
+      "description": "picking eggplant off plate"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.992063492063492,
+      "t_end_sec": 8.690476190476192,
+      "control_step_start": 302,
+      "control_step_end": 437,
+      "policy_step_start": 50,
+      "policy_step_end": 72,
+      "phrase": "placing eggplant and retracting",
+      "description": "placing eggplant and retracting"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/eggplant/ep046/labels.npz b/eggplant/ep046/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..6a7a42b7cbb23587530a556816a6dc5e730c38d1
--- /dev/null
+++ b/eggplant/ep046/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e8d5595763242ff4b75ba93cf0b51015f4a2ed921fde2c7c6c335fe121791198
+size 3076
diff --git a/eggplant/ep047/example.md b/eggplant/ep047/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..fe240e96df437db00d2bd133af8d74b4e56ae250
--- /dev/null
+++ b/eggplant/ep047/example.md
@@ -0,0 +1,33 @@
+# t004__ep047__eggplant
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the eggplant off the plate
+- **success:** False
+- **control steps:** 360  (7.1s at 50 Hz)
+- **policy inferences:** 62  (replan_steps=6)
+- **video:** `eggplant/ep047/rollout.mp4`
+- **features:** `62` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching for eggplant on plate |
+| 3.0–6.0 | 152–303 | 25–50 | lifting empty gripper |
+| 6.0–7.1 | 304–359 | 50–59 | hovering above table |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 2.9306930693069306 s (00:02.9)
+- **frame / control_step:** 148 (policy_step=24, chunk_index=4)
+- **coarse onset:** 3.5 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot lowered its end-effector over the eggplant and closed its gripper, but failed to securely grasp the object, leaving the eggplant on the plate when lifting up.
+  Summary: failed grasp attempt; gripper missed eggplant
+- **recovery:** Re-align the gripper directly over the eggplant, lower down to enclose it, and close the gripper securely before lifting it off the plate.
+  Summary: realign gripper over eggplant; grasp securely and lift
+- **prevention:** Center the gripper precisely over the eggplant and lower to a sufficient depth before closing the jaws to establish a firm hold.
+  Summary: center gripper over eggplant; ensure depth before closing
+- **justification:** At around 3.5 seconds, the robot closes its gripper around the eggplant on the plate but fails to capture it. When the arm retracts upwards, the eggplant is left behind on the plate.
diff --git a/eggplant/ep047/labels.json b/eggplant/ep047/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..6d48454ddca4219ac609c7d64a36b2320efced78
--- /dev/null
+++ b/eggplant/ep047/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t004__ep047__eggplant",
+  "scene_variant": "",
+  "task_id": 4,
+  "task_file": "",
+  "instruction": "Take the eggplant off the plate",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:51:28+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00990099009901,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for eggplant on plate",
+      "description": "reaching for eggplant on plate"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00990099009901,
+      "t_end_sec": 6.01980198019802,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "lifting empty gripper",
+      "description": "lifting empty gripper"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01980198019802,
+      "t_end_sec": 7.128712871287129,
+      "control_step_start": 304,
+      "control_step_end": 359,
+      "policy_step_start": 50,
+      "policy_step_end": 59,
+      "phrase": "hovering above table",
+      "description": "hovering above table"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 2.9306930693069306,
+    "vlm_failure_onset_timestamp": "00:02.9",
+    "vlm_failure_onset_frame": 148,
+    "vlm_failure_onset_step": 148,
+    "vlm_coarse_failure_onset_seconds": 3.5,
+    "vlm_failure_window_start_seconds": 3.0,
+    "vlm_failure_window_end_seconds": 4.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot lowered its end-effector over the eggplant and closed its gripper, but failed to securely grasp the object, leaving the eggplant on the plate when lifting up.",
+    "vlm_failure_reason_summary": "failed grasp attempt; gripper missed eggplant",
+    "vlm_recovery_action": "Re-align the gripper directly over the eggplant, lower down to enclose it, and close the gripper securely before lifting it off the plate.",
+    "vlm_recovery_action_summary": "realign gripper over eggplant; grasp securely and lift",
+    "vlm_prevention_action": "Center the gripper precisely over the eggplant and lower to a sufficient depth before closing the jaws to establish a firm hold.",
+    "vlm_prevention_action_summary": "center gripper over eggplant; ensure depth before closing",
+    "vlm_justification": "At around 3.5 seconds, the robot closes its gripper around the eggplant on the plate but fails to capture it. When the arm retracts upwards, the eggplant is left behind on the plate.",
+    "vlm_temporal_justification": "At 02:03 (clip second 123), the robot finishes closing its gripper fingers around the eggplant, but fails to pinch or secure it, closing fully without grasping the object.",
+    "vlm_video_fps": 50.5,
+    "vlm_video_num_frames": 360,
+    "vlm_video_duration_seconds": 7.108910891089109,
+    "vlm_usage": {
+      "input_tokens": 21514,
+      "output_tokens": 365,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 3.5,\n \"failure_window_start_seconds\": 3.0,\n \"failure_window_end_seconds\": 4.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot lowered its end-effector over the eggplant and closed its gripper, but failed to securely grasp the object, leaving the eggplant on the plate when lifting up.\",\n \"failure_reason_summary\": \"failed grasp attempt; gripper missed eggplant\",\n \"recovery_action\": \"Re-align the gripper directly over the eggplant, lower down to enclose it, and close the gripper securely before lifting it off the plate.\",\n \"recovery_action_summary\": \"realign gripper over eggplant; grasp securely and lift\",\n \"prevention_action\": \"Center the gripper precisely over the eggplant and lower to a sufficient depth before closing the jaws to establish a firm hold.\",\n \"prevention_action_summary\": \"center gripper over eggplant; ensure depth before closing\",\n \"justification\": \"At around 3.5 seconds, the robot closes its gripper around the eggplant on the plate but fails to capture it. When the arm retracts upwards, the eggplant is left behind on the plate.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 148,
+    "failure_sim_step": 148,
+    "failure_policy_step": 24,
+    "failure_chunk_index": 4,
+    "first_post_failure_policy_step": 25,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-align the gripper directly over the eggplant, lower down to enclose it, and close the gripper securely before lifting it off the plate."
+  }
+}
\ No newline at end of file
diff --git a/eggplant/ep047/labels.npz b/eggplant/ep047/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..c3e2495a9a352d855327662533fb0a431ba2f340
--- /dev/null
+++ b/eggplant/ep047/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3a0cc2e28885494d436c05748855848796a07da17975c2eae7f3bb5985a5a3dd
+size 3818
diff --git a/eggplant/ep048/example.md b/eggplant/ep048/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..701471b8c656efb5b255c211f5e1012e63cd8470
--- /dev/null
+++ b/eggplant/ep048/example.md
@@ -0,0 +1,34 @@
+# t004__ep048__eggplant
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the eggplant off the plate
+- **success:** False
+- **control steps:** 575  (11.3s at 51 Hz)
+- **policy inferences:** 100  (replan_steps=5)
+- **video:** `eggplant/ep048/rollout.mp4`
+- **features:** `100` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–30 | approaching eggplant on plate |
+| 3.0–6.0 | 152–303 | 30–60 | closing gripper on eggplant |
+| 6.0–9.0 | 304–455 | 60–91 | holding position near eggplant |
+| 9.0–11.3 | 456–574 | 91–114 | idle over plate |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 3.7672583826429977 s (00:03.8)
+- **frame / control_step:** 191 (policy_step=38, chunk_index=1)
+- **coarse onset:** 3.8 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot lowered its end-effector over the eggplant and closed its gripper, but failed to establish a secure hold, leaving the eggplant resting on the plate.
+  Summary: failed grasp on eggplant; object remains on plate
+- **recovery:** Open the gripper, reposition the end-effector directly over the center of the eggplant, and close the gripper securely before lifting.
+  Summary: re-align gripper; regrasp eggplant; lift off plate
+- **prevention:** Align the gripper precisely with the body of the eggplant and ensure proper finger contact before initiating the lift motion.
+  Summary: center gripper over eggplant; grasp securely
+- **justification:** At approximately 3.8 seconds, the gripper closes on the eggplant but fails to secure it. The robot then remains in place without lifting the eggplant, resulting in a grasp failure.
diff --git a/eggplant/ep048/labels.json b/eggplant/ep048/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..b4c239993583decc4bbe5fb31e8871751d2439eb
--- /dev/null
+++ b/eggplant/ep048/labels.json
@@ -0,0 +1,109 @@
+{
+  "rollout_id": "t004__ep048__eggplant",
+  "scene_variant": "",
+  "task_id": 4,
+  "task_file": "",
+  "instruction": "Take the eggplant off the plate",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:52:01+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9980276134122286,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 30,
+      "phrase": "approaching eggplant on plate",
+      "description": "approaching eggplant on plate"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9980276134122286,
+      "t_end_sec": 5.996055226824457,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 30,
+      "policy_step_end": 60,
+      "phrase": "closing gripper on eggplant",
+      "description": "closing gripper on eggplant"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.996055226824457,
+      "t_end_sec": 8.994082840236686,
+      "control_step_start": 304,
+      "control_step_end": 455,
+      "policy_step_start": 60,
+      "policy_step_end": 91,
+      "phrase": "holding position near eggplant",
+      "description": "holding position near eggplant"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.994082840236686,
+      "t_end_sec": 11.341222879684418,
+      "control_step_start": 456,
+      "control_step_end": 574,
+      "policy_step_start": 91,
+      "policy_step_end": 114,
+      "phrase": "idle over plate",
+      "description": "idle over plate"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 3.7672583826429977,
+    "vlm_failure_onset_timestamp": "00:03.8",
+    "vlm_failure_onset_frame": 191,
+    "vlm_failure_onset_step": 191,
+    "vlm_coarse_failure_onset_seconds": 3.8,
+    "vlm_failure_window_start_seconds": 3.5,
+    "vlm_failure_window_end_seconds": 4.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot lowered its end-effector over the eggplant and closed its gripper, but failed to establish a secure hold, leaving the eggplant resting on the plate.",
+    "vlm_failure_reason_summary": "failed grasp on eggplant; object remains on plate",
+    "vlm_recovery_action": "Open the gripper, reposition the end-effector directly over the center of the eggplant, and close the gripper securely before lifting.",
+    "vlm_recovery_action_summary": "re-align gripper; regrasp eggplant; lift off plate",
+    "vlm_prevention_action": "Align the gripper precisely with the body of the eggplant and ensure proper finger contact before initiating the lift motion.",
+    "vlm_prevention_action_summary": "center gripper over eggplant; grasp securely",
+    "vlm_justification": "At approximately 3.8 seconds, the gripper closes on the eggplant but fails to secure it. The robot then remains in place without lifting the eggplant, resulting in a grasp failure.",
+    "vlm_temporal_justification": "At 02:30 in the magnified clip, the robot's gripper fully closes over the eggplant without securing a hold, leaving the eggplant behind on the plate as the fingers clamp shut above/around it.",
+    "vlm_video_fps": 50.7,
+    "vlm_video_num_frames": 575,
+    "vlm_video_duration_seconds": 11.321499013806706,
+    "vlm_usage": {
+      "input_tokens": 21766,
+      "output_tokens": 359,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 3.8,\n \"failure_window_start_seconds\": 3.5,\n \"failure_window_end_seconds\": 4.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot lowered its end-effector over the eggplant and closed its gripper, but failed to establish a secure hold, leaving the eggplant resting on the plate.\",\n \"failure_reason_summary\": \"failed grasp on eggplant; object remains on plate\",\n \"recovery_action\": \"Open the gripper, reposition the end-effector directly over the center of the eggplant, and close the gripper securely before lifting.\",\n \"recovery_action_summary\": \"re-align gripper; regrasp eggplant; lift off plate\",\n \"prevention_action\": \"Align the gripper precisely with the body of the eggplant and ensure proper finger contact before initiating the lift motion.\",\n \"prevention_action_summary\": \"center gripper over eggplant; grasp securely\",\n \"justification\": \"At approximately 3.8 seconds, the gripper closes on the eggplant but fails to secure it. The robot then remains in place without lifting the eggplant, resulting in a grasp failure.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 191,
+    "failure_sim_step": 191,
+    "failure_policy_step": 38,
+    "failure_chunk_index": 1,
+    "first_post_failure_policy_step": 39,
+    "failure_type": "grasp_failure",
+    "correction_action": "Open the gripper, reposition the end-effector directly over the center of the eggplant, and close the gripper securely before lifting."
+  }
+}
\ No newline at end of file
diff --git a/eggplant/ep048/labels.npz b/eggplant/ep048/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..180abe7ee43e0d7825dcd0b613f14ec530bbd937
--- /dev/null
+++ b/eggplant/ep048/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b86ce0e4cd5b379d7c101db0da6e520f8859da78831feea4c87ac765e9feaafe
+size 3822
diff --git a/eggplant/ep049/example.md b/eggplant/ep049/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..4729121366887a699d627b3e301048e3d28f83b4
--- /dev/null
+++ b/eggplant/ep049/example.md
@@ -0,0 +1,34 @@
+# t004__ep049__eggplant
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the eggplant off the plate
+- **success:** False
+- **control steps:** 502  (9.8s at 51 Hz)
+- **policy inferences:** 87  (replan_steps=6)
+- **video:** `eggplant/ep049/rollout.mp4`
+- **features:** `87` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–153 | 0–25 | hovering in initial position |
+| 3.0–6.0 | 154–307 | 25–51 | descending towards eggplant on plate |
+| 6.0–9.0 | 308–461 | 51–76 | closing gripper on eggplant |
+| 9.0–9.8 | 462–501 | 77–83 | holding position above plate |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 4.6875 s (00:04.7)
+- **frame / control_step:** 240 (policy_step=40, chunk_index=0)
+- **coarse onset:** 7.0 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot descended to the plate and attempted to grab the eggplant, but the gripper closed without establishing a firm grip on the object.
+  Summary: failed grasp on eggplant; closing gripper without secure grip
+- **recovery:** Re-open the gripper, re-position the fingers centrally around the eggplant on the plate, and execute a solid grasp before lifting.
+  Summary: reopen gripper; re-align and grasp eggplant
+- **prevention:** Align the gripper directly above the center of mass of the eggplant before closing the fingers to ensure a secure grip.
+  Summary: align gripper centered over eggplant; grasp securely
+- **justification:** At around 7.0 seconds into the video, the robot's gripper reaches down to the eggplant on the plate and attempts to close, but fails to grip the eggplant properly, leading to a grasp failure.
diff --git a/eggplant/ep049/labels.json b/eggplant/ep049/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..2b6b4076c74653aaaef74b749fb93597510dd530
--- /dev/null
+++ b/eggplant/ep049/labels.json
@@ -0,0 +1,109 @@
+{
+  "rollout_id": "t004__ep049__eggplant",
+  "scene_variant": "",
+  "task_id": 4,
+  "task_file": "",
+  "instruction": "Take the eggplant off the plate",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:52:38+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0078125,
+      "control_step_start": 0,
+      "control_step_end": 153,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "hovering in initial position",
+      "description": "hovering in initial position"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0078125,
+      "t_end_sec": 6.015625,
+      "control_step_start": 154,
+      "control_step_end": 307,
+      "policy_step_start": 25,
+      "policy_step_end": 51,
+      "phrase": "descending towards eggplant on plate",
+      "description": "descending towards eggplant on plate"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.015625,
+      "t_end_sec": 9.0234375,
+      "control_step_start": 308,
+      "control_step_end": 461,
+      "policy_step_start": 51,
+      "policy_step_end": 76,
+      "phrase": "closing gripper on eggplant",
+      "description": "closing gripper on eggplant"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.0234375,
+      "t_end_sec": 9.8046875,
+      "control_step_start": 462,
+      "control_step_end": 501,
+      "policy_step_start": 77,
+      "policy_step_end": 83,
+      "phrase": "holding position above plate",
+      "description": "holding position above plate"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 4.6875,
+    "vlm_failure_onset_timestamp": "00:04.7",
+    "vlm_failure_onset_frame": 240,
+    "vlm_failure_onset_step": 240,
+    "vlm_coarse_failure_onset_seconds": 7.0,
+    "vlm_failure_window_start_seconds": 6.5,
+    "vlm_failure_window_end_seconds": 8.0,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot descended to the plate and attempted to grab the eggplant, but the gripper closed without establishing a firm grip on the object.",
+    "vlm_failure_reason_summary": "failed grasp on eggplant; closing gripper without secure grip",
+    "vlm_recovery_action": "Re-open the gripper, re-position the fingers centrally around the eggplant on the plate, and execute a solid grasp before lifting.",
+    "vlm_recovery_action_summary": "reopen gripper; re-align and grasp eggplant",
+    "vlm_prevention_action": "Align the gripper directly above the center of mass of the eggplant before closing the fingers to ensure a secure grip.",
+    "vlm_prevention_action_summary": "align gripper centered over eggplant; grasp securely",
+    "vlm_justification": "At around 7.0 seconds into the video, the robot's gripper reaches down to the eggplant on the plate and attempts to close, but fails to grip the eggplant properly, leading to a grasp failure.",
+    "vlm_temporal_justification": "At 00:36 in the clip, the robot gripper fully closes its fingers without successfully grasping the eggplant slice on the dish, resulting in a grasp failure.",
+    "vlm_video_fps": 51.2,
+    "vlm_video_num_frames": 502,
+    "vlm_video_duration_seconds": 9.78515625,
+    "vlm_usage": {
+      "input_tokens": 21258,
+      "output_tokens": 352,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 7.0,\n \"failure_window_start_seconds\": 6.5,\n \"failure_window_end_seconds\": 8.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot descended to the plate and attempted to grab the eggplant, but the gripper closed without establishing a firm grip on the object.\",\n \"failure_reason_summary\": \"failed grasp on eggplant; closing gripper without secure grip\",\n \"recovery_action\": \"Re-open the gripper, re-position the fingers centrally around the eggplant on the plate, and execute a solid grasp before lifting.\",\n \"recovery_action_summary\": \"reopen gripper; re-align and grasp eggplant\",\n \"prevention_action\": \"Align the gripper directly above the center of mass of the eggplant before closing the fingers to ensure a secure grip.\",\n \"prevention_action_summary\": \"align gripper centered over eggplant; grasp securely\",\n \"justification\": \"At around 7.0 seconds into the video, the robot's gripper reaches down to the eggplant on the plate and attempts to close, but fails to grip the eggplant properly, leading to a grasp failure.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 240,
+    "failure_sim_step": 240,
+    "failure_policy_step": 40,
+    "failure_chunk_index": 0,
+    "first_post_failure_policy_step": 41,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-open the gripper, re-position the fingers centrally around the eggplant on the plate, and execute a solid grasp before lifting."
+  }
+}
\ No newline at end of file
diff --git a/eggplant/ep049/labels.npz b/eggplant/ep049/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..daf11e731493f09ad785f06829aac719b5746571
--- /dev/null
+++ b/eggplant/ep049/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c4cf7ea2cb026f1cf8d396ad282564669166842c322595b47a1d69efd1b1f316
+size 3837
diff --git a/manifest.csv b/manifest.csv
new file mode 100644
index 0000000000000000000000000000000000000000..019d4a9e3afd9a0efa660972c500ab81af1a9aab
--- /dev/null
+++ b/manifest.csv
@@ -0,0 +1,400 @@
+scene_variant,suite,task_id,task,prompt,episode,rollout_id,success,n_policy,n_control,control_hz,replan_steps,sim_failure_category,failing_predicate,labeled_captions,labeled_failure,labeler_model,vlm_failure_onset_frame,dir
+,,,,Flip the basket upright,0,t002__ep000__basket,False,74,430,50.9,6,,,True,True,gemini-3.5-flash,160,basket/ep000
+,,,,Flip the basket upright,1,t002__ep001__basket,False,66,388,49.8,6,,,True,True,gemini-3.5-flash,284,basket/ep001
+,,,,Flip the basket upright,2,t002__ep002__basket,False,65,378,51.1,6,,,True,True,gemini-3.5-flash,231,basket/ep002
+,,,,Flip the basket upright,3,t002__ep003__basket,False,54,311,50.6,6,,,True,True,gemini-3.5-flash,162,basket/ep003
+,,,,Flip the basket upright,4,t002__ep004__basket,False,82,480,51.0,6,,,True,True,gemini-3.5-flash,201,basket/ep004
+,,,,Flip the basket upright,5,t002__ep005__basket,False,58,334,50.2,6,,,True,True,gemini-3.5-flash,128,basket/ep005
+,,,,Flip the basket upright,6,t002__ep006__basket,False,57,328,50.1,5,,,True,True,gemini-3.5-flash,95,basket/ep006
+,,,,Flip the basket upright,7,t002__ep007__basket,False,50,286,50.7,6,,,True,True,gemini-3.5-flash,135,basket/ep007
+,,,,Flip the basket upright,8,t002__ep008__basket,True,89,522,50.7,6,,,True,False,gemini-3.5-flash,,basket/ep008
+,,,,Flip the basket upright,9,t002__ep009__basket,False,47,275,50.6,6,,,True,True,gemini-3.5-flash,157,basket/ep009
+,,,,Flip the basket upright,10,t002__ep010__basket,False,55,320,50.8,6,,,True,True,gemini-3.5-flash,244,basket/ep010
+,,,,Flip the basket upright,11,t002__ep011__basket,False,56,330,50.6,6,,,True,True,gemini-3.5-flash,162,basket/ep011
+,,,,Flip the basket upright,12,t002__ep012__basket,False,43,249,51.1,6,,,True,True,gemini-3.5-flash,92,basket/ep012
+,,,,Flip the basket upright,13,t002__ep013__basket,True,57,323,50.1,6,,,True,False,gemini-3.5-flash,,basket/ep013
+,,,,Flip the basket upright,14,t002__ep014__basket,False,47,274,50.4,6,,,True,True,gemini-3.5-flash,178,basket/ep014
+,,,,Flip the basket upright,15,t002__ep015__basket,False,80,463,50.8,6,,,True,True,gemini-3.5-flash,136,basket/ep015
+,,,,Flip the basket upright,16,t002__ep016__basket,False,58,332,50.5,6,,,True,True,gemini-3.5-flash,172,basket/ep016
+,,,,Flip the basket upright,17,t002__ep017__basket,False,67,386,50.5,6,,,True,True,gemini-3.5-flash,94,basket/ep017
+,,,,Flip the basket upright,18,t002__ep018__basket,False,75,431,50.5,6,,,True,True,gemini-3.5-flash,125,basket/ep018
+,,,,Flip the basket upright,19,t002__ep019__basket,False,81,473,50.1,6,,,True,True,gemini-3.5-flash,240,basket/ep019
+,,,,Flip the basket upright,20,t002__ep020__basket,False,42,242,50.3,6,,,True,True,gemini-3.5-flash,125,basket/ep020
+,,,,Flip the basket upright,21,t002__ep021__basket,False,53,311,50.9,6,,,True,True,gemini-3.5-flash,152,basket/ep021
+,,,,Flip the basket upright,22,t002__ep022__basket,False,71,413,50.5,6,,,True,True,gemini-3.5-flash,172,basket/ep022
+,,,,Flip the basket upright,23,t002__ep023__basket,False,69,405,50.4,6,,,True,True,gemini-3.5-flash,170,basket/ep023
+,,,,Flip the basket upright,24,t002__ep024__basket,False,62,355,50.6,6,,,True,True,gemini-3.5-flash,118,basket/ep024
+,,,,Flip the basket upright,25,t002__ep025__basket,False,57,325,50.7,6,,,True,True,gemini-3.5-flash,80,basket/ep025
+,,,,Flip the basket upright,26,t002__ep026__basket,False,62,362,51.3,6,,,True,True,gemini-3.5-flash,131,basket/ep026
+,,,,Flip the basket upright,27,t002__ep027__basket,False,60,342,50.3,6,,,True,True,gemini-3.5-flash,144,basket/ep027
+,,,,Flip the basket upright,28,t002__ep028__basket,False,97,560,50.3,6,,,True,True,gemini-3.5-flash,151,basket/ep028
+,,,,Flip the basket upright,29,t002__ep029__basket,False,64,371,50.3,6,,,True,True,gemini-3.5-flash,124,basket/ep029
+,,,,Flip the basket upright,30,t002__ep030__basket,False,60,351,51.0,6,,,True,True,gemini-3.5-flash,129,basket/ep030
+,,,,Flip the basket upright,31,t002__ep031__basket,False,109,639,50.4,6,,,True,True,gemini-3.5-flash,89,basket/ep031
+,,,,Flip the basket upright,32,t002__ep032__basket,False,74,427,50.6,6,,,True,True,gemini-3.5-flash,152,basket/ep032
+,,,,Flip the basket upright,33,t002__ep033__basket,True,100,580,50.4,6,,,True,False,gemini-3.5-flash,,basket/ep033
+,,,,Flip the basket upright,34,t002__ep034__basket,False,64,369,50.7,6,,,True,True,gemini-3.5-flash,207,basket/ep034
+,,,,Flip the basket upright,35,t002__ep035__basket,True,55,321,50.5,6,,,True,False,gemini-3.5-flash,,basket/ep035
+,,,,Flip the basket upright,36,t002__ep036__basket,False,53,304,50.5,6,,,True,True,gemini-3.5-flash,151,basket/ep036
+,,,,Flip the basket upright,37,t002__ep037__basket,True,56,325,50.4,6,,,True,False,gemini-3.5-flash,,basket/ep037
+,,,,Flip the basket upright,38,t002__ep038__basket,False,52,302,50.3,6,,,True,True,gemini-3.5-flash,194,basket/ep038
+,,,,Flip the basket upright,39,t002__ep039__basket,True,55,317,50.3,6,,,True,False,gemini-3.5-flash,,basket/ep039
+,,,,Flip the basket upright,40,t002__ep040__basket,False,80,469,50.5,6,,,True,True,gemini-3.5-flash,158,basket/ep040
+,,,,Flip the basket upright,41,t002__ep041__basket,False,52,303,51.0,6,,,True,True,gemini-3.5-flash,162,basket/ep041
+,,,,Flip the basket upright,42,t002__ep042__basket,False,46,263,51.6,6,,,True,True,gemini-3.5-flash,134,basket/ep042
+,,,,Flip the basket upright,43,t002__ep043__basket,False,88,510,50.1,6,,,True,True,gemini-3.5-flash,225,basket/ep043
+,,,,Flip the basket upright,44,t002__ep044__basket,False,74,432,50.4,6,,,True,True,gemini-3.5-flash,165,basket/ep044
+,,,,Flip the basket upright,45,t002__ep045__basket,False,63,361,51.2,6,,,True,True,gemini-3.5-flash,135,basket/ep045
+,,,,Flip the basket upright,46,t002__ep046__basket,False,86,499,50.3,6,,,True,True,gemini-3.5-flash,156,basket/ep046
+,,,,Flip the basket upright,47,t002__ep047__basket,False,50,289,50.4,6,,,True,True,gemini-3.5-flash,135,basket/ep047
+,,,,Flip the basket upright,48,t002__ep048__basket,True,84,489,50.2,6,,,True,False,gemini-3.5-flash,,basket/ep048
+,,,,Flip the basket upright,49,t002__ep049__basket,True,86,497,50.4,6,,,True,False,gemini-3.5-flash,,basket/ep049
+,,,,Stack the yellow block on the wooden block,0,t000__ep000__blocks,False,160,929,50.9,6,,,True,True,gemini-3.5-flash,238,blocks/ep000
+,,,,Stack the yellow block on the wooden block,1,t000__ep001__blocks,False,136,790,50.9,6,,,True,True,gemini-3.5-flash,300,blocks/ep001
+,,,,Stack the yellow block on the wooden block,2,t000__ep002__blocks,False,55,316,50.2,6,,,True,True,gemini-3.5-flash,210,blocks/ep002
+,,,,Stack the yellow block on the wooden block,3,t000__ep003__blocks,False,337,1955,50.4,6,,,True,True,gemini-3.5-flash,756,blocks/ep003
+,,,,Stack the yellow block on the wooden block,4,t000__ep004__blocks,False,60,349,50.8,6,,,True,True,gemini-3.5-flash,216,blocks/ep004
+,,,,Stack the yellow block on the wooden block,5,t000__ep005__blocks,False,108,622,50.3,6,,,True,True,gemini-3.5-flash,540,blocks/ep005
+,,,,Stack the yellow block on the wooden block,6,t000__ep006__blocks,True,130,760,50.6,6,,,True,False,gemini-3.5-flash,,blocks/ep006
+,,,,Stack the yellow block on the wooden block,7,t000__ep007__blocks,False,60,342,50.5,6,,,True,True,gemini-3.5-flash,181,blocks/ep007
+,,,,Stack the yellow block on the wooden block,8,t000__ep008__blocks,False,66,381,51.6,6,,,True,True,gemini-3.5-flash,209,blocks/ep008
+,,,,Stack the yellow block on the wooden block,9,t000__ep009__blocks,True,115,670,51.1,6,,,True,False,gemini-3.5-flash,,blocks/ep009
+,,,,Stack the yellow block on the wooden block,10,t000__ep010__blocks,False,61,359,50.8,6,,,True,True,gemini-3.5-flash,231,blocks/ep010
+,,,,Stack the yellow block on the wooden block,11,t000__ep011__blocks,False,73,417,51.2,6,,,True,True,gemini-3.5-flash,3,blocks/ep011
+,,,,Stack the yellow block on the wooden block,12,t000__ep012__blocks,False,52,297,50.0,5,,,True,True,gemini-3.5-flash,251,blocks/ep012
+,,,,Stack the yellow block on the wooden block,13,t000__ep013__blocks,False,275,1610,50.4,6,,,True,True,gemini-3.5-flash,1129,blocks/ep013
+,,,,Stack the yellow block on the wooden block,14,t000__ep014__blocks,False,75,443,50.2,6,,,True,True,gemini-3.5-flash,183,blocks/ep014
+,,,,Stack the yellow block on the wooden block,15,t000__ep015__blocks,False,185,1080,50.6,6,,,True,True,gemini-3.5-flash,491,blocks/ep015
+,,,,Stack the yellow block on the wooden block,16,t000__ep016__blocks,False,162,932,50.5,6,,,True,True,gemini-3.5-flash,177,blocks/ep016
+,,,,Stack the yellow block on the wooden block,17,t000__ep017__blocks,False,82,475,50.2,6,,,True,True,gemini-3.5-flash,261,blocks/ep017
+,,,,Stack the yellow block on the wooden block,18,t000__ep018__blocks,False,99,571,50.6,6,,,True,True,gemini-3.5-flash,439,blocks/ep018
+,,,,Stack the yellow block on the wooden block,19,t000__ep019__blocks,True,165,964,50.7,6,,,True,False,gemini-3.5-flash,,blocks/ep019
+,,,,Stack the yellow block on the wooden block,20,t000__ep020__blocks,True,371,2153,50.4,6,,,True,False,gemini-3.5-flash,,blocks/ep020
+,,,,Stack the yellow block on the wooden block,21,t000__ep021__blocks,False,88,505,50.5,6,,,True,True,gemini-3.5-flash,256,blocks/ep021
+,,,,Stack the yellow block on the wooden block,22,t000__ep022__blocks,False,96,563,50.6,6,,,True,True,gemini-3.5-flash,253,blocks/ep022
+,,,,Stack the yellow block on the wooden block,23,t000__ep023__blocks,False,88,511,50.4,6,,,True,True,gemini-3.5-flash,303,blocks/ep023
+,,,,Stack the yellow block on the wooden block,24,t000__ep024__blocks,False,82,481,50.5,6,,,True,True,gemini-3.5-flash,317,blocks/ep024
+,,,,Stack the yellow block on the wooden block,25,t000__ep025__blocks,False,178,1029,50.3,6,,,True,True,gemini-3.5-flash,314,blocks/ep025
+,,,,Stack the yellow block on the wooden block,26,t000__ep026__blocks,False,99,577,50.6,6,,,True,True,gemini-3.5-flash,375,blocks/ep026
+,,,,Stack the yellow block on the wooden block,27,t000__ep027__blocks,False,112,645,50.2,6,,,True,True,gemini-3.5-flash,254,blocks/ep027
+,,,,Stack the yellow block on the wooden block,28,t000__ep028__blocks,False,160,934,50.4,6,,,True,True,gemini-3.5-flash,835,blocks/ep028
+,,,,Stack the yellow block on the wooden block,29,t000__ep029__blocks,False,79,461,50.7,6,,,True,True,gemini-3.5-flash,3,blocks/ep029
+,,,,Stack the yellow block on the wooden block,30,t000__ep030__blocks,False,68,394,50.7,6,,,True,True,gemini-3.5-flash,214,blocks/ep030
+,,,,Stack the yellow block on the wooden block,31,t000__ep031__blocks,False,63,364,50.0,6,,,True,True,gemini-3.5-flash,203,blocks/ep031
+,,,,Stack the yellow block on the wooden block,32,t000__ep032__blocks,True,206,1200,50.4,6,,,True,False,gemini-3.5-flash,,blocks/ep032
+,,,,Stack the yellow block on the wooden block,33,t000__ep033__blocks,False,169,981,50.1,6,,,True,True,gemini-3.5-flash,375,blocks/ep033
+,,,,Stack the yellow block on the wooden block,34,t000__ep034__blocks,False,160,930,50.4,6,,,True,True,gemini-3.5-flash,257,blocks/ep034
+,,,,Stack the yellow block on the wooden block,35,t000__ep035__blocks,False,200,1165,50.4,6,,,True,True,gemini-3.5-flash,608,blocks/ep035
+,,,,Stack the yellow block on the wooden block,36,t000__ep036__blocks,False,108,618,50.2,6,,,True,True,gemini-3.5-flash,242,blocks/ep036
+,,,,Stack the yellow block on the wooden block,37,t000__ep037__blocks,True,149,871,50.4,6,,,True,False,gemini-3.5-flash,,blocks/ep037
+,,,,Stack the yellow block on the wooden block,38,t000__ep038__blocks,False,301,1749,50.4,6,,,True,True,gemini-3.5-flash,412,blocks/ep038
+,,,,Stack the yellow block on the wooden block,39,t000__ep039__blocks,False,109,635,50.6,6,,,True,True,gemini-3.5-flash,521,blocks/ep039
+,,,,Stack the yellow block on the wooden block,40,t000__ep040__blocks,False,74,426,50.7,6,,,True,True,gemini-3.5-flash,239,blocks/ep040
+,,,,Stack the yellow block on the wooden block,41,t000__ep041__blocks,False,151,877,50.2,6,,,True,True,gemini-3.5-flash,752,blocks/ep041
+,,,,Stack the yellow block on the wooden block,42,t000__ep042__blocks,False,220,1285,50.3,6,,,True,True,gemini-3.5-flash,679,blocks/ep042
+,,,,Stack the yellow block on the wooden block,43,t000__ep043__blocks,False,118,679,50.6,6,,,True,True,gemini-3.5-flash,209,blocks/ep043
+,,,,Stack the yellow block on the wooden block,44,t000__ep044__blocks,False,76,439,50.8,6,,,True,True,gemini-3.5-flash,250,blocks/ep044
+,,,,Stack the yellow block on the wooden block,45,t000__ep045__blocks,False,119,695,50.4,6,,,True,True,gemini-3.5-flash,254,blocks/ep045
+,,,,Stack the yellow block on the wooden block,46,t000__ep046__blocks,False,144,836,50.4,6,,,True,True,gemini-3.5-flash,202,blocks/ep046
+,,,,Stack the yellow block on the wooden block,47,t000__ep047__blocks,False,142,833,51.0,6,,,True,True,gemini-3.5-flash,448,blocks/ep047
+,,,,Stack the yellow block on the wooden block,48,t000__ep048__blocks,False,221,1281,50.5,6,,,True,True,gemini-3.5-flash,354,blocks/ep048
+,,,,Stack the yellow block on the wooden block,49,t000__ep049__blocks,False,147,844,50.2,6,,,True,True,gemini-3.5-flash,559,blocks/ep049
+,,,,Close the drawer,0,t006__ep000__closer_drawer,True,108,624,50.8,5,,,True,False,gemini-3.5-flash,,closer_drawer/ep000
+,,,,Close the drawer,1,t006__ep001__closer_drawer,True,112,646,50.4,6,,,True,False,gemini-3.5-flash,,closer_drawer/ep001
+,,,,Close the drawer,2,t006__ep002__closer_drawer,True,118,683,50.3,6,,,True,False,gemini-3.5-flash,,closer_drawer/ep002
+,,,,Close the drawer,3,t006__ep003__closer_drawer,True,99,575,50.4,6,,,True,False,gemini-3.5-flash,,closer_drawer/ep003
+,,,,Close the drawer,4,t006__ep004__closer_drawer,True,122,704,50.3,6,,,True,False,gemini-3.5-flash,,closer_drawer/ep004
+,,,,Close the drawer,5,t006__ep005__closer_drawer,True,124,715,50.4,6,,,True,False,gemini-3.5-flash,,closer_drawer/ep005
+,,,,Close the drawer,6,t006__ep006__closer_drawer,True,118,692,51.5,6,,,True,False,gemini-3.5-flash,,closer_drawer/ep006
+,,,,Close the drawer,7,t006__ep007__closer_drawer,True,118,687,50.2,6,,,True,False,gemini-3.5-flash,,closer_drawer/ep007
+,,,,Close the drawer,8,t006__ep008__closer_drawer,True,117,681,50.4,6,,,True,False,gemini-3.5-flash,,closer_drawer/ep008
+,,,,Close the drawer,9,t006__ep009__closer_drawer,True,115,671,50.6,6,,,True,False,gemini-3.5-flash,,closer_drawer/ep009
+,,,,Close the drawer,10,t006__ep010__closer_drawer,False,70,409,51.0,6,,,True,True,gemini-3.5-flash,408,closer_drawer/ep010
+,,,,Close the drawer,11,t006__ep011__closer_drawer,True,58,336,50.2,6,,,True,False,gemini-3.5-flash,,closer_drawer/ep011
+,,,,Close the drawer,12,t006__ep012__closer_drawer,True,61,350,50.5,6,,,True,False,gemini-3.5-flash,,closer_drawer/ep012
+,,,,Close the drawer,13,t006__ep013__closer_drawer,False,49,282,50.3,6,,,True,True,gemini-3.5-flash,281,closer_drawer/ep013
+,,,,Close the drawer,14,t006__ep014__closer_drawer,True,64,374,50.7,6,,,True,False,gemini-3.5-flash,,closer_drawer/ep014
+,,,,Close the drawer,15,t006__ep015__closer_drawer,True,53,307,51.2,6,,,True,False,gemini-3.5-flash,,closer_drawer/ep015
+,,,,Close the drawer,16,t006__ep016__closer_drawer,True,49,285,50.5,6,,,True,False,gemini-3.5-flash,,closer_drawer/ep016
+,,,,Close the drawer,17,t006__ep017__closer_drawer,True,60,346,50.1,5,,,True,False,gemini-3.5-flash,,closer_drawer/ep017
+,,,,Close the drawer,18,t006__ep018__closer_drawer,False,42,247,50.7,6,,,True,True,gemini-3.5-flash,158,closer_drawer/ep018
+,,,,Close the drawer,19,t006__ep019__closer_drawer,False,40,226,50.5,6,,,True,True,gemini-3.5-flash,54,closer_drawer/ep019
+,,,,Close the drawer,20,t006__ep020__closer_drawer,True,52,305,50.4,6,,,True,False,gemini-3.5-flash,,closer_drawer/ep020
+,,,,Close the drawer,21,t006__ep021__closer_drawer,False,59,343,50.6,6,,,True,True,gemini-3.5-flash,238,closer_drawer/ep021
+,,,,Close the drawer,22,t006__ep022__closer_drawer,False,88,515,50.4,6,,,True,True,gemini-3.5-flash,274,closer_drawer/ep022
+,,,,Close the drawer,23,t006__ep023__closer_drawer,True,145,851,50.6,6,,,True,False,gemini-3.5-flash,,closer_drawer/ep023
+,,,,Close the drawer,24,t006__ep024__closer_drawer,False,101,586,50.4,6,,,True,True,gemini-3.5-flash,390,closer_drawer/ep024
+,,,,Close the drawer,25,t006__ep025__closer_drawer,False,106,617,50.2,6,,,True,True,gemini-3.5-flash,387,closer_drawer/ep025
+,,,,Close the drawer,26,t006__ep026__closer_drawer,True,96,557,51.0,6,,,True,False,gemini-3.5-flash,,closer_drawer/ep026
+,,,,Close the drawer,27,t006__ep027__closer_drawer,False,101,590,50.0,6,,,True,True,gemini-3.5-flash,210,closer_drawer/ep027
+,,,,Close the drawer,28,t006__ep028__closer_drawer,True,125,725,50.5,6,,,True,False,gemini-3.5-flash,,closer_drawer/ep028
+,,,,Close the drawer,29,t006__ep029__closer_drawer,False,44,251,50.9,5,,,True,True,gemini-3.5-flash,138,closer_drawer/ep029
+,,,,Close the drawer,30,t006__ep030__closer_drawer,False,97,569,50.7,6,,,True,True,gemini-3.5-flash,250,closer_drawer/ep030
+,,,,Close the drawer,31,t006__ep031__closer_drawer,False,44,253,50.6,6,,,True,True,gemini-3.5-flash,112,closer_drawer/ep031
+,,,,Close the drawer,32,t006__ep032__closer_drawer,True,48,279,50.4,6,,,True,False,gemini-3.5-flash,,closer_drawer/ep032
+,,,,Close the drawer,33,t006__ep033__closer_drawer,False,54,311,50.1,6,,,True,True,gemini-3.5-flash,81,closer_drawer/ep033
+,,,,Close the drawer,34,t006__ep034__closer_drawer,False,32,181,50.3,6,,,True,True,gemini-3.5-flash,120,closer_drawer/ep034
+,,,,Close the drawer,35,t006__ep035__closer_drawer,False,39,226,50.1,6,,,True,True,gemini-3.5-flash,130,closer_drawer/ep035
+,,,,Close the drawer,36,t006__ep036__closer_drawer,False,35,209,50.8,6,,,True,True,gemini-3.5-flash,105,closer_drawer/ep036
+,,,,Close the drawer,37,t006__ep037__closer_drawer,False,44,259,51.1,6,,,True,True,gemini-3.5-flash,138,closer_drawer/ep037
+,,,,Close the drawer,38,t006__ep038__closer_drawer,True,53,311,50.5,6,,,True,False,gemini-3.5-flash,,closer_drawer/ep038
+,,,,Close the drawer,39,t006__ep039__closer_drawer,True,44,257,50.1,6,,,True,False,gemini-3.5-flash,,closer_drawer/ep039
+,,,,Close the drawer,40,t006__ep040__closer_drawer,True,35,206,50.3,5,,,True,False,gemini-3.5-flash,,closer_drawer/ep040
+,,,,Close the drawer,41,t006__ep041__closer_drawer,True,57,332,49.8,6,,,True,False,gemini-3.5-flash,,closer_drawer/ep041
+,,,,Close the drawer,42,t006__ep042__closer_drawer,False,52,299,50.9,6,,,True,True,gemini-3.5-flash,151,closer_drawer/ep042
+,,,,Close the drawer,43,t006__ep043__closer_drawer,True,47,279,50.4,6,,,True,False,gemini-3.5-flash,,closer_drawer/ep043
+,,,,Close the drawer,44,t006__ep044__closer_drawer,True,37,209,50.9,5,,,True,False,gemini-3.5-flash,,closer_drawer/ep044
+,,,,Close the drawer,45,t006__ep045__closer_drawer,False,46,267,50.9,6,,,True,True,gemini-3.5-flash,190,closer_drawer/ep045
+,,,,Close the drawer,46,t006__ep046__closer_drawer,False,52,304,50.1,6,,,True,True,gemini-3.5-flash,200,closer_drawer/ep046
+,,,,Close the drawer,47,t006__ep047__closer_drawer,False,54,311,50.9,6,,,True,True,gemini-3.5-flash,175,closer_drawer/ep047
+,,,,Close the drawer,48,t006__ep048__closer_drawer,False,49,286,50.1,6,,,True,True,gemini-3.5-flash,193,closer_drawer/ep048
+,,,,Close the drawer,49,t006__ep049__closer_drawer,True,55,318,50.6,6,,,True,False,gemini-3.5-flash,,closer_drawer/ep049
+,,,,Take the eggplant off the plate,0,t004__ep000__eggplant,False,72,411,50.2,5,,,True,True,gemini-3.5-flash,180,eggplant/ep000
+,,,,Take the eggplant off the plate,1,t004__ep001__eggplant,True,66,379,50.2,6,,,True,False,gemini-3.5-flash,,eggplant/ep001
+,,,,Take the eggplant off the plate,2,t004__ep002__eggplant,True,72,423,51.0,6,,,True,False,gemini-3.5-flash,,eggplant/ep002
+,,,,Take the eggplant off the plate,3,t004__ep003__eggplant,False,55,320,51.1,6,,,True,True,gemini-3.5-flash,181,eggplant/ep003
+,,,,Take the eggplant off the plate,4,t004__ep004__eggplant,True,72,420,51.2,6,,,True,False,gemini-3.5-flash,,eggplant/ep004
+,,,,Take the eggplant off the plate,5,t004__ep005__eggplant,True,61,347,50.4,6,,,True,False,gemini-3.5-flash,,eggplant/ep005
+,,,,Take the eggplant off the plate,6,t004__ep006__eggplant,True,69,400,50.4,6,,,True,False,gemini-3.5-flash,,eggplant/ep006
+,,,,Take the eggplant off the plate,7,t004__ep007__eggplant,False,60,340,50.4,6,,,True,True,gemini-3.5-flash,234,eggplant/ep007
+,,,,Take the eggplant off the plate,8,t004__ep008__eggplant,True,67,388,50.0,6,,,True,False,gemini-3.5-flash,,eggplant/ep008
+,,,,Take the eggplant off the plate,9,t004__ep009__eggplant,True,80,463,50.6,6,,,True,False,gemini-3.5-flash,,eggplant/ep009
+,,,,Take the eggplant off the plate,10,t004__ep010__eggplant,True,73,423,50.7,6,,,True,False,gemini-3.5-flash,,eggplant/ep010
+,,,,Take the eggplant off the plate,11,t004__ep011__eggplant,True,71,411,50.6,6,,,True,False,gemini-3.5-flash,,eggplant/ep011
+,,,,Take the eggplant off the plate,12,t004__ep012__eggplant,True,70,410,50.9,6,,,True,False,gemini-3.5-flash,,eggplant/ep012
+,,,,Take the eggplant off the plate,13,t004__ep013__eggplant,True,62,366,51.0,6,,,True,False,gemini-3.5-flash,,eggplant/ep013
+,,,,Take the eggplant off the plate,14,t004__ep014__eggplant,True,72,418,50.0,6,,,True,False,gemini-3.5-flash,,eggplant/ep014
+,,,,Take the eggplant off the plate,15,t004__ep015__eggplant,True,70,400,50.1,6,,,True,False,gemini-3.5-flash,,eggplant/ep015
+,,,,Take the eggplant off the plate,16,t004__ep016__eggplant,True,66,380,50.7,6,,,True,False,gemini-3.5-flash,,eggplant/ep016
+,,,,Take the eggplant off the plate,17,t004__ep017__eggplant,False,47,271,50.2,5,,,True,True,gemini-3.5-flash,167,eggplant/ep017
+,,,,Take the eggplant off the plate,18,t004__ep018__eggplant,True,62,360,49.7,6,,,True,False,gemini-3.5-flash,,eggplant/ep018
+,,,,Take the eggplant off the plate,19,t004__ep019__eggplant,False,59,345,51.0,6,,,True,True,gemini-3.5-flash,219,eggplant/ep019
+,,,,Take the eggplant off the plate,20,t004__ep020__eggplant,True,62,356,50.8,6,,,True,False,gemini-3.5-flash,,eggplant/ep020
+,,,,Take the eggplant off the plate,21,t004__ep021__eggplant,False,76,437,50.1,6,,,True,True,gemini-3.5-flash,261,eggplant/ep021
+,,,,Take the eggplant off the plate,22,t004__ep022__eggplant,False,53,300,50.3,5,,,True,True,gemini-3.5-flash,141,eggplant/ep022
+,,,,Take the eggplant off the plate,23,t004__ep023__eggplant,False,117,674,50.7,6,,,True,True,gemini-3.5-flash,252,eggplant/ep023
+,,,,Take the eggplant off the plate,24,t004__ep024__eggplant,False,87,501,50.8,6,,,True,True,gemini-3.5-flash,168,eggplant/ep024
+,,,,Take the eggplant off the plate,25,t004__ep025__eggplant,True,73,419,50.7,6,,,True,False,gemini-3.5-flash,,eggplant/ep025
+,,,,Take the eggplant off the plate,26,t004__ep026__eggplant,True,79,463,50.5,6,,,True,False,gemini-3.5-flash,,eggplant/ep026
+,,,,Take the eggplant off the plate,27,t004__ep027__eggplant,False,50,287,50.7,6,,,True,True,gemini-3.5-flash,162,eggplant/ep027
+,,,,Take the eggplant off the plate,28,t004__ep028__eggplant,False,54,316,50.7,6,,,True,True,gemini-3.5-flash,209,eggplant/ep028
+,,,,Take the eggplant off the plate,29,t004__ep029__eggplant,False,67,384,50.3,6,,,True,True,gemini-3.5-flash,274,eggplant/ep029
+,,,,Take the eggplant off the plate,30,t004__ep030__eggplant,True,100,577,50.7,6,,,True,False,gemini-3.5-flash,,eggplant/ep030
+,,,,Take the eggplant off the plate,31,t004__ep031__eggplant,False,83,484,50.2,6,,,True,True,gemini-3.5-flash,190,eggplant/ep031
+,,,,Take the eggplant off the plate,32,t004__ep032__eggplant,False,63,367,50.5,6,,,True,True,gemini-3.5-flash,164,eggplant/ep032
+,,,,Take the eggplant off the plate,33,t004__ep033__eggplant,True,71,407,50.4,6,,,True,False,gemini-3.5-flash,,eggplant/ep033
+,,,,Take the eggplant off the plate,34,t004__ep034__eggplant,False,70,406,51.5,6,,,True,True,gemini-3.5-flash,206,eggplant/ep034
+,,,,Take the eggplant off the plate,35,t004__ep035__eggplant,False,55,321,51.0,6,,,True,True,gemini-3.5-flash,160,eggplant/ep035
+,,,,Take the eggplant off the plate,36,t004__ep036__eggplant,True,65,377,50.4,6,,,True,False,gemini-3.5-flash,,eggplant/ep036
+,,,,Take the eggplant off the plate,37,t004__ep037__eggplant,True,69,402,50.4,6,,,True,False,gemini-3.5-flash,,eggplant/ep037
+,,,,Take the eggplant off the plate,38,t004__ep038__eggplant,True,89,515,50.5,6,,,True,False,gemini-3.5-flash,,eggplant/ep038
+,,,,Take the eggplant off the plate,39,t004__ep039__eggplant,False,144,837,50.4,6,,,True,True,gemini-3.5-flash,158,eggplant/ep039
+,,,,Take the eggplant off the plate,40,t004__ep040__eggplant,False,67,392,50.5,6,,,True,True,gemini-3.5-flash,150,eggplant/ep040
+,,,,Take the eggplant off the plate,41,t004__ep041__eggplant,False,101,576,50.3,6,,,True,True,gemini-3.5-flash,431,eggplant/ep041
+,,,,Take the eggplant off the plate,42,t004__ep042__eggplant,False,149,861,50.4,6,,,True,True,gemini-3.5-flash,164,eggplant/ep042
+,,,,Take the eggplant off the plate,43,t004__ep043__eggplant,False,60,344,50.2,6,,,True,True,gemini-3.5-flash,168,eggplant/ep043
+,,,,Take the eggplant off the plate,44,t004__ep044__eggplant,True,75,436,50.5,6,,,True,False,gemini-3.5-flash,,eggplant/ep044
+,,,,Take the eggplant off the plate,45,t004__ep045__eggplant,False,78,458,50.6,6,,,True,True,gemini-3.5-flash,243,eggplant/ep045
+,,,,Take the eggplant off the plate,46,t004__ep046__eggplant,True,76,438,50.4,6,,,True,False,gemini-3.5-flash,,eggplant/ep046
+,,,,Take the eggplant off the plate,47,t004__ep047__eggplant,False,62,360,50.5,6,,,True,True,gemini-3.5-flash,148,eggplant/ep047
+,,,,Take the eggplant off the plate,48,t004__ep048__eggplant,False,100,575,50.7,5,,,True,True,gemini-3.5-flash,191,eggplant/ep048
+,,,,Take the eggplant off the plate,49,t004__ep049__eggplant,False,87,502,51.2,6,,,True,True,gemini-3.5-flash,240,eggplant/ep049
+,,,,Put the mushroom on the plate,0,t001__ep000__mushroom,False,88,508,50.1,6,,,True,True,gemini-3.5-flash,276,mushroom/ep000
+,,,,Put the mushroom on the plate,1,t001__ep001__mushroom,False,80,464,50.4,6,,,True,True,gemini-3.5-flash,286,mushroom/ep001
+,,,,Put the mushroom on the plate,2,t001__ep002__mushroom,False,90,519,50.2,6,,,True,True,gemini-3.5-flash,201,mushroom/ep002
+,,,,Put the mushroom on the plate,3,t001__ep003__mushroom,False,73,423,50.5,6,,,True,True,gemini-3.5-flash,346,mushroom/ep003
+,,,,Put the mushroom on the plate,4,t001__ep004__mushroom,False,82,475,50.6,6,,,True,True,gemini-3.5-flash,210,mushroom/ep004
+,,,,Put the mushroom on the plate,5,t001__ep005__mushroom,False,67,386,50.5,6,,,True,True,gemini-3.5-flash,177,mushroom/ep005
+,,,,Put the mushroom on the plate,6,t001__ep006__mushroom,True,68,392,50.4,6,,,True,False,gemini-3.5-flash,,mushroom/ep006
+,,,,Put the mushroom on the plate,7,t001__ep007__mushroom,False,69,393,50.3,5,,,True,True,gemini-3.5-flash,191,mushroom/ep007
+,,,,Put the mushroom on the plate,8,t001__ep008__mushroom,False,64,368,50.2,6,,,True,True,gemini-3.5-flash,128,mushroom/ep008
+,,,,Put the mushroom on the plate,9,t001__ep009__mushroom,False,53,306,50.2,6,,,True,True,gemini-3.5-flash,185,mushroom/ep009
+,,,,Put the mushroom on the plate,10,t001__ep010__mushroom,False,80,469,50.9,6,,,True,True,gemini-3.5-flash,279,mushroom/ep010
+,,,,Put the mushroom on the plate,11,t001__ep011__mushroom,False,73,415,50.2,6,,,True,True,gemini-3.5-flash,165,mushroom/ep011
+,,,,Put the mushroom on the plate,12,t001__ep012__mushroom,False,70,408,50.8,6,,,True,True,gemini-3.5-flash,156,mushroom/ep012
+,,,,Put the mushroom on the plate,13,t001__ep013__mushroom,False,70,405,50.9,6,,,True,True,gemini-3.5-flash,309,mushroom/ep013
+,,,,Put the mushroom on the plate,14,t001__ep014__mushroom,False,80,462,50.3,6,,,True,True,gemini-3.5-flash,341,mushroom/ep014
+,,,,Put the mushroom on the plate,15,t001__ep015__mushroom,True,64,370,50.3,6,,,True,False,gemini-3.5-flash,,mushroom/ep015
+,,,,Put the mushroom on the plate,16,t001__ep016__mushroom,True,69,401,51.3,6,,,True,False,gemini-3.5-flash,,mushroom/ep016
+,,,,Put the mushroom on the plate,17,t001__ep017__mushroom,False,53,306,50.4,6,,,True,True,gemini-3.5-flash,94,mushroom/ep017
+,,,,Put the mushroom on the plate,18,t001__ep018__mushroom,False,55,315,50.1,6,,,True,True,gemini-3.5-flash,241,mushroom/ep018
+,,,,Put the mushroom on the plate,19,t001__ep019__mushroom,True,59,340,50.0,6,,,True,False,gemini-3.5-flash,,mushroom/ep019
+,,,,Put the mushroom on the plate,20,t001__ep020__mushroom,False,97,568,50.8,6,,,True,True,gemini-3.5-flash,239,mushroom/ep020
+,,,,Put the mushroom on the plate,21,t001__ep021__mushroom,False,54,306,50.7,6,,,True,True,gemini-3.5-flash,229,mushroom/ep021
+,,,,Put the mushroom on the plate,22,t001__ep022__mushroom,True,69,403,49.9,6,,,True,False,gemini-3.5-flash,,mushroom/ep022
+,,,,Put the mushroom on the plate,23,t001__ep023__mushroom,False,89,520,51.3,6,,,True,True,gemini-3.5-flash,232,mushroom/ep023
+,,,,Put the mushroom on the plate,24,t001__ep024__mushroom,False,59,336,49.9,6,,,True,True,gemini-3.5-flash,253,mushroom/ep024
+,,,,Put the mushroom on the plate,25,t001__ep025__mushroom,False,65,370,50.7,5,,,True,True,gemini-3.5-flash,301,mushroom/ep025
+,,,,Put the mushroom on the plate,26,t001__ep026__mushroom,False,62,360,50.3,6,,,True,True,gemini-3.5-flash,278,mushroom/ep026
+,,,,Put the mushroom on the plate,27,t001__ep027__mushroom,False,100,582,50.2,6,,,True,True,gemini-3.5-flash,315,mushroom/ep027
+,,,,Put the mushroom on the plate,28,t001__ep028__mushroom,True,63,360,49.9,6,,,True,False,gemini-3.5-flash,,mushroom/ep028
+,,,,Put the mushroom on the plate,29,t001__ep029__mushroom,True,65,382,50.6,6,,,True,False,gemini-3.5-flash,,mushroom/ep029
+,,,,Put the mushroom on the plate,30,t001__ep030__mushroom,True,72,413,50.6,6,,,True,False,gemini-3.5-flash,,mushroom/ep030
+,,,,Put the mushroom on the plate,31,t001__ep031__mushroom,True,100,581,51.0,6,,,True,False,gemini-3.5-flash,,mushroom/ep031
+,,,,Put the mushroom on the plate,32,t001__ep032__mushroom,False,66,380,50.8,6,,,True,True,gemini-3.5-flash,193,mushroom/ep032
+,,,,Put the mushroom on the plate,33,t001__ep033__mushroom,False,106,614,50.1,6,,,True,True,gemini-3.5-flash,503,mushroom/ep033
+,,,,Put the mushroom on the plate,34,t001__ep034__mushroom,False,59,334,50.1,5,,,True,True,gemini-3.5-flash,259,mushroom/ep034
+,,,,Put the mushroom on the plate,35,t001__ep035__mushroom,False,51,295,50.3,6,,,True,True,gemini-3.5-flash,178,mushroom/ep035
+,,,,Put the mushroom on the plate,36,t001__ep036__mushroom,False,65,379,50.8,6,,,True,True,gemini-3.5-flash,265,mushroom/ep036
+,,,,Put the mushroom on the plate,37,t001__ep037__mushroom,False,126,725,50.5,6,,,True,True,gemini-3.5-flash,185,mushroom/ep037
+,,,,Put the mushroom on the plate,38,t001__ep038__mushroom,False,100,581,50.5,6,,,True,True,gemini-3.5-flash,417,mushroom/ep038
+,,,,Put the mushroom on the plate,39,t001__ep039__mushroom,False,52,303,50.1,6,,,True,True,gemini-3.5-flash,218,mushroom/ep039
+,,,,Put the mushroom on the plate,40,t001__ep040__mushroom,True,69,398,50.6,6,,,True,False,gemini-3.5-flash,,mushroom/ep040
+,,,,Put the mushroom on the plate,41,t001__ep041__mushroom,False,54,308,50.4,5,,,True,True,gemini-3.5-flash,194,mushroom/ep041
+,,,,Put the mushroom on the plate,42,t001__ep042__mushroom,False,70,407,50.7,6,,,True,True,gemini-3.5-flash,323,mushroom/ep042
+,,,,Put the mushroom on the plate,43,t001__ep043__mushroom,False,97,563,50.5,6,,,True,True,gemini-3.5-flash,230,mushroom/ep043
+,,,,Put the mushroom on the plate,44,t001__ep044__mushroom,True,71,413,50.1,5,,,True,False,gemini-3.5-flash,,mushroom/ep044
+,,,,Put the mushroom on the plate,45,t001__ep045__mushroom,False,59,338,49.9,6,,,True,True,gemini-3.5-flash,98,mushroom/ep045
+,,,,Put the mushroom on the plate,46,t001__ep046__mushroom,False,98,567,50.2,6,,,True,True,gemini-3.5-flash,343,mushroom/ep046
+,,,,Put the mushroom on the plate,47,t001__ep047__mushroom,True,88,507,50.2,6,,,True,False,gemini-3.5-flash,,mushroom/ep047
+,,,,Put the mushroom on the plate,48,t001__ep048__mushroom,True,86,493,50.0,6,,,True,False,gemini-3.5-flash,,mushroom/ep048
+,,,,Take the pineapple out of the drawer,0,t007__ep000__pineapple_out_of_drawer,False,98,573,51.0,6,,,True,True,gemini-3.5-flash,136,pineapple_out_of_drawer/ep000
+,,,,Take the pineapple out of the drawer,1,t007__ep001__pineapple_out_of_drawer,False,167,975,50.5,6,,,True,True,gemini-3.5-flash,76,pineapple_out_of_drawer/ep001
+,,,,Take the pineapple out of the drawer,2,t007__ep002__pineapple_out_of_drawer,False,87,502,50.3,6,,,True,True,gemini-3.5-flash,213,pineapple_out_of_drawer/ep002
+,,,,Take the pineapple out of the drawer,3,t007__ep003__pineapple_out_of_drawer,False,108,626,51.0,6,,,True,True,gemini-3.5-flash,287,pineapple_out_of_drawer/ep003
+,,,,Take the pineapple out of the drawer,4,t007__ep004__pineapple_out_of_drawer,False,128,744,50.4,6,,,True,True,gemini-3.5-flash,216,pineapple_out_of_drawer/ep004
+,,,,Take the pineapple out of the drawer,5,t007__ep005__pineapple_out_of_drawer,False,59,342,50.0,6,,,True,True,gemini-3.5-flash,267,pineapple_out_of_drawer/ep005
+,,,,Take the pineapple out of the drawer,6,t007__ep006__pineapple_out_of_drawer,True,121,699,50.5,6,,,True,False,gemini-3.5-flash,,pineapple_out_of_drawer/ep006
+,,,,Take the pineapple out of the drawer,7,t007__ep007__pineapple_out_of_drawer,True,93,537,50.6,6,,,True,False,gemini-3.5-flash,,pineapple_out_of_drawer/ep007
+,,,,Take the pineapple out of the drawer,8,t007__ep008__pineapple_out_of_drawer,True,103,594,50.4,6,,,True,False,gemini-3.5-flash,,pineapple_out_of_drawer/ep008
+,,,,Take the pineapple out of the drawer,9,t007__ep009__pineapple_out_of_drawer,False,74,426,50.6,6,,,True,True,gemini-3.5-flash,248,pineapple_out_of_drawer/ep009
+,,,,Take the pineapple out of the drawer,10,t007__ep010__pineapple_out_of_drawer,False,112,648,51.2,6,,,True,True,gemini-3.5-flash,265,pineapple_out_of_drawer/ep010
+,,,,Take the pineapple out of the drawer,11,t007__ep011__pineapple_out_of_drawer,False,97,564,50.4,6,,,True,True,gemini-3.5-flash,249,pineapple_out_of_drawer/ep011
+,,,,Take the pineapple out of the drawer,12,t007__ep012__pineapple_out_of_drawer,False,128,744,50.4,6,,,True,True,gemini-3.5-flash,269,pineapple_out_of_drawer/ep012
+,,,,Take the pineapple out of the drawer,13,t007__ep013__pineapple_out_of_drawer,True,209,1228,50.7,6,,,True,False,gemini-3.5-flash,,pineapple_out_of_drawer/ep013
+,,,,Take the pineapple out of the drawer,14,t007__ep014__pineapple_out_of_drawer,False,187,1083,50.5,6,,,True,True,gemini-3.5-flash,328,pineapple_out_of_drawer/ep014
+,,,,Take the pineapple out of the drawer,15,t007__ep015__pineapple_out_of_drawer,False,121,691,50.3,6,,,True,True,gemini-3.5-flash,302,pineapple_out_of_drawer/ep015
+,,,,Take the pineapple out of the drawer,16,t007__ep016__pineapple_out_of_drawer,False,177,1025,50.3,6,,,True,True,gemini-3.5-flash,609,pineapple_out_of_drawer/ep016
+,,,,Take the pineapple out of the drawer,17,t007__ep017__pineapple_out_of_drawer,False,155,902,50.7,6,,,True,True,gemini-3.5-flash,458,pineapple_out_of_drawer/ep017
+,,,,Take the pineapple out of the drawer,18,t007__ep018__pineapple_out_of_drawer,True,186,1072,50.4,6,,,True,False,gemini-3.5-flash,,pineapple_out_of_drawer/ep018
+,,,,Take the pineapple out of the drawer,19,t007__ep019__pineapple_out_of_drawer,True,246,1427,50.5,6,,,True,False,gemini-3.5-flash,,pineapple_out_of_drawer/ep019
+,,,,Take the pineapple out of the drawer,20,t007__ep020__pineapple_out_of_drawer,False,162,946,50.4,6,,,True,True,gemini-3.5-flash,515,pineapple_out_of_drawer/ep020
+,,,,Take the pineapple out of the drawer,21,t007__ep021__pineapple_out_of_drawer,False,192,1118,50.4,6,,,True,True,gemini-3.5-flash,380,pineapple_out_of_drawer/ep021
+,,,,Take the pineapple out of the drawer,22,t007__ep022__pineapple_out_of_drawer,False,166,967,50.7,6,,,True,True,gemini-3.5-flash,574,pineapple_out_of_drawer/ep022
+,,,,Take the pineapple out of the drawer,23,t007__ep023__pineapple_out_of_drawer,False,143,832,50.5,6,,,True,True,gemini-3.5-flash,388,pineapple_out_of_drawer/ep023
+,,,,Take the pineapple out of the drawer,24,t007__ep024__pineapple_out_of_drawer,False,78,455,50.5,6,,,True,True,gemini-3.5-flash,147,pineapple_out_of_drawer/ep024
+,,,,Take the pineapple out of the drawer,25,t007__ep025__pineapple_out_of_drawer,False,93,540,51.3,6,,,True,True,gemini-3.5-flash,439,pineapple_out_of_drawer/ep025
+,,,,Take the pineapple out of the drawer,26,t007__ep026__pineapple_out_of_drawer,True,123,715,50.8,6,,,True,False,gemini-3.5-flash,,pineapple_out_of_drawer/ep026
+,,,,Take the pineapple out of the drawer,27,t007__ep027__pineapple_out_of_drawer,False,118,685,50.4,6,,,True,True,gemini-3.5-flash,181,pineapple_out_of_drawer/ep027
+,,,,Take the pineapple out of the drawer,28,t007__ep028__pineapple_out_of_drawer,False,107,620,50.4,6,,,True,True,gemini-3.5-flash,353,pineapple_out_of_drawer/ep028
+,,,,Take the pineapple out of the drawer,29,t007__ep029__pineapple_out_of_drawer,False,147,857,50.6,6,,,True,True,gemini-3.5-flash,243,pineapple_out_of_drawer/ep029
+,,,,Take the pineapple out of the drawer,30,t007__ep030__pineapple_out_of_drawer,True,161,941,50.3,6,,,True,False,gemini-3.5-flash,,pineapple_out_of_drawer/ep030
+,,,,Take the pineapple out of the drawer,31,t007__ep031__pineapple_out_of_drawer,True,211,1226,50.4,6,,,True,False,gemini-3.5-flash,,pineapple_out_of_drawer/ep031
+,,,,Take the pineapple out of the drawer,32,t007__ep032__pineapple_out_of_drawer,True,155,898,50.5,6,,,True,False,gemini-3.5-flash,,pineapple_out_of_drawer/ep032
+,,,,Take the pineapple out of the drawer,33,t007__ep033__pineapple_out_of_drawer,False,156,903,50.4,6,,,True,True,gemini-3.5-flash,362,pineapple_out_of_drawer/ep033
+,,,,Take the pineapple out of the drawer,34,t007__ep034__pineapple_out_of_drawer,False,163,944,50.4,6,,,True,True,gemini-3.5-flash,302,pineapple_out_of_drawer/ep034
+,,,,Take the pineapple out of the drawer,35,t007__ep035__pineapple_out_of_drawer,False,205,1184,50.9,6,,,True,True,gemini-3.5-flash,404,pineapple_out_of_drawer/ep035
+,,,,Take the pineapple out of the drawer,36,t007__ep036__pineapple_out_of_drawer,False,87,507,50.7,6,,,True,True,gemini-3.5-flash,395,pineapple_out_of_drawer/ep036
+,,,,Take the pineapple out of the drawer,37,t007__ep037__pineapple_out_of_drawer,False,85,495,50.6,6,,,True,True,gemini-3.5-flash,155,pineapple_out_of_drawer/ep037
+,,,,Take the pineapple out of the drawer,38,t007__ep038__pineapple_out_of_drawer,False,98,570,50.2,6,,,True,True,gemini-3.5-flash,341,pineapple_out_of_drawer/ep038
+,,,,Take the pineapple out of the drawer,39,t007__ep039__pineapple_out_of_drawer,False,103,594,50.5,6,,,True,True,gemini-3.5-flash,404,pineapple_out_of_drawer/ep039
+,,,,Take the pineapple out of the drawer,40,t007__ep040__pineapple_out_of_drawer,False,146,849,50.6,6,,,True,True,gemini-3.5-flash,632,pineapple_out_of_drawer/ep040
+,,,,Take the pineapple out of the drawer,41,t007__ep041__pineapple_out_of_drawer,False,171,998,51.1,6,,,True,True,gemini-3.5-flash,577,pineapple_out_of_drawer/ep041
+,,,,Take the pineapple out of the drawer,42,t007__ep042__pineapple_out_of_drawer,False,175,1016,50.4,6,,,True,True,gemini-3.5-flash,751,pineapple_out_of_drawer/ep042
+,,,,Take the pineapple out of the drawer,43,t007__ep043__pineapple_out_of_drawer,False,90,522,50.8,6,,,True,True,gemini-3.5-flash,437,pineapple_out_of_drawer/ep043
+,,,,Take the pineapple out of the drawer,44,t007__ep044__pineapple_out_of_drawer,False,116,671,50.9,6,,,True,True,gemini-3.5-flash,412,pineapple_out_of_drawer/ep044
+,,,,Take the pineapple out of the drawer,45,t007__ep045__pineapple_out_of_drawer,True,111,650,51.0,6,,,True,False,gemini-3.5-flash,,pineapple_out_of_drawer/ep045
+,,,,Take the pineapple out of the drawer,46,t007__ep046__pineapple_out_of_drawer,False,113,654,50.1,6,,,True,True,gemini-3.5-flash,306,pineapple_out_of_drawer/ep046
+,,,,Take the pineapple out of the drawer,47,t007__ep047__pineapple_out_of_drawer,False,94,548,50.3,6,,,True,True,gemini-3.5-flash,377,pineapple_out_of_drawer/ep047
+,,,,Take the pineapple out of the drawer,48,t007__ep048__pineapple_out_of_drawer,False,135,778,50.7,6,,,True,True,gemini-3.5-flash,194,pineapple_out_of_drawer/ep048
+,,,,Take the pineapple out of the drawer,49,t007__ep049__pineapple_out_of_drawer,True,163,942,50.3,6,,,True,False,gemini-3.5-flash,,pineapple_out_of_drawer/ep049
+,,,,Push the pineapple off the stove,0,t005__ep000__push_pineapple,True,49,288,50.2,6,,,True,False,gemini-3.5-flash,,push_pineapple/ep000
+,,,,Push the pineapple off the stove,1,t005__ep001__push_pineapple,True,62,361,50.4,6,,,True,False,gemini-3.5-flash,,push_pineapple/ep001
+,,,,Push the pineapple off the stove,2,t005__ep002__push_pineapple,True,56,319,50.7,6,,,True,False,gemini-3.5-flash,,push_pineapple/ep002
+,,,,Push the pineapple off the stove,3,t005__ep003__push_pineapple,True,141,820,50.6,6,,,True,False,gemini-3.5-flash,,push_pineapple/ep003
+,,,,Push the pineapple off the stove,4,t005__ep004__push_pineapple,True,99,577,51.0,6,,,True,False,gemini-3.5-flash,,push_pineapple/ep004
+,,,,Push the pineapple off the stove,5,t005__ep005__push_pineapple,True,57,336,50.6,6,,,True,False,gemini-3.5-flash,,push_pineapple/ep005
+,,,,Push the pineapple off the stove,6,t005__ep006__push_pineapple,True,56,323,50.5,6,,,True,False,gemini-3.5-flash,,push_pineapple/ep006
+,,,,Push the pineapple off the stove,7,t005__ep007__push_pineapple,True,66,387,50.7,6,,,True,False,gemini-3.5-flash,,push_pineapple/ep007
+,,,,Push the pineapple off the stove,8,t005__ep008__push_pineapple,True,71,411,50.1,6,,,True,False,gemini-3.5-flash,,push_pineapple/ep008
+,,,,Push the pineapple off the stove,9,t005__ep009__push_pineapple,True,72,416,50.5,6,,,True,False,gemini-3.5-flash,,push_pineapple/ep009
+,,,,Push the pineapple off the stove,10,t005__ep010__push_pineapple,True,110,639,50.8,6,,,True,False,gemini-3.5-flash,,push_pineapple/ep010
+,,,,Push the pineapple off the stove,11,t005__ep011__push_pineapple,True,68,392,50.4,6,,,True,False,gemini-3.5-flash,,push_pineapple/ep011
+,,,,Push the pineapple off the stove,12,t005__ep012__push_pineapple,True,78,445,50.9,6,,,True,False,gemini-3.5-flash,,push_pineapple/ep012
+,,,,Push the pineapple off the stove,13,t005__ep013__push_pineapple,True,55,318,50.4,6,,,True,False,gemini-3.5-flash,,push_pineapple/ep013
+,,,,Push the pineapple off the stove,14,t005__ep014__push_pineapple,True,58,334,50.5,6,,,True,False,gemini-3.5-flash,,push_pineapple/ep014
+,,,,Push the pineapple off the stove,15,t005__ep015__push_pineapple,True,96,556,50.4,6,,,True,False,gemini-3.5-flash,,push_pineapple/ep015
+,,,,Push the pineapple off the stove,16,t005__ep016__push_pineapple,True,43,249,50.4,6,,,True,False,gemini-3.5-flash,,push_pineapple/ep016
+,,,,Push the pineapple off the stove,17,t005__ep017__push_pineapple,True,69,401,50.2,6,,,True,False,gemini-3.5-flash,,push_pineapple/ep017
+,,,,Push the pineapple off the stove,18,t005__ep018__push_pineapple,True,42,243,50.2,5,,,True,False,gemini-3.5-flash,,push_pineapple/ep018
+,,,,Push the pineapple off the stove,19,t005__ep019__push_pineapple,True,47,271,50.6,6,,,True,False,gemini-3.5-flash,,push_pineapple/ep019
+,,,,Push the pineapple off the stove,20,t005__ep020__push_pineapple,False,75,437,51.0,6,,,True,True,gemini-3.5-flash,197,push_pineapple/ep020
+,,,,Push the pineapple off the stove,21,t005__ep021__push_pineapple,False,128,752,50.6,6,,,True,True,gemini-3.5-flash,152,push_pineapple/ep021
+,,,,Push the pineapple off the stove,22,t005__ep022__push_pineapple,False,71,415,50.6,6,,,True,True,gemini-3.5-flash,186,push_pineapple/ep022
+,,,,Push the pineapple off the stove,23,t005__ep023__push_pineapple,True,42,246,50.4,6,,,True,False,gemini-3.5-flash,,push_pineapple/ep023
+,,,,Push the pineapple off the stove,24,t005__ep024__push_pineapple,True,82,478,50.7,6,,,True,False,gemini-3.5-flash,,push_pineapple/ep024
+,,,,Push the pineapple off the stove,25,t005__ep025__push_pineapple,True,46,265,50.2,6,,,True,False,gemini-3.5-flash,,push_pineapple/ep025
+,,,,Push the pineapple off the stove,26,t005__ep026__push_pineapple,False,43,246,50.2,6,,,True,True,gemini-3.5-flash,118,push_pineapple/ep026
+,,,,Push the pineapple off the stove,27,t005__ep027__push_pineapple,False,91,527,50.6,6,,,True,True,gemini-3.5-flash,193,push_pineapple/ep027
+,,,,Push the pineapple off the stove,28,t005__ep028__push_pineapple,True,98,571,51.2,6,,,True,False,gemini-3.5-flash,,push_pineapple/ep028
+,,,,Push the pineapple off the stove,29,t005__ep029__push_pineapple,False,48,283,51.0,6,,,True,True,gemini-3.5-flash,168,push_pineapple/ep029
+,,,,Push the pineapple off the stove,30,t005__ep030__push_pineapple,False,48,277,50.5,6,,,True,True,gemini-3.5-flash,150,push_pineapple/ep030
+,,,,Push the pineapple off the stove,31,t005__ep031__push_pineapple,True,124,720,50.5,6,,,True,False,gemini-3.5-flash,,push_pineapple/ep031
+,,,,Push the pineapple off the stove,32,t005__ep032__push_pineapple,True,79,463,50.5,6,,,True,False,gemini-3.5-flash,,push_pineapple/ep032
+,,,,Push the pineapple off the stove,33,t005__ep033__push_pineapple,False,60,354,50.3,6,,,True,True,gemini-3.5-flash,185,push_pineapple/ep033
+,,,,Push the pineapple off the stove,34,t005__ep034__push_pineapple,False,68,392,50.6,6,,,True,True,gemini-3.5-flash,245,push_pineapple/ep034
+,,,,Push the pineapple off the stove,35,t005__ep035__push_pineapple,True,96,564,50.7,6,,,True,False,gemini-3.5-flash,,push_pineapple/ep035
+,,,,Push the pineapple off the stove,36,t005__ep036__push_pineapple,False,62,355,50.8,6,,,True,True,gemini-3.5-flash,105,push_pineapple/ep036
+,,,,Push the pineapple off the stove,37,t005__ep037__push_pineapple,False,50,291,50.8,6,,,True,True,gemini-3.5-flash,290,push_pineapple/ep037
+,,,,Push the pineapple off the stove,38,t005__ep038__push_pineapple,False,47,267,50.7,6,,,True,True,gemini-3.5-flash,266,push_pineapple/ep038
+,,,,Push the pineapple off the stove,39,t005__ep039__push_pineapple,False,45,257,50.5,6,,,True,True,gemini-3.5-flash,210,push_pineapple/ep039
+,,,,Push the pineapple off the stove,40,t005__ep040__push_pineapple,False,55,316,50.4,6,,,True,True,gemini-3.5-flash,156,push_pineapple/ep040
+,,,,Push the pineapple off the stove,41,t005__ep041__push_pineapple,False,61,352,50.8,6,,,True,True,gemini-3.5-flash,161,push_pineapple/ep041
+,,,,Push the pineapple off the stove,42,t005__ep042__push_pineapple,False,76,446,50.8,6,,,True,True,gemini-3.5-flash,176,push_pineapple/ep042
+,,,,Push the pineapple off the stove,43,t005__ep043__push_pineapple,False,108,630,50.4,6,,,True,True,gemini-3.5-flash,388,push_pineapple/ep043
+,,,,Push the pineapple off the stove,44,t005__ep044__push_pineapple,False,111,639,50.5,6,,,True,True,gemini-3.5-flash,203,push_pineapple/ep044
+,,,,Push the pineapple off the stove,45,t005__ep045__push_pineapple,False,129,756,50.7,6,,,True,True,gemini-3.5-flash,641,push_pineapple/ep045
+,,,,Push the pineapple off the stove,46,t005__ep046__push_pineapple,False,73,424,50.2,6,,,True,True,gemini-3.5-flash,194,push_pineapple/ep046
+,,,,Push the pineapple off the stove,47,t005__ep047__push_pineapple,False,112,657,50.6,6,,,True,True,gemini-3.5-flash,395,push_pineapple/ep047
+,,,,Push the pineapple off the stove,48,t005__ep048__push_pineapple,False,129,749,50.0,6,,,True,True,gemini-3.5-flash,56,push_pineapple/ep048
+,,,,Push the pineapple off the stove,49,t005__ep049__push_pineapple,False,113,670,51.4,6,,,True,True,gemini-3.5-flash,3,push_pineapple/ep049
+,,,,Put the tomato in the bin,0,t003__ep000__tomato,True,86,503,50.5,6,,,True,False,gemini-3.5-flash,,tomato/ep000
+,,,,Put the tomato in the bin,1,t003__ep001__tomato,True,76,434,50.2,6,,,True,False,gemini-3.5-flash,,tomato/ep001
+,,,,Put the tomato in the bin,2,t003__ep002__tomato,True,67,386,50.1,5,,,True,False,gemini-3.5-flash,,tomato/ep002
+,,,,Put the tomato in the bin,3,t003__ep003__tomato,True,99,572,50.2,6,,,True,False,gemini-3.5-flash,,tomato/ep003
+,,,,Put the tomato in the bin,4,t003__ep004__tomato,True,88,513,50.4,6,,,True,False,gemini-3.5-flash,,tomato/ep004
+,,,,Put the tomato in the bin,5,t003__ep005__tomato,False,72,421,51.1,6,,,True,True,gemini-3.5-flash,281,tomato/ep005
+,,,,Put the tomato in the bin,6,t003__ep006__tomato,True,92,531,49.8,6,,,True,False,gemini-3.5-flash,,tomato/ep006
+,,,,Put the tomato in the bin,7,t003__ep007__tomato,False,69,402,50.2,6,,,True,True,gemini-3.5-flash,362,tomato/ep007
+,,,,Put the tomato in the bin,8,t003__ep008__tomato,True,113,659,50.5,6,,,True,False,gemini-3.5-flash,,tomato/ep008
+,,,,Put the tomato in the bin,9,t003__ep009__tomato,True,89,516,50.4,6,,,True,False,gemini-3.5-flash,,tomato/ep009
+,,,,Put the tomato in the bin,10,t003__ep010__tomato,False,124,713,50.3,6,,,True,True,gemini-3.5-flash,52,tomato/ep010
+,,,,Put the tomato in the bin,11,t003__ep011__tomato,True,83,477,50.4,6,,,True,False,gemini-3.5-flash,,tomato/ep011
+,,,,Put the tomato in the bin,12,t003__ep012__tomato,True,86,500,50.8,6,,,True,False,gemini-3.5-flash,,tomato/ep012
+,,,,Put the tomato in the bin,13,t003__ep013__tomato,False,71,409,50.8,6,,,True,True,gemini-3.5-flash,273,tomato/ep013
+,,,,Put the tomato in the bin,14,t003__ep014__tomato,True,119,693,50.7,6,,,True,False,gemini-3.5-flash,,tomato/ep014
+,,,,Put the tomato in the bin,15,t003__ep015__tomato,True,150,875,50.6,6,,,True,False,gemini-3.5-flash,,tomato/ep015
+,,,,Put the tomato in the bin,16,t003__ep016__tomato,False,61,354,50.7,6,,,True,True,gemini-3.5-flash,0,tomato/ep016
+,,,,Put the tomato in the bin,17,t003__ep017__tomato,False,110,639,50.2,6,,,True,True,gemini-3.5-flash,110,tomato/ep017
+,,,,Put the tomato in the bin,18,t003__ep018__tomato,False,123,708,50.4,6,,,True,True,gemini-3.5-flash,655,tomato/ep018
+,,,,Put the tomato in the bin,19,t003__ep019__tomato,True,97,564,50.2,6,,,True,False,gemini-3.5-flash,,tomato/ep019
+,,,,Put the tomato in the bin,20,t003__ep020__tomato,True,110,639,50.5,6,,,True,False,gemini-3.5-flash,,tomato/ep020
+,,,,Put the tomato in the bin,21,t003__ep021__tomato,False,97,555,50.2,6,,,True,True,gemini-3.5-flash,326,tomato/ep021
+,,,,Put the tomato in the bin,22,t003__ep022__tomato,True,103,603,50.6,6,,,True,False,gemini-3.5-flash,,tomato/ep022
+,,,,Put the tomato in the bin,23,t003__ep023__tomato,True,89,519,50.2,6,,,True,False,gemini-3.5-flash,,tomato/ep023
+,,,,Put the tomato in the bin,24,t003__ep024__tomato,True,117,679,50.4,6,,,True,False,gemini-3.5-flash,,tomato/ep024
+,,,,Put the tomato in the bin,25,t003__ep025__tomato,True,80,463,50.4,6,,,True,False,gemini-3.5-flash,,tomato/ep025
+,,,,Put the tomato in the bin,26,t003__ep026__tomato,True,82,476,50.2,6,,,True,False,gemini-3.5-flash,,tomato/ep026
+,,,,Put the tomato in the bin,27,t003__ep027__tomato,True,156,909,50.0,6,,,True,False,gemini-3.5-flash,,tomato/ep027
+,,,,Put the tomato in the bin,28,t003__ep028__tomato,False,70,404,50.6,6,,,True,True,gemini-3.5-flash,156,tomato/ep028
+,,,,Put the tomato in the bin,29,t003__ep029__tomato,False,87,506,50.8,6,,,True,True,gemini-3.5-flash,114,tomato/ep029
+,,,,Put the tomato in the bin,30,t003__ep030__tomato,False,83,481,50.5,6,,,True,True,gemini-3.5-flash,294,tomato/ep030
+,,,,Put the tomato in the bin,31,t003__ep031__tomato,False,150,877,50.1,6,,,True,True,gemini-3.5-flash,130,tomato/ep031
+,,,,Put the tomato in the bin,32,t003__ep032__tomato,False,95,560,50.7,6,,,True,True,gemini-3.5-flash,301,tomato/ep032
+,,,,Put the tomato in the bin,33,t003__ep033__tomato,False,174,1010,50.1,6,,,True,True,gemini-3.5-flash,65,tomato/ep033
+,,,,Put the tomato in the bin,34,t003__ep034__tomato,False,131,753,50.1,6,,,True,True,gemini-3.5-flash,42,tomato/ep034
+,,,,Put the tomato in the bin,35,t003__ep035__tomato,False,139,807,50.4,6,,,True,True,gemini-3.5-flash,203,tomato/ep035
+,,,,Put the tomato in the bin,36,t003__ep036__tomato,False,161,925,50.2,6,,,True,True,gemini-3.5-flash,310,tomato/ep036
+,,,,Put the tomato in the bin,37,t003__ep037__tomato,False,83,488,50.8,6,,,True,True,gemini-3.5-flash,181,tomato/ep037
+,,,,Put the tomato in the bin,38,t003__ep038__tomato,True,132,764,50.3,6,,,True,False,gemini-3.5-flash,,tomato/ep038
+,,,,Put the tomato in the bin,39,t003__ep039__tomato,True,107,626,50.8,6,,,True,False,gemini-3.5-flash,,tomato/ep039
+,,,,Put the tomato in the bin,40,t003__ep040__tomato,True,117,678,50.5,6,,,True,False,gemini-3.5-flash,,tomato/ep040
+,,,,Put the tomato in the bin,41,t003__ep041__tomato,True,80,458,50.3,6,,,True,False,gemini-3.5-flash,,tomato/ep041
+,,,,Put the tomato in the bin,42,t003__ep042__tomato,True,85,496,50.2,6,,,True,False,gemini-3.5-flash,,tomato/ep042
+,,,,Put the tomato in the bin,43,t003__ep043__tomato,False,73,432,51.1,6,,,True,True,gemini-3.5-flash,319,tomato/ep043
+,,,,Put the tomato in the bin,44,t003__ep044__tomato,False,70,403,50.0,6,,,True,True,gemini-3.5-flash,223,tomato/ep044
+,,,,Put the tomato in the bin,45,t003__ep045__tomato,False,92,532,50.5,6,,,True,True,gemini-3.5-flash,235,tomato/ep045
+,,,,Put the tomato in the bin,46,t003__ep046__tomato,False,106,606,50.1,6,,,True,True,gemini-3.5-flash,260,tomato/ep046
+,,,,Put the tomato in the bin,47,t003__ep047__tomato,True,108,632,50.2,6,,,True,False,gemini-3.5-flash,,tomato/ep047
+,,,,Put the tomato in the bin,48,t003__ep048__tomato,False,67,376,50.1,6,,,True,True,gemini-3.5-flash,85,tomato/ep048
+,,,,Put the tomato in the bin,49,t003__ep049__tomato,False,97,562,50.4,6,,,True,True,gemini-3.5-flash,184,tomato/ep049
diff --git a/mushroom/ep000/example.md b/mushroom/ep000/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..f6f652ae95f1d59faea9ae2213cfe022a82bcdde
--- /dev/null
+++ b/mushroom/ep000/example.md
@@ -0,0 +1,34 @@
+# t001__ep000__mushroom
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the mushroom on the plate
+- **success:** False
+- **control steps:** 508  (10.1s at 50 Hz)
+- **policy inferences:** 88  (replan_steps=6)
+- **video:** `mushroom/ep000/rollout.mp4`
+- **features:** `88` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–149 | 0–24 | reaching for mushroom |
+| 3.0–6.0 | 150–299 | 25–49 | grasping and dropping mushroom |
+| 6.0–9.0 | 300–449 | 50–74 | moving empty gripper over plate |
+| 9.0–10.1 | 450–507 | 75–84 | hovering empty gripper above table |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 5.508982035928144 s (00:05.5)
+- **frame / control_step:** 276 (policy_step=46, chunk_index=0)
+- **coarse onset:** 5.5 s (refined=False)
+- **mode:** object_displacement
+- **reason:** The robot successfully grasped and lifted the mushroom, but dropped it onto the table while transporting it toward the green plate.
+  Summary: dropping mushroom during transport; unstable grasp
+- **recovery:** Reposition the arm over the dropped mushroom, execute a firm grasp, and transfer it onto the green plate.
+  Summary: re-grasp dropped mushroom; place on plate
+- **prevention:** Maintain adequate force and gripper alignment during transport to ensure the mushroom does not slip out before reaching the plate.
+  Summary: secure grasp on mushroom; carry smoothly to plate
+- **justification:** At 00:04-00:05, the robot picks up the mushroom off the table. However, as the arm moves left toward the green plate around 00:05-00:06, the mushroom slips out of the gripper and falls to the table, causing the rollout to fail.
diff --git a/mushroom/ep000/labels.json b/mushroom/ep000/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..3fbb668a9ee758ab46a0da1abb403990958b9267
--- /dev/null
+++ b/mushroom/ep000/labels.json
@@ -0,0 +1,109 @@
+{
+  "rollout_id": "t001__ep000__mushroom",
+  "scene_variant": "",
+  "task_id": 1,
+  "task_file": "",
+  "instruction": "Put the mushroom on the plate",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:53:08+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9940119760479043,
+      "control_step_start": 0,
+      "control_step_end": 149,
+      "policy_step_start": 0,
+      "policy_step_end": 24,
+      "phrase": "reaching for mushroom",
+      "description": "reaching for mushroom"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9940119760479043,
+      "t_end_sec": 5.9880239520958085,
+      "control_step_start": 150,
+      "control_step_end": 299,
+      "policy_step_start": 25,
+      "policy_step_end": 49,
+      "phrase": "grasping and dropping mushroom",
+      "description": "grasping and dropping mushroom"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.9880239520958085,
+      "t_end_sec": 8.982035928143713,
+      "control_step_start": 300,
+      "control_step_end": 449,
+      "policy_step_start": 50,
+      "policy_step_end": 74,
+      "phrase": "moving empty gripper over plate",
+      "description": "moving empty gripper over plate"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.982035928143713,
+      "t_end_sec": 10.139720558882235,
+      "control_step_start": 450,
+      "control_step_end": 507,
+      "policy_step_start": 75,
+      "policy_step_end": 84,
+      "phrase": "hovering empty gripper above table",
+      "description": "hovering empty gripper above table"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "object_displacement",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 5.508982035928144,
+    "vlm_failure_onset_timestamp": "00:05.5",
+    "vlm_failure_onset_frame": 276,
+    "vlm_failure_onset_step": 276,
+    "vlm_coarse_failure_onset_seconds": 5.5,
+    "vlm_failure_window_start_seconds": 5.0,
+    "vlm_failure_window_end_seconds": 6.0,
+    "vlm_temporal_refined": false,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot successfully grasped and lifted the mushroom, but dropped it onto the table while transporting it toward the green plate.",
+    "vlm_failure_reason_summary": "dropping mushroom during transport; unstable grasp",
+    "vlm_recovery_action": "Reposition the arm over the dropped mushroom, execute a firm grasp, and transfer it onto the green plate.",
+    "vlm_recovery_action_summary": "re-grasp dropped mushroom; place on plate",
+    "vlm_prevention_action": "Maintain adequate force and gripper alignment during transport to ensure the mushroom does not slip out before reaching the plate.",
+    "vlm_prevention_action_summary": "secure grasp on mushroom; carry smoothly to plate",
+    "vlm_justification": "At 00:04-00:05, the robot picks up the mushroom off the table. However, as the arm moves left toward the green plate around 00:05-00:06, the mushroom slips out of the gripper and falls to the table, causing the rollout to fail.",
+    "vlm_temporal_justification": "At timestamp 04:31, the robot arm lowers the mushroom and drops/places it directly onto the table surface instead of transporting it onto the green plate.",
+    "vlm_video_fps": 50.1,
+    "vlm_video_num_frames": 508,
+    "vlm_video_duration_seconds": 10.119760479041917,
+    "vlm_usage": {
+      "input_tokens": 21446,
+      "output_tokens": 368,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"object_displacement\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 5.5,\n \"failure_window_start_seconds\": 5.0,\n \"failure_window_end_seconds\": 6.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot successfully grasped and lifted the mushroom, but dropped it onto the table while transporting it toward the green plate.\",\n \"failure_reason_summary\": \"dropping mushroom during transport; unstable grasp\",\n \"recovery_action\": \"Reposition the arm over the dropped mushroom, execute a firm grasp, and transfer it onto the green plate.\",\n \"recovery_action_summary\": \"re-grasp dropped mushroom; place on plate\",\n \"prevention_action\": \"Maintain adequate force and gripper alignment during transport to ensure the mushroom does not slip out before reaching the plate.\",\n \"prevention_action_summary\": \"secure grasp on mushroom; carry smoothly to plate\",\n \"justification\": \"At 00:04-00:05, the robot picks up the mushroom off the table. However, as the arm moves left toward the green plate around 00:05-00:06, the mushroom slips out of the gripper and falls to the table, causing the rollout to fail.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 276,
+    "failure_sim_step": 276,
+    "failure_policy_step": 46,
+    "failure_chunk_index": 0,
+    "first_post_failure_policy_step": 47,
+    "failure_type": "object_displacement",
+    "correction_action": "Reposition the arm over the dropped mushroom, execute a firm grasp, and transfer it onto the green plate."
+  }
+}
\ No newline at end of file
diff --git a/mushroom/ep000/labels.npz b/mushroom/ep000/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..0f20d4adfa1b8bbba249f8965c31c7776e2bd821
--- /dev/null
+++ b/mushroom/ep000/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a231470c6c1d1eb760081f7c0e8034e4c293b8a640b74b3a3c3c199f9e95a858
+size 3814
diff --git a/mushroom/ep001/example.md b/mushroom/ep001/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..d73e6892547708b89e07cbca00b3e1776a826c36
--- /dev/null
+++ b/mushroom/ep001/example.md
@@ -0,0 +1,34 @@
+# t001__ep001__mushroom
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the mushroom on the plate
+- **success:** False
+- **control steps:** 464  (9.2s at 50 Hz)
+- **policy inferences:** 80  (replan_steps=6)
+- **video:** `mushroom/ep001/rollout.mp4`
+- **features:** `80` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | lowering gripper toward mushroom |
+| 3.0–6.0 | 151–301 | 25–50 | attempting grasp and lifting empty |
+| 6.0–9.0 | 302–452 | 50–75 | hovering above table |
+| 9.0–9.2 | 453–463 | 75–77 | remaining stationary |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 5.674603174603175 s (00:05.7)
+- **frame / control_step:** 286 (policy_step=47, chunk_index=4)
+- **coarse onset:** 5.0 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the mushroom, but its gripper failed to secure a firm hold around the stem, causing the gripper to release and pull away empty.
+  Summary: failed grasp on mushroom; gripper slipped off
+- **recovery:** Re-align the gripper directly over the mushroom, descend to securely clamp around its stem, lift it, and transport it onto the green plate.
+  Summary: re-align gripper; grasp mushroom securely; place on plate
+- **prevention:** Lower the end-effector closer to the table and align the gripper fingers precisely around the stem before closing to ensure a secure grip.
+  Summary: center gripper over stem; close firmly before lifting
+- **justification:** At around 5.0 seconds, the robot closes its gripper on the mushroom but fails to secure it, lifting up empty while the mushroom remains on the tabletop.
diff --git a/mushroom/ep001/labels.json b/mushroom/ep001/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..f9ab50e8fe5820e761235e3cb60e5f4099af8d78
--- /dev/null
+++ b/mushroom/ep001/labels.json
@@ -0,0 +1,109 @@
+{
+  "rollout_id": "t001__ep001__mushroom",
+  "scene_variant": "",
+  "task_id": 1,
+  "task_file": "",
+  "instruction": "Put the mushroom on the plate",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:53:41+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.996031746031746,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "lowering gripper toward mushroom",
+      "description": "lowering gripper toward mushroom"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.996031746031746,
+      "t_end_sec": 5.992063492063492,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "attempting grasp and lifting empty",
+      "description": "attempting grasp and lifting empty"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.992063492063492,
+      "t_end_sec": 8.988095238095239,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "hovering above table",
+      "description": "hovering above table"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.988095238095239,
+      "t_end_sec": 9.206349206349207,
+      "control_step_start": 453,
+      "control_step_end": 463,
+      "policy_step_start": 75,
+      "policy_step_end": 77,
+      "phrase": "remaining stationary",
+      "description": "remaining stationary"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 5.674603174603175,
+    "vlm_failure_onset_timestamp": "00:05.7",
+    "vlm_failure_onset_frame": 286,
+    "vlm_failure_onset_step": 286,
+    "vlm_coarse_failure_onset_seconds": 5.0,
+    "vlm_failure_window_start_seconds": 4.5,
+    "vlm_failure_window_end_seconds": 5.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the mushroom, but its gripper failed to secure a firm hold around the stem, causing the gripper to release and pull away empty.",
+    "vlm_failure_reason_summary": "failed grasp on mushroom; gripper slipped off",
+    "vlm_recovery_action": "Re-align the gripper directly over the mushroom, descend to securely clamp around its stem, lift it, and transport it onto the green plate.",
+    "vlm_recovery_action_summary": "re-align gripper; grasp mushroom securely; place on plate",
+    "vlm_prevention_action": "Lower the end-effector closer to the table and align the gripper fingers precisely around the stem before closing to ensure a secure grip.",
+    "vlm_prevention_action_summary": "center gripper over stem; close firmly before lifting",
+    "vlm_justification": "At around 5.0 seconds, the robot closes its gripper on the mushroom but fails to secure it, lifting up empty while the mushroom remains on the tabletop.",
+    "vlm_temporal_justification": "At timestamp 03:05 (frame 185), the robot's gripper fingers begin opening and releasing the mushroom stem rather than lifting it, making the grasp failure visibly apparent.",
+    "vlm_video_fps": 50.4,
+    "vlm_video_num_frames": 464,
+    "vlm_video_duration_seconds": 9.186507936507937,
+    "vlm_usage": {
+      "input_tokens": 21515,
+      "output_tokens": 361,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 5.0,\n \"failure_window_start_seconds\": 4.5,\n \"failure_window_end_seconds\": 5.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the mushroom, but its gripper failed to secure a firm hold around the stem, causing the gripper to release and pull away empty.\",\n \"failure_reason_summary\": \"failed grasp on mushroom; gripper slipped off\",\n \"recovery_action\": \"Re-align the gripper directly over the mushroom, descend to securely clamp around its stem, lift it, and transport it onto the green plate.\",\n \"recovery_action_summary\": \"re-align gripper; grasp mushroom securely; place on plate\",\n \"prevention_action\": \"Lower the end-effector closer to the table and align the gripper fingers precisely around the stem before closing to ensure a secure grip.\",\n \"prevention_action_summary\": \"center gripper over stem; close firmly before lifting\",\n \"justification\": \"At around 5.0 seconds, the robot closes its gripper on the mushroom but fails to secure it, lifting up empty while the mushroom remains on the tabletop.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 286,
+    "failure_sim_step": 286,
+    "failure_policy_step": 47,
+    "failure_chunk_index": 4,
+    "first_post_failure_policy_step": 48,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-align the gripper directly over the mushroom, descend to securely clamp around its stem, lift it, and transport it onto the green plate."
+  }
+}
\ No newline at end of file
diff --git a/mushroom/ep001/labels.npz b/mushroom/ep001/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..d3d489b434740d09e81e046820cb20afea70c3bf
--- /dev/null
+++ b/mushroom/ep001/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2eb9de611e8d307b71a89c45b65b3bfe2f35d9cedfffe8c55f1cffe4861aa8d0
+size 3873
diff --git a/mushroom/ep002/example.md b/mushroom/ep002/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..b696a8b12a3e8bf0844f9f04a99085fa41e2bd57
--- /dev/null
+++ b/mushroom/ep002/example.md
@@ -0,0 +1,34 @@
+# t001__ep002__mushroom
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the mushroom on the plate
+- **success:** False
+- **control steps:** 519  (10.3s at 50 Hz)
+- **policy inferences:** 90  (replan_steps=6)
+- **video:** `mushroom/ep002/rollout.mp4`
+- **features:** `90` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | reaching for mushroom |
+| 3.0–6.0 | 151–301 | 25–50 | attempting grasp; lifting empty |
+| 6.0–9.0 | 302–452 | 50–75 | moving towards green plate |
+| 9.0–10.3 | 453–518 | 75–86 | hovering empty above plate |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 4.00398406374502 s (00:04.0)
+- **frame / control_step:** 201 (policy_step=33, chunk_index=3)
+- **coarse onset:** 4.0 s (refined=False)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the mushroom on the table but failed to secure a firm grip, lifting up empty and proceeding toward the green plate without holding the mushroom.
+  Summary: failed grasp attempt; moving empty gripper to target
+- **recovery:** Re-orient and center the end-effector over the mushroom, execute a secure grasp, and transport it onto the green plate.
+  Summary: re-center gripper; regrasp mushroom; move to plate
+- **prevention:** Align the gripper carefully over the center of mass of the mushroom and lower it sufficiently prior to closing the fingers to ensure a firm grasp.
+  Summary: center gripper over mushroom; lower fully before closing
+- **justification:** At around 4.0 seconds, the robot closes its gripper on the mushroom but fails to secure it, lifting up empty by 5.0 seconds and continuing to execute the placement trajectory without the object.
diff --git a/mushroom/ep002/labels.json b/mushroom/ep002/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..f1bb5eb4037ceac5d281b7919c7e49156b8258be
--- /dev/null
+++ b/mushroom/ep002/labels.json
@@ -0,0 +1,109 @@
+{
+  "rollout_id": "t001__ep002__mushroom",
+  "scene_variant": "",
+  "task_id": 1,
+  "task_file": "",
+  "instruction": "Put the mushroom on the plate",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:54:11+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00796812749004,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for mushroom",
+      "description": "reaching for mushroom"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00796812749004,
+      "t_end_sec": 6.01593625498008,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "attempting grasp; lifting empty",
+      "description": "attempting grasp; lifting empty"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01593625498008,
+      "t_end_sec": 9.02390438247012,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "moving towards green plate",
+      "description": "moving towards green plate"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.02390438247012,
+      "t_end_sec": 10.338645418326692,
+      "control_step_start": 453,
+      "control_step_end": 518,
+      "policy_step_start": 75,
+      "policy_step_end": 86,
+      "phrase": "hovering empty above plate",
+      "description": "hovering empty above plate"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 4.00398406374502,
+    "vlm_failure_onset_timestamp": "00:04.0",
+    "vlm_failure_onset_frame": 201,
+    "vlm_failure_onset_step": 201,
+    "vlm_coarse_failure_onset_seconds": 4.0,
+    "vlm_failure_window_start_seconds": 3.5,
+    "vlm_failure_window_end_seconds": 5.0,
+    "vlm_temporal_refined": false,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the mushroom on the table but failed to secure a firm grip, lifting up empty and proceeding toward the green plate without holding the mushroom.",
+    "vlm_failure_reason_summary": "failed grasp attempt; moving empty gripper to target",
+    "vlm_recovery_action": "Re-orient and center the end-effector over the mushroom, execute a secure grasp, and transport it onto the green plate.",
+    "vlm_recovery_action_summary": "re-center gripper; regrasp mushroom; move to plate",
+    "vlm_prevention_action": "Align the gripper carefully over the center of mass of the mushroom and lower it sufficiently prior to closing the fingers to ensure a firm grasp.",
+    "vlm_prevention_action_summary": "center gripper over mushroom; lower fully before closing",
+    "vlm_justification": "At around 4.0 seconds, the robot closes its gripper on the mushroom but fails to secure it, lifting up empty by 5.0 seconds and continuing to execute the placement trajectory without the object.",
+    "vlm_temporal_justification": "At 03:35, the robot's gripper fingers fully close together without enclosing the mushroom, leaving the object on the table and making the grasp failure visibly apparent.",
+    "vlm_video_fps": 50.2,
+    "vlm_video_num_frames": 519,
+    "vlm_video_duration_seconds": 10.318725099601593,
+    "vlm_usage": {
+      "input_tokens": 21578,
+      "output_tokens": 368,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 4.0,\n \"failure_window_start_seconds\": 3.5,\n \"failure_window_end_seconds\": 5.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the mushroom on the table but failed to secure a firm grip, lifting up empty and proceeding toward the green plate without holding the mushroom.\",\n \"failure_reason_summary\": \"failed grasp attempt; moving empty gripper to target\",\n \"recovery_action\": \"Re-orient and center the end-effector over the mushroom, execute a secure grasp, and transport it onto the green plate.\",\n \"recovery_action_summary\": \"re-center gripper; regrasp mushroom; move to plate\",\n \"prevention_action\": \"Align the gripper carefully over the center of mass of the mushroom and lower it sufficiently prior to closing the fingers to ensure a firm grasp.\",\n \"prevention_action_summary\": \"center gripper over mushroom; lower fully before closing\",\n \"justification\": \"At around 4.0 seconds, the robot closes its gripper on the mushroom but fails to secure it, lifting up empty by 5.0 seconds and continuing to execute the placement trajectory without the object.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 201,
+    "failure_sim_step": 201,
+    "failure_policy_step": 33,
+    "failure_chunk_index": 3,
+    "first_post_failure_policy_step": 34,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-orient and center the end-effector over the mushroom, execute a secure grasp, and transport it onto the green plate."
+  }
+}
\ No newline at end of file
diff --git a/mushroom/ep002/labels.npz b/mushroom/ep002/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..2a305e25480642305660f788a614e38e90eb47a8
--- /dev/null
+++ b/mushroom/ep002/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:83d3723a695fe06c9ffbb316705e2cd648bbf61cf13b54fca6b94f6df0378301
+size 3886
diff --git a/mushroom/ep003/example.md b/mushroom/ep003/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..9e587e1f89b84f5e876815c640f066829b4ccf55
--- /dev/null
+++ b/mushroom/ep003/example.md
@@ -0,0 +1,33 @@
+# t001__ep003__mushroom
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the mushroom on the plate
+- **success:** False
+- **control steps:** 423  (8.4s at 50 Hz)
+- **policy inferences:** 73  (replan_steps=6)
+- **video:** `mushroom/ep003/rollout.mp4`
+- **features:** `73` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching down for mushroom |
+| 3.0–6.0 | 152–303 | 25–50 | grasping and moving mushroom |
+| 6.0–8.4 | 304–422 | 50–70 | releasing mushroom; raising arm |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 6.851485148514851 s (00:06.9)
+- **frame / control_step:** 346 (policy_step=57, chunk_index=4)
+- **coarse onset:** 5.8 s (refined=True)
+- **mode:** placement_or_insertion_failure
+- **reason:** The robot brought the mushroom toward the green plate but released it prematurely near the front edge of the plate, causing the mushroom to fall onto the table surface.
+  Summary: premature release; misaligned placement on plate edge
+- **recovery:** Grasp the mushroom again from the table surface in front of the plate and place it securely inside the plate.
+  Summary: regrasp mushroom; place centered on plate
+- **prevention:** Carry the mushroom fully over the center of the green plate before opening the gripper to ensure a stable placement.
+  Summary: center mushroom over plate; release object
+- **justification:** The robot successfully grasps the mushroom around 4.0s and carries it toward the green plate, but releases its grip prematurely at 5.8s right over the edge of the plate, resulting in the mushroom falling off onto the table.
diff --git a/mushroom/ep003/labels.json b/mushroom/ep003/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..b7f4718921f6ecbfc2ad3735a0a27d91df6db58b
--- /dev/null
+++ b/mushroom/ep003/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t001__ep003__mushroom",
+  "scene_variant": "",
+  "task_id": 1,
+  "task_file": "",
+  "instruction": "Put the mushroom on the plate",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:54:42+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00990099009901,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching down for mushroom",
+      "description": "reaching down for mushroom"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00990099009901,
+      "t_end_sec": 6.01980198019802,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "grasping and moving mushroom",
+      "description": "grasping and moving mushroom"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01980198019802,
+      "t_end_sec": 8.376237623762377,
+      "control_step_start": 304,
+      "control_step_end": 422,
+      "policy_step_start": 50,
+      "policy_step_end": 70,
+      "phrase": "releasing mushroom; raising arm",
+      "description": "releasing mushroom; raising arm"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "placement_or_insertion_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 6.851485148514851,
+    "vlm_failure_onset_timestamp": "00:06.9",
+    "vlm_failure_onset_frame": 346,
+    "vlm_failure_onset_step": 346,
+    "vlm_coarse_failure_onset_seconds": 5.8,
+    "vlm_failure_window_start_seconds": 5.3,
+    "vlm_failure_window_end_seconds": 6.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot brought the mushroom toward the green plate but released it prematurely near the front edge of the plate, causing the mushroom to fall onto the table surface.",
+    "vlm_failure_reason_summary": "premature release; misaligned placement on plate edge",
+    "vlm_recovery_action": "Grasp the mushroom again from the table surface in front of the plate and place it securely inside the plate.",
+    "vlm_recovery_action_summary": "regrasp mushroom; place centered on plate",
+    "vlm_prevention_action": "Carry the mushroom fully over the center of the green plate before opening the gripper to ensure a stable placement.",
+    "vlm_prevention_action_summary": "center mushroom over plate; release object",
+    "vlm_justification": "The robot successfully grasps the mushroom around 4.0s and carries it toward the green plate, but releases its grip prematurely at 5.8s right over the edge of the plate, resulting in the mushroom falling off onto the table.",
+    "vlm_temporal_justification": "At timestamp 03:25 (second 205 of the clip), the robot gripper begins opening to release the mushroom prematurely before it is positioned over the green plate, causing the mushroom to fall onto the table surface.",
+    "vlm_video_fps": 50.5,
+    "vlm_video_num_frames": 423,
+    "vlm_video_duration_seconds": 8.356435643564357,
+    "vlm_usage": {
+      "input_tokens": 20133,
+      "output_tokens": 373,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"placement_or_insertion_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 5.8,\n \"failure_window_start_seconds\": 5.3,\n \"failure_window_end_seconds\": 6.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot brought the mushroom toward the green plate but released it prematurely near the front edge of the plate, causing the mushroom to fall onto the table surface.\",\n \"failure_reason_summary\": \"premature release; misaligned placement on plate edge\",\n \"recovery_action\": \"Grasp the mushroom again from the table surface in front of the plate and place it securely inside the plate.\",\n \"recovery_action_summary\": \"regrasp mushroom; place centered on plate\",\n \"prevention_action\": \"Carry the mushroom fully over the center of the green plate before opening the gripper to ensure a stable placement.\",\n \"prevention_action_summary\": \"center mushroom over plate; release object\",\n \"justification\": \"The robot successfully grasps the mushroom around 4.0s and carries it toward the green plate, but releases its grip prematurely at 5.8s right over the edge of the plate, resulting in the mushroom falling off onto the table.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 346,
+    "failure_sim_step": 346,
+    "failure_policy_step": 57,
+    "failure_chunk_index": 4,
+    "first_post_failure_policy_step": 58,
+    "failure_type": "placement_or_insertion_failure",
+    "correction_action": "Grasp the mushroom again from the table surface in front of the plate and place it securely inside the plate."
+  }
+}
\ No newline at end of file
diff --git a/mushroom/ep003/labels.npz b/mushroom/ep003/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..785255923b2d53b0b60174898d5adbc11dc5de7a
--- /dev/null
+++ b/mushroom/ep003/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e3c1d0991b3df8959997486782f528ba06a39e423537236b1b1c8a574eff318d
+size 3761
diff --git a/mushroom/ep004/example.md b/mushroom/ep004/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..22ac90c4f6009c943ea65b56767d700fb07a2c27
--- /dev/null
+++ b/mushroom/ep004/example.md
@@ -0,0 +1,34 @@
+# t001__ep004__mushroom
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the mushroom on the plate
+- **success:** False
+- **control steps:** 475  (9.4s at 51 Hz)
+- **policy inferences:** 82  (replan_steps=6)
+- **video:** `mushroom/ep004/rollout.mp4`
+- **features:** `82` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | approaching mushroom |
+| 3.0–6.0 | 152–303 | 25–50 | closing gripper near mushroom |
+| 6.0–9.0 | 304–455 | 50–75 | lifting empty gripper |
+| 9.0–9.4 | 456–474 | 76–79 | hovering near plate |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 4.150197628458498 s (00:04.2)
+- **frame / control_step:** 210 (policy_step=35, chunk_index=0)
+- **coarse onset:** 4.5 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to pick up the mushroom, but the gripper closed prematurely and missed the object, lifting up empty.
+  Summary: failed grasp on mushroom; empty gripper closure
+- **recovery:** Re-open the gripper, descend directly over the mushroom to establish a proper grip, and then lift it onto the green plate.
+  Summary: re-align gripper over mushroom; secure grasp
+- **prevention:** Align the end-effector precisely over the mushroom's center of mass before closing the gripper fingers.
+  Summary: center gripper precisely; secure grasp
+- **justification:** Around 4.5 seconds, the robot closes its gripper without successfully grasping the mushroom. It then lifts away empty, leaving the mushroom behind on the table.
diff --git a/mushroom/ep004/labels.json b/mushroom/ep004/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..3d9a0b2481bdbd1404595fb63ee1277e45cdb7d4
--- /dev/null
+++ b/mushroom/ep004/labels.json
@@ -0,0 +1,109 @@
+{
+  "rollout_id": "t001__ep004__mushroom",
+  "scene_variant": "",
+  "task_id": 1,
+  "task_file": "",
+  "instruction": "Put the mushroom on the plate",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:55:13+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0039525691699605,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "approaching mushroom",
+      "description": "approaching mushroom"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0039525691699605,
+      "t_end_sec": 6.007905138339921,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "closing gripper near mushroom",
+      "description": "closing gripper near mushroom"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.007905138339921,
+      "t_end_sec": 9.011857707509881,
+      "control_step_start": 304,
+      "control_step_end": 455,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "lifting empty gripper",
+      "description": "lifting empty gripper"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.011857707509881,
+      "t_end_sec": 9.387351778656127,
+      "control_step_start": 456,
+      "control_step_end": 474,
+      "policy_step_start": 76,
+      "policy_step_end": 79,
+      "phrase": "hovering near plate",
+      "description": "hovering near plate"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 4.150197628458498,
+    "vlm_failure_onset_timestamp": "00:04.2",
+    "vlm_failure_onset_frame": 210,
+    "vlm_failure_onset_step": 210,
+    "vlm_coarse_failure_onset_seconds": 4.5,
+    "vlm_failure_window_start_seconds": 4.0,
+    "vlm_failure_window_end_seconds": 5.0,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to pick up the mushroom, but the gripper closed prematurely and missed the object, lifting up empty.",
+    "vlm_failure_reason_summary": "failed grasp on mushroom; empty gripper closure",
+    "vlm_recovery_action": "Re-open the gripper, descend directly over the mushroom to establish a proper grip, and then lift it onto the green plate.",
+    "vlm_recovery_action_summary": "re-align gripper over mushroom; secure grasp",
+    "vlm_prevention_action": "Align the end-effector precisely over the mushroom's center of mass before closing the gripper fingers.",
+    "vlm_prevention_action_summary": "center gripper precisely; secure grasp",
+    "vlm_justification": "Around 4.5 seconds, the robot closes its gripper without successfully grasping the mushroom. It then lifts away empty, leaving the mushroom behind on the table.",
+    "vlm_temporal_justification": "At clip timestamp 02:14 (second 134), the gripper begins closing prematurely above the mushroom, missing the object entirely.",
+    "vlm_video_fps": 50.6,
+    "vlm_video_num_frames": 475,
+    "vlm_video_duration_seconds": 9.367588932806324,
+    "vlm_usage": {
+      "input_tokens": 21630,
+      "output_tokens": 327,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 4.5,\n \"failure_window_start_seconds\": 4.0,\n \"failure_window_end_seconds\": 5.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to pick up the mushroom, but the gripper closed prematurely and missed the object, lifting up empty.\",\n \"failure_reason_summary\": \"failed grasp on mushroom; empty gripper closure\",\n \"recovery_action\": \"Re-open the gripper, descend directly over the mushroom to establish a proper grip, and then lift it onto the green plate.\",\n \"recovery_action_summary\": \"re-align gripper over mushroom; secure grasp\",\n \"prevention_action\": \"Align the end-effector precisely over the mushroom's center of mass before closing the gripper fingers.\",\n \"prevention_action_summary\": \"center gripper precisely; secure grasp\",\n \"justification\": \"Around 4.5 seconds, the robot closes its gripper without successfully grasping the mushroom. It then lifts away empty, leaving the mushroom behind on the table.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 210,
+    "failure_sim_step": 210,
+    "failure_policy_step": 35,
+    "failure_chunk_index": 0,
+    "first_post_failure_policy_step": 36,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-open the gripper, descend directly over the mushroom to establish a proper grip, and then lift it onto the green plate."
+  }
+}
\ No newline at end of file
diff --git a/mushroom/ep004/labels.npz b/mushroom/ep004/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..70841149de9b596376c0ff8b383075da26e50f52
--- /dev/null
+++ b/mushroom/ep004/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c137c63b14cad2c6801f16cae32fa3e35e4d494f1e75e734f02efc02051b67c9
+size 3758
diff --git a/mushroom/ep005/example.md b/mushroom/ep005/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..2e4105f134ce8a2c61774c3787d59f7460c6fa66
--- /dev/null
+++ b/mushroom/ep005/example.md
@@ -0,0 +1,33 @@
+# t001__ep005__mushroom
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the mushroom on the plate
+- **success:** False
+- **control steps:** 386  (7.6s at 50 Hz)
+- **policy inferences:** 67  (replan_steps=6)
+- **video:** `mushroom/ep005/rollout.mp4`
+- **features:** `67` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching down toward mushroom |
+| 3.0–6.0 | 152–303 | 25–50 | closing gripper and moving toward plate |
+| 6.0–7.6 | 304–385 | 50–64 | hovering empty gripper above plate |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 3.504950495049505 s (00:03.5)
+- **frame / control_step:** 177 (policy_step=29, chunk_index=3)
+- **coarse onset:** 3.5 s (refined=False)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the mushroom, but closed its gripper without securing a firm hold, moving away toward the green plate while leaving the mushroom behind on the table.
+  Summary: failed grasp attempt; moving empty gripper to target
+- **recovery:** Return the end effector to the mushroom, align the gripper properly over its stem to establish a secure grasp, and then transfer it onto the green plate.
+  Summary: re-align over mushroom; secure grasp; place on plate
+- **prevention:** Ensure the gripper descends sufficiently and centers around the mushroom before closing fingers to secure a firm grasp before moving.
+  Summary: center gripper over mushroom; close securely; move to plate
+- **justification:** Around 3.5 seconds into the rollout, the robot closes its gripper on the mushroom but fails to hold it. The arm then moves away empty-handed to the green plate, leaving the mushroom behind.
diff --git a/mushroom/ep005/labels.json b/mushroom/ep005/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..158ed7b14bba7df4439c133f342d2c2d75e47c6f
--- /dev/null
+++ b/mushroom/ep005/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t001__ep005__mushroom",
+  "scene_variant": "",
+  "task_id": 1,
+  "task_file": "",
+  "instruction": "Put the mushroom on the plate",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:55:43+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00990099009901,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching down toward mushroom",
+      "description": "reaching down toward mushroom"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00990099009901,
+      "t_end_sec": 6.01980198019802,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "closing gripper and moving toward plate",
+      "description": "closing gripper and moving toward plate"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01980198019802,
+      "t_end_sec": 7.643564356435643,
+      "control_step_start": 304,
+      "control_step_end": 385,
+      "policy_step_start": 50,
+      "policy_step_end": 64,
+      "phrase": "hovering empty gripper above plate",
+      "description": "hovering empty gripper above plate"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 3.504950495049505,
+    "vlm_failure_onset_timestamp": "00:03.5",
+    "vlm_failure_onset_frame": 177,
+    "vlm_failure_onset_step": 177,
+    "vlm_coarse_failure_onset_seconds": 3.5,
+    "vlm_failure_window_start_seconds": 3.0,
+    "vlm_failure_window_end_seconds": 4.5,
+    "vlm_temporal_refined": false,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the mushroom, but closed its gripper without securing a firm hold, moving away toward the green plate while leaving the mushroom behind on the table.",
+    "vlm_failure_reason_summary": "failed grasp attempt; moving empty gripper to target",
+    "vlm_recovery_action": "Return the end effector to the mushroom, align the gripper properly over its stem to establish a secure grasp, and then transfer it onto the green plate.",
+    "vlm_recovery_action_summary": "re-align over mushroom; secure grasp; place on plate",
+    "vlm_prevention_action": "Ensure the gripper descends sufficiently and centers around the mushroom before closing fingers to secure a firm grasp before moving.",
+    "vlm_prevention_action_summary": "center gripper over mushroom; close securely; move to plate",
+    "vlm_justification": "Around 3.5 seconds into the rollout, the robot closes its gripper on the mushroom but fails to hold it. The arm then moves away empty-handed to the green plate, leaving the mushroom behind.",
+    "vlm_temporal_justification": "At 03:14, the robot arm begins moving upward after fully closing its gripper, leaving the mushroom behind on its stand and confirming that the grasp attempt failed.",
+    "vlm_video_fps": 50.5,
+    "vlm_video_num_frames": 386,
+    "vlm_video_duration_seconds": 7.623762376237623,
+    "vlm_usage": {
+      "input_tokens": 21577,
+      "output_tokens": 368,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 3.5,\n \"failure_window_start_seconds\": 3.0,\n \"failure_window_end_seconds\": 4.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the mushroom, but closed its gripper without securing a firm hold, moving away toward the green plate while leaving the mushroom behind on the table.\",\n \"failure_reason_summary\": \"failed grasp attempt; moving empty gripper to target\",\n \"recovery_action\": \"Return the end effector to the mushroom, align the gripper properly over its stem to establish a secure grasp, and then transfer it onto the green plate.\",\n \"recovery_action_summary\": \"re-align over mushroom; secure grasp; place on plate\",\n \"prevention_action\": \"Ensure the gripper descends sufficiently and centers around the mushroom before closing fingers to secure a firm grasp before moving.\",\n \"prevention_action_summary\": \"center gripper over mushroom; close securely; move to plate\",\n \"justification\": \"Around 3.5 seconds into the rollout, the robot closes its gripper on the mushroom but fails to hold it. The arm then moves away empty-handed to the green plate, leaving the mushroom behind.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 177,
+    "failure_sim_step": 177,
+    "failure_policy_step": 29,
+    "failure_chunk_index": 3,
+    "first_post_failure_policy_step": 30,
+    "failure_type": "grasp_failure",
+    "correction_action": "Return the end effector to the mushroom, align the gripper properly over its stem to establish a secure grasp, and then transfer it onto the green plate."
+  }
+}
\ No newline at end of file
diff --git a/mushroom/ep005/labels.npz b/mushroom/ep005/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..b083108433d590231101912fab25a1ff9d909e54
--- /dev/null
+++ b/mushroom/ep005/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6c26c05f8a30a56663df4eb9a56e4414cb8097ce5c03b871988818e9d6a5eb2b
+size 3876
diff --git a/mushroom/ep006/example.md b/mushroom/ep006/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..0416b9f77720b9734ed5f1298ac9cfb846458f37
--- /dev/null
+++ b/mushroom/ep006/example.md
@@ -0,0 +1,21 @@
+# t001__ep006__mushroom
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the mushroom on the plate
+- **success:** True
+- **control steps:** 392  (7.8s at 50 Hz)
+- **policy inferences:** 68  (replan_steps=6)
+- **video:** `mushroom/ep006/rollout.mp4`
+- **features:** `68` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | reaching and grasping mushroom |
+| 3.0–6.0 | 151–301 | 25–50 | placing mushroom on plate |
+| 6.0–7.8 | 302–391 | 50–65 | retracting arm |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/mushroom/ep006/labels.json b/mushroom/ep006/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..0126b1c3719901c5e2a320452514ecf9f49fade2
--- /dev/null
+++ b/mushroom/ep006/labels.json
@@ -0,0 +1,68 @@
+{
+  "rollout_id": "t001__ep006__mushroom",
+  "scene_variant": "",
+  "task_id": 1,
+  "task_file": "",
+  "instruction": "Put the mushroom on the plate",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:55:52+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.996031746031746,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching and grasping mushroom",
+      "description": "reaching and grasping mushroom"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.996031746031746,
+      "t_end_sec": 5.992063492063492,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "placing mushroom on plate",
+      "description": "placing mushroom on plate"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.992063492063492,
+      "t_end_sec": 7.777777777777778,
+      "control_step_start": 302,
+      "control_step_end": 391,
+      "policy_step_start": 50,
+      "policy_step_end": 65,
+      "phrase": "retracting arm",
+      "description": "retracting arm"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/mushroom/ep006/labels.npz b/mushroom/ep006/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..223d85906c2277298cb07f12c91a3f9cc91264ee
--- /dev/null
+++ b/mushroom/ep006/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b945efd9bcb93fdb7b2f46e42c1810eaf1fc02bc3fc33a1f0b2f213990c57f34
+size 3069
diff --git a/mushroom/ep007/example.md b/mushroom/ep007/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..2e0e4e9ba714d7df840babc8454b766c7d3c972d
--- /dev/null
+++ b/mushroom/ep007/example.md
@@ -0,0 +1,33 @@
+# t001__ep007__mushroom
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the mushroom on the plate
+- **success:** False
+- **control steps:** 393  (7.8s at 50 Hz)
+- **policy inferences:** 69  (replan_steps=5)
+- **video:** `mushroom/ep007/rollout.mp4`
+- **features:** `69` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–30 | descending and attempting to grasp mushroom |
+| 3.0–6.0 | 151–301 | 30–60 | lifting empty gripper toward plate |
+| 6.0–7.8 | 302–392 | 60–78 | pressing against the green plate |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 3.797216699801193 s (00:03.8)
+- **frame / control_step:** 191 (policy_step=38, chunk_index=1)
+- **coarse onset:** 2.5 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the mushroom but failed to securely close its gripper around it, lifting its end-effector empty and subsequently moving toward the plate without the object.
+  Summary: unsuccessful grasp of mushroom; moving empty gripper to plate
+- **recovery:** Re-align the gripper over the mushroom on the table, close the fingers firmly around it, and then carry it to the green plate.
+  Summary: re-align over mushroom; grasp mushroom; place on plate
+- **prevention:** Accurately position the end-effector over the center of the mushroom and ensure a solid grip before initiating the lifting and transfer motion toward the plate.
+  Summary: align gripper over mushroom; close gripper firmly; transport to plate
+- **justification:** At around 2.5 seconds, the robot arm descends to pick up the mushroom, but fails to securely grasp it. It lifts its empty gripper at 3.0 seconds and moves over to collide with the plate without carrying the object.
diff --git a/mushroom/ep007/labels.json b/mushroom/ep007/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..93ae8d9f45f5eda2363e9ef9ecb82001721353a7
--- /dev/null
+++ b/mushroom/ep007/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t001__ep007__mushroom",
+  "scene_variant": "",
+  "task_id": 1,
+  "task_file": "",
+  "instruction": "Put the mushroom on the plate",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:56:34+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.001988071570577,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 30,
+      "phrase": "descending and attempting to grasp mushroom",
+      "description": "descending and attempting to grasp mushroom"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.001988071570577,
+      "t_end_sec": 6.003976143141154,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 30,
+      "policy_step_end": 60,
+      "phrase": "lifting empty gripper toward plate",
+      "description": "lifting empty gripper toward plate"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.003976143141154,
+      "t_end_sec": 7.813121272365805,
+      "control_step_start": 302,
+      "control_step_end": 392,
+      "policy_step_start": 60,
+      "policy_step_end": 78,
+      "phrase": "pressing against the green plate",
+      "description": "pressing against the green plate"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 3.797216699801193,
+    "vlm_failure_onset_timestamp": "00:03.8",
+    "vlm_failure_onset_frame": 191,
+    "vlm_failure_onset_step": 191,
+    "vlm_coarse_failure_onset_seconds": 2.5,
+    "vlm_failure_window_start_seconds": 2.0,
+    "vlm_failure_window_end_seconds": 3.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the mushroom but failed to securely close its gripper around it, lifting its end-effector empty and subsequently moving toward the plate without the object.",
+    "vlm_failure_reason_summary": "unsuccessful grasp of mushroom; moving empty gripper to plate",
+    "vlm_recovery_action": "Re-align the gripper over the mushroom on the table, close the fingers firmly around it, and then carry it to the green plate.",
+    "vlm_recovery_action_summary": "re-align over mushroom; grasp mushroom; place on plate",
+    "vlm_prevention_action": "Accurately position the end-effector over the center of the mushroom and ensure a solid grip before initiating the lifting and transfer motion toward the plate.",
+    "vlm_prevention_action_summary": "align gripper over mushroom; close gripper firmly; transport to plate",
+    "vlm_justification": "At around 2.5 seconds, the robot arm descends to pick up the mushroom, but fails to securely grasp it. It lifts its empty gripper at 3.0 seconds and moves over to collide with the plate without carrying the object.",
+    "vlm_temporal_justification": "The robot closes its gripper around the mushroom around 02:58, but fails to secure a grip. At 03:11 (second 191), the arm begins lifting upward empty, leaving the mushroom standing untouched on the table surface.",
+    "vlm_video_fps": 50.3,
+    "vlm_video_num_frames": 393,
+    "vlm_video_duration_seconds": 7.79324055666004,
+    "vlm_usage": {
+      "input_tokens": 19873,
+      "output_tokens": 402,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 2.5,\n \"failure_window_start_seconds\": 2.0,\n \"failure_window_end_seconds\": 3.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the mushroom but failed to securely close its gripper around it, lifting its end-effector empty and subsequently moving toward the plate without the object.\",\n \"failure_reason_summary\": \"unsuccessful grasp of mushroom; moving empty gripper to plate\",\n \"recovery_action\": \"Re-align the gripper over the mushroom on the table, close the fingers firmly around it, and then carry it to the green plate.\",\n \"recovery_action_summary\": \"re-align over mushroom; grasp mushroom; place on plate\",\n \"prevention_action\": \"Accurately position the end-effector over the center of the mushroom and ensure a solid grip before initiating the lifting and transfer motion toward the plate.\",\n \"prevention_action_summary\": \"align gripper over mushroom; close gripper firmly; transport to plate\",\n \"justification\": \"At around 2.5 seconds, the robot arm descends to pick up the mushroom, but fails to securely grasp it. It lifts its empty gripper at 3.0 seconds and moves over to collide with the plate without carrying the object.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 191,
+    "failure_sim_step": 191,
+    "failure_policy_step": 38,
+    "failure_chunk_index": 1,
+    "first_post_failure_policy_step": 39,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-align the gripper over the mushroom on the table, close the fingers firmly around it, and then carry it to the green plate."
+  }
+}
\ No newline at end of file
diff --git a/mushroom/ep007/labels.npz b/mushroom/ep007/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..99d533d08e865e9e8a2e9bb4c2fb38c14f9a85b1
--- /dev/null
+++ b/mushroom/ep007/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8d3250731b113a46d48a736dcdb408dfc9a4c751a448f92e6d6f17851b5efc2a
+size 3927
diff --git a/mushroom/ep008/example.md b/mushroom/ep008/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..e1e29c6f635fbe9df5b56abc98728dc8ab98069e
--- /dev/null
+++ b/mushroom/ep008/example.md
@@ -0,0 +1,33 @@
+# t001__ep008__mushroom
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the mushroom on the plate
+- **success:** False
+- **control steps:** 368  (7.3s at 50 Hz)
+- **policy inferences:** 64  (replan_steps=6)
+- **video:** `mushroom/ep008/rollout.mp4`
+- **features:** `64` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | descending and grasping mushroom |
+| 3.0–6.0 | 151–301 | 25–50 | carrying and dropping mushroom |
+| 6.0–7.3 | 302–367 | 50–61 | hovering empty gripper above plate |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 2.549800796812749 s (00:02.5)
+- **frame / control_step:** 128 (policy_step=21, chunk_index=2)
+- **coarse onset:** 5.5 s (refined=True)
+- **mode:** object_displacement
+- **reason:** The robot successfully grasped the mushroom and began carrying it toward the green plate, but dropped the mushroom onto the table mid-transit.
+  Summary: dropping held mushroom; failed transport to plate
+- **recovery:** Re-grasp the dropped mushroom on the table, transport it over the green plate, and release it onto the plate.
+  Summary: re-grasp dropped mushroom; place mushroom on plate
+- **prevention:** Maintain a firm grip on the mushroom while moving it toward the green plate to prevent it from slipping out of the gripper.
+  Summary: maintain firm grasp; transport mushroom carefully
+- **justification:** Around 5.5 seconds into the video, as the robot arm moves leftward toward the green plate, the mushroom slips out of the gripper and falls onto the tabletop, preventing task completion.
diff --git a/mushroom/ep008/labels.json b/mushroom/ep008/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..70602fa9998f464f01d2c6b5553751aebff7ac4a
--- /dev/null
+++ b/mushroom/ep008/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t001__ep008__mushroom",
+  "scene_variant": "",
+  "task_id": 1,
+  "task_file": "",
+  "instruction": "Put the mushroom on the plate",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:57:01+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00796812749004,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "descending and grasping mushroom",
+      "description": "descending and grasping mushroom"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00796812749004,
+      "t_end_sec": 6.01593625498008,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "carrying and dropping mushroom",
+      "description": "carrying and dropping mushroom"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01593625498008,
+      "t_end_sec": 7.330677290836653,
+      "control_step_start": 302,
+      "control_step_end": 367,
+      "policy_step_start": 50,
+      "policy_step_end": 61,
+      "phrase": "hovering empty gripper above plate",
+      "description": "hovering empty gripper above plate"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "object_displacement",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 2.549800796812749,
+    "vlm_failure_onset_timestamp": "00:02.5",
+    "vlm_failure_onset_frame": 128,
+    "vlm_failure_onset_step": 128,
+    "vlm_coarse_failure_onset_seconds": 5.5,
+    "vlm_failure_window_start_seconds": 5.0,
+    "vlm_failure_window_end_seconds": 6.0,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot successfully grasped the mushroom and began carrying it toward the green plate, but dropped the mushroom onto the table mid-transit.",
+    "vlm_failure_reason_summary": "dropping held mushroom; failed transport to plate",
+    "vlm_recovery_action": "Re-grasp the dropped mushroom on the table, transport it over the green plate, and release it onto the plate.",
+    "vlm_recovery_action_summary": "re-grasp dropped mushroom; place mushroom on plate",
+    "vlm_prevention_action": "Maintain a firm grip on the mushroom while moving it toward the green plate to prevent it from slipping out of the gripper.",
+    "vlm_prevention_action_summary": "maintain firm grasp; transport mushroom carefully",
+    "vlm_justification": "Around 5.5 seconds into the video, as the robot arm moves leftward toward the green plate, the mushroom slips out of the gripper and falls onto the tabletop, preventing task completion.",
+    "vlm_temporal_justification": "The robot carries the mushroom toward the green plate until 02:59, when the gripper prematurely opens/loses grip and the mushroom falls onto the table.",
+    "vlm_video_fps": 50.2,
+    "vlm_video_num_frames": 368,
+    "vlm_video_duration_seconds": 7.3107569721115535,
+    "vlm_usage": {
+      "input_tokens": 17604,
+      "output_tokens": 346,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"object_displacement\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 5.5,\n \"failure_window_start_seconds\": 5.0,\n \"failure_window_end_seconds\": 6.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot successfully grasped the mushroom and began carrying it toward the green plate, but dropped the mushroom onto the table mid-transit.\",\n \"failure_reason_summary\": \"dropping held mushroom; failed transport to plate\",\n \"recovery_action\": \"Re-grasp the dropped mushroom on the table, transport it over the green plate, and release it onto the plate.\",\n \"recovery_action_summary\": \"re-grasp dropped mushroom; place mushroom on plate\",\n \"prevention_action\": \"Maintain a firm grip on the mushroom while moving it toward the green plate to prevent it from slipping out of the gripper.\",\n \"prevention_action_summary\": \"maintain firm grasp; transport mushroom carefully\",\n \"justification\": \"Around 5.5 seconds into the video, as the robot arm moves leftward toward the green plate, the mushroom slips out of the gripper and falls onto the tabletop, preventing task completion.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 128,
+    "failure_sim_step": 128,
+    "failure_policy_step": 21,
+    "failure_chunk_index": 2,
+    "first_post_failure_policy_step": 22,
+    "failure_type": "object_displacement",
+    "correction_action": "Re-grasp the dropped mushroom on the table, transport it over the green plate, and release it onto the plate."
+  }
+}
\ No newline at end of file
diff --git a/mushroom/ep008/labels.npz b/mushroom/ep008/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..22abfadb0964068ea794e710db7ed49d729eca23
--- /dev/null
+++ b/mushroom/ep008/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:54ea40c53e2ef7e2e81b9d75b000b4a5e38bbcfda1459418a4f3a77daf23f14e
+size 3770
diff --git a/mushroom/ep009/example.md b/mushroom/ep009/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..150cce5f4863fb5fe4c70a34caaaa9039866b4f7
--- /dev/null
+++ b/mushroom/ep009/example.md
@@ -0,0 +1,33 @@
+# t001__ep009__mushroom
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the mushroom on the plate
+- **success:** False
+- **control steps:** 306  (6.1s at 50 Hz)
+- **policy inferences:** 53  (replan_steps=6)
+- **video:** `mushroom/ep009/rollout.mp4`
+- **features:** `53` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | descending and grasping at mushroom |
+| 3.0–6.0 | 151–301 | 25–50 | lifting and moving toward plate |
+| 6.0–6.1 | 302–305 | 50–50 | hovering empty above plate |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 3.685258964143426 s (00:03.7)
+- **frame / control_step:** 185 (policy_step=30, chunk_index=5)
+- **coarse onset:** 3.0 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the mushroom, but the gripper failed to secure it, lifting up empty while leaving the mushroom on the table.
+  Summary: failed grasp on mushroom; lifting empty gripper
+- **recovery:** Lower the gripper back over the mushroom, align the jaws around the object to achieve a firm grasp, and then carry it to the green plate.
+  Summary: re-align gripper over mushroom; secure grasp; place on plate
+- **prevention:** Center the end effector precisely over the mushroom and ensure a solid grip before initiating the upward motion.
+  Summary: center gripper over mushroom; verify secure grasp before lifting
+- **justification:** At around 3.0 seconds, the robot closes its gripper on the mushroom and lifts up, but fails to hold the mushroom, proceeding to move toward the plate with empty fingers.
diff --git a/mushroom/ep009/labels.json b/mushroom/ep009/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..186868b73781e988ee8a4dbbdccf387e0700df8a
--- /dev/null
+++ b/mushroom/ep009/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t001__ep009__mushroom",
+  "scene_variant": "",
+  "task_id": 1,
+  "task_file": "",
+  "instruction": "Put the mushroom on the plate",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:57:37+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00796812749004,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "descending and grasping at mushroom",
+      "description": "descending and grasping at mushroom"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00796812749004,
+      "t_end_sec": 6.01593625498008,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "lifting and moving toward plate",
+      "description": "lifting and moving toward plate"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01593625498008,
+      "t_end_sec": 6.095617529880478,
+      "control_step_start": 302,
+      "control_step_end": 305,
+      "policy_step_start": 50,
+      "policy_step_end": 50,
+      "phrase": "hovering empty above plate",
+      "description": "hovering empty above plate"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 3.685258964143426,
+    "vlm_failure_onset_timestamp": "00:03.7",
+    "vlm_failure_onset_frame": 185,
+    "vlm_failure_onset_step": 185,
+    "vlm_coarse_failure_onset_seconds": 3.0,
+    "vlm_failure_window_start_seconds": 2.8,
+    "vlm_failure_window_end_seconds": 3.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the mushroom, but the gripper failed to secure it, lifting up empty while leaving the mushroom on the table.",
+    "vlm_failure_reason_summary": "failed grasp on mushroom; lifting empty gripper",
+    "vlm_recovery_action": "Lower the gripper back over the mushroom, align the jaws around the object to achieve a firm grasp, and then carry it to the green plate.",
+    "vlm_recovery_action_summary": "re-align gripper over mushroom; secure grasp; place on plate",
+    "vlm_prevention_action": "Center the end effector precisely over the mushroom and ensure a solid grip before initiating the upward motion.",
+    "vlm_prevention_action_summary": "center gripper over mushroom; verify secure grasp before lifting",
+    "vlm_justification": "At around 3.0 seconds, the robot closes its gripper on the mushroom and lifts up, but fails to hold the mushroom, proceeding to move toward the plate with empty fingers.",
+    "vlm_temporal_justification": "At clip second 185, the robot's gripper finishes closing completely without securing the mushroom, failing the grasp attempt as the arm begins to lift up empty while the mushroom remains on the table.",
+    "vlm_video_fps": 50.2,
+    "vlm_video_num_frames": 306,
+    "vlm_video_duration_seconds": 6.075697211155378,
+    "vlm_usage": {
+      "input_tokens": 21316,
+      "output_tokens": 357,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 3.0,\n \"failure_window_start_seconds\": 2.8,\n \"failure_window_end_seconds\": 3.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the mushroom, but the gripper failed to secure it, lifting up empty while leaving the mushroom on the table.\",\n \"failure_reason_summary\": \"failed grasp on mushroom; lifting empty gripper\",\n \"recovery_action\": \"Lower the gripper back over the mushroom, align the jaws around the object to achieve a firm grasp, and then carry it to the green plate.\",\n \"recovery_action_summary\": \"re-align gripper over mushroom; secure grasp; place on plate\",\n \"prevention_action\": \"Center the end effector precisely over the mushroom and ensure a solid grip before initiating the upward motion.\",\n \"prevention_action_summary\": \"center gripper over mushroom; verify secure grasp before lifting\",\n \"justification\": \"At around 3.0 seconds, the robot closes its gripper on the mushroom and lifts up, but fails to hold the mushroom, proceeding to move toward the plate with empty fingers.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 185,
+    "failure_sim_step": 185,
+    "failure_policy_step": 30,
+    "failure_chunk_index": 5,
+    "first_post_failure_policy_step": 31,
+    "failure_type": "grasp_failure",
+    "correction_action": "Lower the gripper back over the mushroom, align the jaws around the object to achieve a firm grasp, and then carry it to the green plate."
+  }
+}
\ No newline at end of file
diff --git a/mushroom/ep009/labels.npz b/mushroom/ep009/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..a4a9d644284df13ee0dde6307992a66e4312e3b1
--- /dev/null
+++ b/mushroom/ep009/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f2a6a1cdfeb959bd55b483fc0b089028a20f1e2f7b41924f644f3b91a703eb62
+size 3804
diff --git a/mushroom/ep010/example.md b/mushroom/ep010/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..36df85402f3476bc3b9c3acf791037e9f86c3763
--- /dev/null
+++ b/mushroom/ep010/example.md
@@ -0,0 +1,34 @@
+# t001__ep010__mushroom
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the mushroom on the plate
+- **success:** False
+- **control steps:** 469  (9.2s at 51 Hz)
+- **policy inferences:** 80  (replan_steps=6)
+- **video:** `mushroom/ep010/rollout.mp4`
+- **features:** `80` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–152 | 0–25 | lowering gripper toward mushroom |
+| 3.0–6.0 | 153–305 | 25–50 | attempting grasp and lifting empty |
+| 6.0–9.0 | 306–458 | 51–76 | moving empty gripper over plate |
+| 9.0–9.2 | 459–468 | 76–78 | hovering above plate |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 5.481335952848723 s (00:05.5)
+- **frame / control_step:** 279 (policy_step=46, chunk_index=3)
+- **coarse onset:** 3.0 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to close its gripper around the mushroom but failed to establish a firm grip, subsequently lifting its end-effector empty and moving to the plate.
+  Summary: failed grasp on mushroom; moving empty-handed
+- **recovery:** Lower the gripper back down, re-align over the mushroom stem, establish a secure grasp, and place it on the green plate.
+  Summary: re-align gripper; grasp mushroom; place on plate
+- **prevention:** Position the gripper directly around the mushroom before closing, ensuring a firm hold before lifting and moving towards the plate.
+  Summary: center gripper on mushroom; secure grasp; transport to plate
+- **justification:** At around 3.0 seconds, the robot's gripper attempts to grasp the mushroom but misses/slips, failing to acquire the object. The robot then lifts up empty at 5.0 seconds and moves toward the plate without the mushroom.
diff --git a/mushroom/ep010/labels.json b/mushroom/ep010/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..09ff614d5c4870ed99ce75b8523ebe1b12306a64
--- /dev/null
+++ b/mushroom/ep010/labels.json
@@ -0,0 +1,109 @@
+{
+  "rollout_id": "t001__ep010__mushroom",
+  "scene_variant": "",
+  "task_id": 1,
+  "task_file": "",
+  "instruction": "Put the mushroom on the plate",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:58:07+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.005893909626719,
+      "control_step_start": 0,
+      "control_step_end": 152,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "lowering gripper toward mushroom",
+      "description": "lowering gripper toward mushroom"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.005893909626719,
+      "t_end_sec": 6.011787819253438,
+      "control_step_start": 153,
+      "control_step_end": 305,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "attempting grasp and lifting empty",
+      "description": "attempting grasp and lifting empty"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.011787819253438,
+      "t_end_sec": 9.017681728880158,
+      "control_step_start": 306,
+      "control_step_end": 458,
+      "policy_step_start": 51,
+      "policy_step_end": 76,
+      "phrase": "moving empty gripper over plate",
+      "description": "moving empty gripper over plate"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.017681728880158,
+      "t_end_sec": 9.214145383104126,
+      "control_step_start": 459,
+      "control_step_end": 468,
+      "policy_step_start": 76,
+      "policy_step_end": 78,
+      "phrase": "hovering above plate",
+      "description": "hovering above plate"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 5.481335952848723,
+    "vlm_failure_onset_timestamp": "00:05.5",
+    "vlm_failure_onset_frame": 279,
+    "vlm_failure_onset_step": 279,
+    "vlm_coarse_failure_onset_seconds": 3.0,
+    "vlm_failure_window_start_seconds": 2.5,
+    "vlm_failure_window_end_seconds": 4.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to close its gripper around the mushroom but failed to establish a firm grip, subsequently lifting its end-effector empty and moving to the plate.",
+    "vlm_failure_reason_summary": "failed grasp on mushroom; moving empty-handed",
+    "vlm_recovery_action": "Lower the gripper back down, re-align over the mushroom stem, establish a secure grasp, and place it on the green plate.",
+    "vlm_recovery_action_summary": "re-align gripper; grasp mushroom; place on plate",
+    "vlm_prevention_action": "Position the gripper directly around the mushroom before closing, ensuring a firm hold before lifting and moving towards the plate.",
+    "vlm_prevention_action_summary": "center gripper on mushroom; secure grasp; transport to plate",
+    "vlm_justification": "At around 3.0 seconds, the robot's gripper attempts to grasp the mushroom but misses/slips, failing to acquire the object. The robot then lifts up empty at 5.0 seconds and moves toward the plate without the mushroom.",
+    "vlm_temporal_justification": "At refined second 279 (04:39), the robot's gripper completely closes without securing the mushroom, leaving the mushroom behind on the table as the end-effector begins to lift up empty at second 280 (04:40).",
+    "vlm_video_fps": 50.9,
+    "vlm_video_num_frames": 469,
+    "vlm_video_duration_seconds": 9.194499017681729,
+    "vlm_usage": {
+      "input_tokens": 21761,
+      "output_tokens": 390,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 3.0,\n \"failure_window_start_seconds\": 2.5,\n \"failure_window_end_seconds\": 4.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to close its gripper around the mushroom but failed to establish a firm grip, subsequently lifting its end-effector empty and moving to the plate.\",\n \"failure_reason_summary\": \"failed grasp on mushroom; moving empty-handed\",\n \"recovery_action\": \"Lower the gripper back down, re-align over the mushroom stem, establish a secure grasp, and place it on the green plate.\",\n \"recovery_action_summary\": \"re-align gripper; grasp mushroom; place on plate\",\n \"prevention_action\": \"Position the gripper directly around the mushroom before closing, ensuring a firm hold before lifting and moving towards the plate.\",\n \"prevention_action_summary\": \"center gripper on mushroom; secure grasp; transport to plate\",\n \"justification\": \"At around 3.0 seconds, the robot's gripper attempts to grasp the mushroom but misses/slips, failing to acquire the object. The robot then lifts up empty at 5.0 seconds and moves toward the plate without the mushroom.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 279,
+    "failure_sim_step": 279,
+    "failure_policy_step": 46,
+    "failure_chunk_index": 3,
+    "first_post_failure_policy_step": 47,
+    "failure_type": "grasp_failure",
+    "correction_action": "Lower the gripper back down, re-align over the mushroom stem, establish a secure grasp, and place it on the green plate."
+  }
+}
\ No newline at end of file
diff --git a/mushroom/ep010/labels.npz b/mushroom/ep010/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..eb4f119d8a20eacb367b59bbc5f626470078a346
--- /dev/null
+++ b/mushroom/ep010/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c5f67d313c6b77fe735f8b50b1582549858d5df0bb6c8dc2eb7c30e7633cab30
+size 3871
diff --git a/mushroom/ep011/example.md b/mushroom/ep011/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..1851df7ee32ae7c07a775febf924bc8399e6ea01
--- /dev/null
+++ b/mushroom/ep011/example.md
@@ -0,0 +1,33 @@
+# t001__ep011__mushroom
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the mushroom on the plate
+- **success:** False
+- **control steps:** 415  (8.3s at 50 Hz)
+- **policy inferences:** 73  (replan_steps=6)
+- **video:** `mushroom/ep011/rollout.mp4`
+- **features:** `73` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | reaching for and grasping mushroom |
+| 3.0–6.0 | 151–301 | 25–50 | lifting empty gripper toward plate |
+| 6.0–8.3 | 302–414 | 50–69 | lowering empty gripper over plate |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 3.286852589641434 s (00:03.3)
+- **frame / control_step:** 165 (policy_step=27, chunk_index=3)
+- **coarse onset:** 3.0 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the mushroom but failed to establish a secure grip, causing the mushroom to slip out and stay on the table as the end-effector lifted away empty.
+  Summary: unsuccessful grasp attempt; lifting empty gripper
+- **recovery:** Lower the end-effector back down to the mushroom, align the gripper fingers around its body, close the gripper securely, and transport it to the plate.
+  Summary: re-align gripper; regrasp mushroom; move to plate
+- **prevention:** Position the gripper lower and align it carefully over the mushroom before closing the fingers to guarantee a secure hold before lifting.
+  Summary: center gripper low; secure firm grasp
+- **justification:** At approximately 3.0 seconds, the robot closes its gripper around the mushroom and begins moving upward, but the mushroom slips out and remains on the tabletop while the gripper moves to the plate empty.
diff --git a/mushroom/ep011/labels.json b/mushroom/ep011/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..e501ba3e5efd570b6ea32475f66392501ffdc6ba
--- /dev/null
+++ b/mushroom/ep011/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t001__ep011__mushroom",
+  "scene_variant": "",
+  "task_id": 1,
+  "task_file": "",
+  "instruction": "Put the mushroom on the plate",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:58:41+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00796812749004,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for and grasping mushroom",
+      "description": "reaching for and grasping mushroom"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00796812749004,
+      "t_end_sec": 6.01593625498008,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "lifting empty gripper toward plate",
+      "description": "lifting empty gripper toward plate"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01593625498008,
+      "t_end_sec": 8.266932270916334,
+      "control_step_start": 302,
+      "control_step_end": 414,
+      "policy_step_start": 50,
+      "policy_step_end": 69,
+      "phrase": "lowering empty gripper over plate",
+      "description": "lowering empty gripper over plate"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 3.286852589641434,
+    "vlm_failure_onset_timestamp": "00:03.3",
+    "vlm_failure_onset_frame": 165,
+    "vlm_failure_onset_step": 165,
+    "vlm_coarse_failure_onset_seconds": 3.0,
+    "vlm_failure_window_start_seconds": 2.5,
+    "vlm_failure_window_end_seconds": 3.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the mushroom but failed to establish a secure grip, causing the mushroom to slip out and stay on the table as the end-effector lifted away empty.",
+    "vlm_failure_reason_summary": "unsuccessful grasp attempt; lifting empty gripper",
+    "vlm_recovery_action": "Lower the end-effector back down to the mushroom, align the gripper fingers around its body, close the gripper securely, and transport it to the plate.",
+    "vlm_recovery_action_summary": "re-align gripper; regrasp mushroom; move to plate",
+    "vlm_prevention_action": "Position the gripper lower and align it carefully over the mushroom before closing the fingers to guarantee a secure hold before lifting.",
+    "vlm_prevention_action_summary": "center gripper low; secure firm grasp",
+    "vlm_justification": "At approximately 3.0 seconds, the robot closes its gripper around the mushroom and begins moving upward, but the mushroom slips out and remains on the tabletop while the gripper moves to the plate empty.",
+    "vlm_temporal_justification": "At 02:45 (frame 165 of the rollout), the robot's end-effector finishes closing its gripper around the mushroom without securing a grip and begins lifting away empty, leaving the mushroom behind on the table.",
+    "vlm_video_fps": 50.2,
+    "vlm_video_num_frames": 415,
+    "vlm_video_duration_seconds": 8.247011952191235,
+    "vlm_usage": {
+      "input_tokens": 21450,
+      "output_tokens": 378,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 3.0,\n \"failure_window_start_seconds\": 2.5,\n \"failure_window_end_seconds\": 3.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the mushroom but failed to establish a secure grip, causing the mushroom to slip out and stay on the table as the end-effector lifted away empty.\",\n \"failure_reason_summary\": \"unsuccessful grasp attempt; lifting empty gripper\",\n \"recovery_action\": \"Lower the end-effector back down to the mushroom, align the gripper fingers around its body, close the gripper securely, and transport it to the plate.\",\n \"recovery_action_summary\": \"re-align gripper; regrasp mushroom; move to plate\",\n \"prevention_action\": \"Position the gripper lower and align it carefully over the mushroom before closing the fingers to guarantee a secure hold before lifting.\",\n \"prevention_action_summary\": \"center gripper low; secure firm grasp\",\n \"justification\": \"At approximately 3.0 seconds, the robot closes its gripper around the mushroom and begins moving upward, but the mushroom slips out and remains on the tabletop while the gripper moves to the plate empty.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 165,
+    "failure_sim_step": 165,
+    "failure_policy_step": 27,
+    "failure_chunk_index": 3,
+    "first_post_failure_policy_step": 28,
+    "failure_type": "grasp_failure",
+    "correction_action": "Lower the end-effector back down to the mushroom, align the gripper fingers around its body, close the gripper securely, and transport it to the plate."
+  }
+}
\ No newline at end of file
diff --git a/mushroom/ep011/labels.npz b/mushroom/ep011/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..f49d228fbf003f27151dee19c9a8db34c0975c8a
--- /dev/null
+++ b/mushroom/ep011/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6bdfd6c10acdd65a5ea39cb17a5c00758f1479449d5da0a8ab7d0891a10abc78
+size 3848
diff --git a/mushroom/ep012/example.md b/mushroom/ep012/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..5230b387808822eecb1634cc201d9007f08e4b29
--- /dev/null
+++ b/mushroom/ep012/example.md
@@ -0,0 +1,33 @@
+# t001__ep012__mushroom
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the mushroom on the plate
+- **success:** False
+- **control steps:** 408  (8.0s at 51 Hz)
+- **policy inferences:** 70  (replan_steps=6)
+- **video:** `mushroom/ep012/rollout.mp4`
+- **features:** `70` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching and grasping mushroom |
+| 3.0–6.0 | 152–303 | 25–50 | carrying mushroom to plate |
+| 6.0–8.0 | 304–407 | 50–67 | releasing mushroom; retracting arm |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 3.070866141732284 s (00:03.1)
+- **frame / control_step:** 156 (policy_step=26, chunk_index=0)
+- **coarse onset:** 6.0 s (refined=True)
+- **mode:** placement_or_insertion_failure
+- **reason:** The robot brought the mushroom to the green plate but released it too close to the edge, causing the mushroom to roll off the plate and land back on the table surface.
+  Summary: improper placement position; object rolling off plate
+- **recovery:** Re-grasp the mushroom from the table, move over to the center of the green plate, and gently release it.
+  Summary: re-grasp mushroom; place in center of plate
+- **prevention:** Center the robot gripper directly above the plate and lower the mushroom closer to the plate surface before opening the gripper.
+  Summary: center gripper over plate; lower object before releasing
+- **justification:** At 6.0 seconds, the robot opens its gripper to drop the mushroom, but because it is positioned over the rim of the green plate, the mushroom slides off onto the table. This is an obvious placement failure.
diff --git a/mushroom/ep012/labels.json b/mushroom/ep012/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..bc6079f9b677a14ebfbc2f1e85330f8ba7db3985
--- /dev/null
+++ b/mushroom/ep012/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t001__ep012__mushroom",
+  "scene_variant": "",
+  "task_id": 1,
+  "task_file": "",
+  "instruction": "Put the mushroom on the plate",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:59:14+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9921259842519685,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching and grasping mushroom",
+      "description": "reaching and grasping mushroom"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9921259842519685,
+      "t_end_sec": 5.984251968503937,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "carrying mushroom to plate",
+      "description": "carrying mushroom to plate"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.984251968503937,
+      "t_end_sec": 8.031496062992126,
+      "control_step_start": 304,
+      "control_step_end": 407,
+      "policy_step_start": 50,
+      "policy_step_end": 67,
+      "phrase": "releasing mushroom; retracting arm",
+      "description": "releasing mushroom; retracting arm"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "placement_or_insertion_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 3.070866141732284,
+    "vlm_failure_onset_timestamp": "00:03.1",
+    "vlm_failure_onset_frame": 156,
+    "vlm_failure_onset_step": 156,
+    "vlm_coarse_failure_onset_seconds": 6.0,
+    "vlm_failure_window_start_seconds": 5.5,
+    "vlm_failure_window_end_seconds": 6.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot brought the mushroom to the green plate but released it too close to the edge, causing the mushroom to roll off the plate and land back on the table surface.",
+    "vlm_failure_reason_summary": "improper placement position; object rolling off plate",
+    "vlm_recovery_action": "Re-grasp the mushroom from the table, move over to the center of the green plate, and gently release it.",
+    "vlm_recovery_action_summary": "re-grasp mushroom; place in center of plate",
+    "vlm_prevention_action": "Center the robot gripper directly above the plate and lower the mushroom closer to the plate surface before opening the gripper.",
+    "vlm_prevention_action_summary": "center gripper over plate; lower object before releasing",
+    "vlm_justification": "At 6.0 seconds, the robot opens its gripper to drop the mushroom, but because it is positioned over the rim of the green plate, the mushroom slides off onto the table. This is an obvious placement failure.",
+    "vlm_temporal_justification": "The robot opens its gripper to release the mushroom right over the edge of the plate at second 02:56, causing the mushroom to roll off the plate rim and land on the table.",
+    "vlm_video_fps": 50.8,
+    "vlm_video_num_frames": 408,
+    "vlm_video_duration_seconds": 8.011811023622048,
+    "vlm_usage": {
+      "input_tokens": 18435,
+      "output_tokens": 371,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"placement_or_insertion_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 6.0,\n \"failure_window_start_seconds\": 5.5,\n \"failure_window_end_seconds\": 6.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot brought the mushroom to the green plate but released it too close to the edge, causing the mushroom to roll off the plate and land back on the table surface.\",\n \"failure_reason_summary\": \"improper placement position; object rolling off plate\",\n \"recovery_action\": \"Re-grasp the mushroom from the table, move over to the center of the green plate, and gently release it.\",\n \"recovery_action_summary\": \"re-grasp mushroom; place in center of plate\",\n \"prevention_action\": \"Center the robot gripper directly above the plate and lower the mushroom closer to the plate surface before opening the gripper.\",\n \"prevention_action_summary\": \"center gripper over plate; lower object before releasing\",\n \"justification\": \"At 6.0 seconds, the robot opens its gripper to drop the mushroom, but because it is positioned over the rim of the green plate, the mushroom slides off onto the table. This is an obvious placement failure.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 156,
+    "failure_sim_step": 156,
+    "failure_policy_step": 26,
+    "failure_chunk_index": 0,
+    "first_post_failure_policy_step": 27,
+    "failure_type": "placement_or_insertion_failure",
+    "correction_action": "Re-grasp the mushroom from the table, move over to the center of the green plate, and gently release it."
+  }
+}
\ No newline at end of file
diff --git a/mushroom/ep012/labels.npz b/mushroom/ep012/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..9a0e2d80f0eeb6d695c5c296451dfae633c34696
--- /dev/null
+++ b/mushroom/ep012/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f2e3df63f88e602314a890fb9ed29f85e88d7371c6165f05ae3d7cab5020d480
+size 3795
diff --git a/mushroom/ep013/example.md b/mushroom/ep013/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..df7ffc65e98998f4c0ff9f949bc970617c8a2f0d
--- /dev/null
+++ b/mushroom/ep013/example.md
@@ -0,0 +1,33 @@
+# t001__ep013__mushroom
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the mushroom on the plate
+- **success:** False
+- **control steps:** 405  (8.0s at 51 Hz)
+- **policy inferences:** 70  (replan_steps=6)
+- **video:** `mushroom/ep013/rollout.mp4`
+- **features:** `70` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–152 | 0–25 | reaching for mushroom; positioning gripper |
+| 3.0–6.0 | 153–305 | 25–50 | grasping mushroom; moving toward plate |
+| 6.0–8.0 | 306–404 | 51–67 | dropping mushroom; retracting arm |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 6.070726915520629 s (00:06.1)
+- **frame / control_step:** 309 (policy_step=51, chunk_index=3)
+- **coarse onset:** 5.2 s (refined=True)
+- **mode:** object_displacement
+- **reason:** The robot successfully grasped the mushroom and carried it toward the green plate, but dropped it prematurely onto the table surface outside the plate.
+  Summary: dropping grasped object; missing green plate
+- **recovery:** Re-grasp the mushroom from the table surface and move it onto the center of the green plate.
+  Summary: re-grasp mushroom; place on plate
+- **prevention:** Maintain a secure grip on the mushroom until the end-effector is positioned directly above the center of the green plate before releasing.
+  Summary: hold mushroom securely; release centered over plate
+- **justification:** Around 5.2 seconds into the video, as the robot arm reaches the edge of the plate, it drops the held mushroom onto the table next to the plate rather than placing it onto the plate.
diff --git a/mushroom/ep013/labels.json b/mushroom/ep013/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..9e9b51d319ac0755cca83e8fc66dc0f8028b2efa
--- /dev/null
+++ b/mushroom/ep013/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t001__ep013__mushroom",
+  "scene_variant": "",
+  "task_id": 1,
+  "task_file": "",
+  "instruction": "Put the mushroom on the plate",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T20:59:50+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.005893909626719,
+      "control_step_start": 0,
+      "control_step_end": 152,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for mushroom; positioning gripper",
+      "description": "reaching for mushroom; positioning gripper"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.005893909626719,
+      "t_end_sec": 6.011787819253438,
+      "control_step_start": 153,
+      "control_step_end": 305,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "grasping mushroom; moving toward plate",
+      "description": "grasping mushroom; moving toward plate"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.011787819253438,
+      "t_end_sec": 7.956777996070727,
+      "control_step_start": 306,
+      "control_step_end": 404,
+      "policy_step_start": 51,
+      "policy_step_end": 67,
+      "phrase": "dropping mushroom; retracting arm",
+      "description": "dropping mushroom; retracting arm"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "object_displacement",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 6.070726915520629,
+    "vlm_failure_onset_timestamp": "00:06.1",
+    "vlm_failure_onset_frame": 309,
+    "vlm_failure_onset_step": 309,
+    "vlm_coarse_failure_onset_seconds": 5.2,
+    "vlm_failure_window_start_seconds": 5.0,
+    "vlm_failure_window_end_seconds": 6.0,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot successfully grasped the mushroom and carried it toward the green plate, but dropped it prematurely onto the table surface outside the plate.",
+    "vlm_failure_reason_summary": "dropping grasped object; missing green plate",
+    "vlm_recovery_action": "Re-grasp the mushroom from the table surface and move it onto the center of the green plate.",
+    "vlm_recovery_action_summary": "re-grasp mushroom; place on plate",
+    "vlm_prevention_action": "Maintain a secure grip on the mushroom until the end-effector is positioned directly above the center of the green plate before releasing.",
+    "vlm_prevention_action_summary": "hold mushroom securely; release centered over plate",
+    "vlm_justification": "Around 5.2 seconds into the video, as the robot arm reaches the edge of the plate, it drops the held mushroom onto the table next to the plate rather than placing it onto the plate.",
+    "vlm_temporal_justification": "At timestamp 03:17, the mushroom visibly begins slipping out from between the robot's gripper fingers before reaching the green plate, landing on the table surface at 03:18.",
+    "vlm_video_fps": 50.9,
+    "vlm_video_num_frames": 405,
+    "vlm_video_duration_seconds": 7.93713163064833,
+    "vlm_usage": {
+      "input_tokens": 20817,
+      "output_tokens": 350,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"object_displacement\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 5.2,\n \"failure_window_start_seconds\": 5.0,\n \"failure_window_end_seconds\": 6.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot successfully grasped the mushroom and carried it toward the green plate, but dropped it prematurely onto the table surface outside the plate.\",\n \"failure_reason_summary\": \"dropping grasped object; missing green plate\",\n \"recovery_action\": \"Re-grasp the mushroom from the table surface and move it onto the center of the green plate.\",\n \"recovery_action_summary\": \"re-grasp mushroom; place on plate\",\n \"prevention_action\": \"Maintain a secure grip on the mushroom until the end-effector is positioned directly above the center of the green plate before releasing.\",\n \"prevention_action_summary\": \"hold mushroom securely; release centered over plate\",\n \"justification\": \"Around 5.2 seconds into the video, as the robot arm reaches the edge of the plate, it drops the held mushroom onto the table next to the plate rather than placing it onto the plate.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 309,
+    "failure_sim_step": 309,
+    "failure_policy_step": 51,
+    "failure_chunk_index": 3,
+    "first_post_failure_policy_step": 52,
+    "failure_type": "object_displacement",
+    "correction_action": "Re-grasp the mushroom from the table surface and move it onto the center of the green plate."
+  }
+}
\ No newline at end of file
diff --git a/mushroom/ep013/labels.npz b/mushroom/ep013/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..ea6b863383645b31692602ef6bcaee56f7b601de
--- /dev/null
+++ b/mushroom/ep013/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:25f66e56878cd890c5dd577429c89572795f3c3d354e719c60c6c7b8cae179dc
+size 3802
diff --git a/mushroom/ep014/example.md b/mushroom/ep014/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..a089308b15431fc412e4c591ffb872203de63245
--- /dev/null
+++ b/mushroom/ep014/example.md
@@ -0,0 +1,34 @@
+# t001__ep014__mushroom
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the mushroom on the plate
+- **success:** False
+- **control steps:** 462  (9.2s at 50 Hz)
+- **policy inferences:** 80  (replan_steps=6)
+- **video:** `mushroom/ep014/rollout.mp4`
+- **features:** `80` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | reaching for and grasping mushroom |
+| 3.0–6.0 | 151–301 | 25–50 | lifting and carrying mushroom |
+| 6.0–9.0 | 302–452 | 50–75 | releasing mushroom at plate edge |
+| 9.0–9.2 | 453–461 | 75–76 | idle above table |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 6.779324055666004 s (00:06.8)
+- **frame / control_step:** 341 (policy_step=56, chunk_index=5)
+- **coarse onset:** 6.0 s (refined=True)
+- **mode:** placement_or_insertion_failure
+- **reason:** The robot released the mushroom too close to the edge of the green plate, causing the mushroom to roll off onto the table surface.
+  Summary: improper placement position; mushroom rolling off plate
+- **recovery:** Re-grasp the mushroom from the table surface and place it securely near the center of the green plate.
+  Summary: re-grasp mushroom; place centered on plate
+- **prevention:** Move the end-effector further toward the center of the green plate before releasing the mushroom.
+  Summary: center gripper over plate; release mushroom
+- **justification:** At around 6.0 seconds into the rollout, the robot positions the mushroom right at the front edge of the plate and releases it, causing the mushroom to bounce and roll off onto the table.
diff --git a/mushroom/ep014/labels.json b/mushroom/ep014/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..49884eac669ed2cde15d6a3c084c17bf181b4622
--- /dev/null
+++ b/mushroom/ep014/labels.json
@@ -0,0 +1,109 @@
+{
+  "rollout_id": "t001__ep014__mushroom",
+  "scene_variant": "",
+  "task_id": 1,
+  "task_file": "",
+  "instruction": "Put the mushroom on the plate",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:00:23+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.001988071570577,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for and grasping mushroom",
+      "description": "reaching for and grasping mushroom"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.001988071570577,
+      "t_end_sec": 6.003976143141154,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "lifting and carrying mushroom",
+      "description": "lifting and carrying mushroom"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.003976143141154,
+      "t_end_sec": 9.00596421471173,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "releasing mushroom at plate edge",
+      "description": "releasing mushroom at plate edge"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.00596421471173,
+      "t_end_sec": 9.184890656063619,
+      "control_step_start": 453,
+      "control_step_end": 461,
+      "policy_step_start": 75,
+      "policy_step_end": 76,
+      "phrase": "idle above table",
+      "description": "idle above table"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "placement_or_insertion_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 6.779324055666004,
+    "vlm_failure_onset_timestamp": "00:06.8",
+    "vlm_failure_onset_frame": 341,
+    "vlm_failure_onset_step": 341,
+    "vlm_coarse_failure_onset_seconds": 6.0,
+    "vlm_failure_window_start_seconds": 5.5,
+    "vlm_failure_window_end_seconds": 7.0,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot released the mushroom too close to the edge of the green plate, causing the mushroom to roll off onto the table surface.",
+    "vlm_failure_reason_summary": "improper placement position; mushroom rolling off plate",
+    "vlm_recovery_action": "Re-grasp the mushroom from the table surface and place it securely near the center of the green plate.",
+    "vlm_recovery_action_summary": "re-grasp mushroom; place centered on plate",
+    "vlm_prevention_action": "Move the end-effector further toward the center of the green plate before releasing the mushroom.",
+    "vlm_prevention_action_summary": "center gripper over plate; release mushroom",
+    "vlm_justification": "At around 6.0 seconds into the rollout, the robot positions the mushroom right at the front edge of the plate and releases it, causing the mushroom to bounce and roll off onto the table.",
+    "vlm_temporal_justification": "At 03:10 (second 190 of the clip), the robot fully releases the mushroom near the edge of the plate, causing it to slip and roll off onto the table surface.",
+    "vlm_video_fps": 50.3,
+    "vlm_video_num_frames": 462,
+    "vlm_video_duration_seconds": 9.165009940357853,
+    "vlm_usage": {
+      "input_tokens": 21514,
+      "output_tokens": 349,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"placement_or_insertion_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 6.0,\n \"failure_window_start_seconds\": 5.5,\n \"failure_window_end_seconds\": 7.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot released the mushroom too close to the edge of the green plate, causing the mushroom to roll off onto the table surface.\",\n \"failure_reason_summary\": \"improper placement position; mushroom rolling off plate\",\n \"recovery_action\": \"Re-grasp the mushroom from the table surface and place it securely near the center of the green plate.\",\n \"recovery_action_summary\": \"re-grasp mushroom; place centered on plate\",\n \"prevention_action\": \"Move the end-effector further toward the center of the green plate before releasing the mushroom.\",\n \"prevention_action_summary\": \"center gripper over plate; release mushroom\",\n \"justification\": \"At around 6.0 seconds into the rollout, the robot positions the mushroom right at the front edge of the plate and releases it, causing the mushroom to bounce and roll off onto the table.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 341,
+    "failure_sim_step": 341,
+    "failure_policy_step": 56,
+    "failure_chunk_index": 5,
+    "first_post_failure_policy_step": 57,
+    "failure_type": "placement_or_insertion_failure",
+    "correction_action": "Re-grasp the mushroom from the table surface and place it securely near the center of the green plate."
+  }
+}
\ No newline at end of file
diff --git a/mushroom/ep014/labels.npz b/mushroom/ep014/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..8cdd3771d5fdb1e07b693f83a2c406776158a734
--- /dev/null
+++ b/mushroom/ep014/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b56de37c186f83d5b545f7d49eeca04a131fecc2510431730f02c8ac4def3710
+size 3770
diff --git a/mushroom/ep015/example.md b/mushroom/ep015/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..e5b24547744ae79e4dafe6f6cd6bd3ba41ee4d3c
--- /dev/null
+++ b/mushroom/ep015/example.md
@@ -0,0 +1,21 @@
+# t001__ep015__mushroom
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the mushroom on the plate
+- **success:** True
+- **control steps:** 370  (7.4s at 50 Hz)
+- **policy inferences:** 64  (replan_steps=6)
+- **video:** `mushroom/ep015/rollout.mp4`
+- **features:** `64` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | reaching for and grasping mushroom |
+| 3.0–6.0 | 151–301 | 25–50 | placing mushroom on the plate |
+| 6.0–7.4 | 302–369 | 50–61 | retracting robot arm |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/mushroom/ep015/labels.json b/mushroom/ep015/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..29a0d6812fb5fb3a49ecd56fd780676115a67f25
--- /dev/null
+++ b/mushroom/ep015/labels.json
@@ -0,0 +1,68 @@
+{
+  "rollout_id": "t001__ep015__mushroom",
+  "scene_variant": "",
+  "task_id": 1,
+  "task_file": "",
+  "instruction": "Put the mushroom on the plate",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:00:30+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.001988071570577,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for and grasping mushroom",
+      "description": "reaching for and grasping mushroom"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.001988071570577,
+      "t_end_sec": 6.003976143141154,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "placing mushroom on the plate",
+      "description": "placing mushroom on the plate"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.003976143141154,
+      "t_end_sec": 7.355864811133201,
+      "control_step_start": 302,
+      "control_step_end": 369,
+      "policy_step_start": 50,
+      "policy_step_end": 61,
+      "phrase": "retracting robot arm",
+      "description": "retracting robot arm"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/mushroom/ep015/labels.npz b/mushroom/ep015/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..b9b23c10cc3d118e4c48f5f42c2a44cdbf19b436
--- /dev/null
+++ b/mushroom/ep015/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:234e1292b15d82d1c847bde6e70995a0e273e5f9e2c087dfd9a4335ee6afe70c
+size 3086
diff --git a/mushroom/ep016/example.md b/mushroom/ep016/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..2825eba4713f4d2a19fbbff97980d0b1afbef0e2
--- /dev/null
+++ b/mushroom/ep016/example.md
@@ -0,0 +1,21 @@
+# t001__ep016__mushroom
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the mushroom on the plate
+- **success:** True
+- **control steps:** 401  (7.8s at 51 Hz)
+- **policy inferences:** 69  (replan_steps=6)
+- **video:** `mushroom/ep016/rollout.mp4`
+- **features:** `69` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–153 | 0–25 | reaching and grasping the mushroom |
+| 3.0–6.0 | 154–307 | 25–51 | placing mushroom on the plate |
+| 6.0–7.8 | 308–400 | 51–66 | retracting robot arm |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/mushroom/ep016/labels.json b/mushroom/ep016/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..31e6eaa1a33eb151da50665f57ca427880149391
--- /dev/null
+++ b/mushroom/ep016/labels.json
@@ -0,0 +1,68 @@
+{
+  "rollout_id": "t001__ep016__mushroom",
+  "scene_variant": "",
+  "task_id": 1,
+  "task_file": "",
+  "instruction": "Put the mushroom on the plate",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:00:38+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0019493177387915,
+      "control_step_start": 0,
+      "control_step_end": 153,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching and grasping the mushroom",
+      "description": "reaching and grasping the mushroom"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0019493177387915,
+      "t_end_sec": 6.003898635477583,
+      "control_step_start": 154,
+      "control_step_end": 307,
+      "policy_step_start": 25,
+      "policy_step_end": 51,
+      "phrase": "placing mushroom on the plate",
+      "description": "placing mushroom on the plate"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.003898635477583,
+      "t_end_sec": 7.816764132553606,
+      "control_step_start": 308,
+      "control_step_end": 400,
+      "policy_step_start": 51,
+      "policy_step_end": 66,
+      "phrase": "retracting robot arm",
+      "description": "retracting robot arm"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/mushroom/ep016/labels.npz b/mushroom/ep016/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..52de158c6b83a43f9a00634cc7859184dfee6447
--- /dev/null
+++ b/mushroom/ep016/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ab50b0f3dfc48c9f840f53f6d32f535a0d7c9f89fb0e70bc3d58db095a65a7cf
+size 3086
diff --git a/mushroom/ep017/example.md b/mushroom/ep017/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..2d61e00aecc4050f26e82ed6d7d5ef6a1457f12a
--- /dev/null
+++ b/mushroom/ep017/example.md
@@ -0,0 +1,33 @@
+# t001__ep017__mushroom
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the mushroom on the plate
+- **success:** False
+- **control steps:** 306  (6.1s at 50 Hz)
+- **policy inferences:** 53  (replan_steps=6)
+- **video:** `mushroom/ep017/rollout.mp4`
+- **features:** `53` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | reaching for and grasping mushroom |
+| 3.0–6.0 | 151–301 | 25–50 | moving empty gripper toward plate |
+| 6.0–6.1 | 302–305 | 50–50 | hovering over plate |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 1.8650793650793651 s (00:01.9)
+- **frame / control_step:** 94 (policy_step=15, chunk_index=4)
+- **coarse onset:** 2.5 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot end-effector lowered to grasp the mushroom but failed to secure a firm grip, allowing the mushroom to slip out as the gripper closed.
+  Summary: failed grasp attempt; mushroom slipping from gripper
+- **recovery:** Re-align the end-effector over the mushroom, close the fingers securely around it, and transfer it onto the green plate.
+  Summary: re-align gripper; grasp mushroom; place on plate
+- **prevention:** Position the gripper precisely around the mushroom's center of mass before closing the fingers to ensure a secure grasp.
+  Summary: align gripper over mushroom; establish secure grasp
+- **justification:** At 2.5 seconds, the robot attempts to grasp the mushroom, but the gripper closes improperly and fails to hold the object, resulting in a grasp failure.
diff --git a/mushroom/ep017/labels.json b/mushroom/ep017/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..2401298f7491bab4bb42590ac9bc530c267b9bf5
--- /dev/null
+++ b/mushroom/ep017/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t001__ep017__mushroom",
+  "scene_variant": "",
+  "task_id": 1,
+  "task_file": "",
+  "instruction": "Put the mushroom on the plate",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:01:09+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.996031746031746,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for and grasping mushroom",
+      "description": "reaching for and grasping mushroom"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.996031746031746,
+      "t_end_sec": 5.992063492063492,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "moving empty gripper toward plate",
+      "description": "moving empty gripper toward plate"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.992063492063492,
+      "t_end_sec": 6.071428571428571,
+      "control_step_start": 302,
+      "control_step_end": 305,
+      "policy_step_start": 50,
+      "policy_step_end": 50,
+      "phrase": "hovering over plate",
+      "description": "hovering over plate"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 1.8650793650793651,
+    "vlm_failure_onset_timestamp": "00:01.9",
+    "vlm_failure_onset_frame": 94,
+    "vlm_failure_onset_step": 94,
+    "vlm_coarse_failure_onset_seconds": 2.5,
+    "vlm_failure_window_start_seconds": 2.0,
+    "vlm_failure_window_end_seconds": 3.2,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot end-effector lowered to grasp the mushroom but failed to secure a firm grip, allowing the mushroom to slip out as the gripper closed.",
+    "vlm_failure_reason_summary": "failed grasp attempt; mushroom slipping from gripper",
+    "vlm_recovery_action": "Re-align the end-effector over the mushroom, close the fingers securely around it, and transfer it onto the green plate.",
+    "vlm_recovery_action_summary": "re-align gripper; grasp mushroom; place on plate",
+    "vlm_prevention_action": "Position the gripper precisely around the mushroom's center of mass before closing the fingers to ensure a secure grasp.",
+    "vlm_prevention_action_summary": "align gripper over mushroom; establish secure grasp",
+    "vlm_justification": "At 2.5 seconds, the robot attempts to grasp the mushroom, but the gripper closes improperly and fails to hold the object, resulting in a grasp failure.",
+    "vlm_temporal_justification": "At clip second 01:34 (94 seconds), the gripper closes on the mushroom but fails to secure a firm hold, causing the mushroom to slip out from between the fingers.",
+    "vlm_video_fps": 50.4,
+    "vlm_video_num_frames": 306,
+    "vlm_video_duration_seconds": 6.051587301587301,
+    "vlm_usage": {
+      "input_tokens": 19743,
+      "output_tokens": 348,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 2.5,\n \"failure_window_start_seconds\": 2.0,\n \"failure_window_end_seconds\": 3.2,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot end-effector lowered to grasp the mushroom but failed to secure a firm grip, allowing the mushroom to slip out as the gripper closed.\",\n \"failure_reason_summary\": \"failed grasp attempt; mushroom slipping from gripper\",\n \"recovery_action\": \"Re-align the end-effector over the mushroom, close the fingers securely around it, and transfer it onto the green plate.\",\n \"recovery_action_summary\": \"re-align gripper; grasp mushroom; place on plate\",\n \"prevention_action\": \"Position the gripper precisely around the mushroom's center of mass before closing the fingers to ensure a secure grasp.\",\n \"prevention_action_summary\": \"align gripper over mushroom; establish secure grasp\",\n \"justification\": \"At 2.5 seconds, the robot attempts to grasp the mushroom, but the gripper closes improperly and fails to hold the object, resulting in a grasp failure.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 94,
+    "failure_sim_step": 94,
+    "failure_policy_step": 15,
+    "failure_chunk_index": 4,
+    "first_post_failure_policy_step": 16,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-align the end-effector over the mushroom, close the fingers securely around it, and transfer it onto the green plate."
+  }
+}
\ No newline at end of file
diff --git a/mushroom/ep017/labels.npz b/mushroom/ep017/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..c6199a77652d34ca7a477e3df6291bdabea16988
--- /dev/null
+++ b/mushroom/ep017/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:af710517ef5fb7d33076a4961428907d698fa0e1a0944b3a9dbb262cf2a4b144
+size 3797
diff --git a/mushroom/ep018/example.md b/mushroom/ep018/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..750271b5f5e5f7f78ce9cf37760d24323709170f
--- /dev/null
+++ b/mushroom/ep018/example.md
@@ -0,0 +1,33 @@
+# t001__ep018__mushroom
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the mushroom on the plate
+- **success:** False
+- **control steps:** 315  (6.3s at 50 Hz)
+- **policy inferences:** 55  (replan_steps=6)
+- **video:** `mushroom/ep018/rollout.mp4`
+- **features:** `55` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–149 | 0–24 | lowering gripper toward mushroom |
+| 3.0–6.0 | 150–299 | 25–49 | closing gripper and lifting empty |
+| 6.0–6.3 | 300–314 | 50–52 | hovering arm above table |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 4.810379241516966 s (00:04.8)
+- **frame / control_step:** 241 (policy_step=40, chunk_index=1)
+- **coarse onset:** 2.5 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot lowered its end effector and closed its gripper near the mushroom, but failed to secure a grasp on it, lifting away empty.
+  Summary: failed grasp on mushroom; empty gripper
+- **recovery:** Re-align the end-effector over the mushroom, close the gripper securely around its stem/cap, and carry it onto the green plate.
+  Summary: re-align gripper; grasp mushroom; place on plate
+- **prevention:** Position the open gripper directly over the mushroom's center before closing the fingers to ensure a secure grip.
+  Summary: center gripper over mushroom; establish firm grasp
+- **justification:** Around 2.5 seconds, the robot closes its gripper around the mushroom but fails to grasp it securely. It then lifts its arm at 5.0 seconds without holding the mushroom, leaving the goal unfulfilled.
diff --git a/mushroom/ep018/labels.json b/mushroom/ep018/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..35b65e4c3c31e68f005e185f4e85320b9ef65458
--- /dev/null
+++ b/mushroom/ep018/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t001__ep018__mushroom",
+  "scene_variant": "",
+  "task_id": 1,
+  "task_file": "",
+  "instruction": "Put the mushroom on the plate",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:01:41+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9940119760479043,
+      "control_step_start": 0,
+      "control_step_end": 149,
+      "policy_step_start": 0,
+      "policy_step_end": 24,
+      "phrase": "lowering gripper toward mushroom",
+      "description": "lowering gripper toward mushroom"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9940119760479043,
+      "t_end_sec": 5.9880239520958085,
+      "control_step_start": 150,
+      "control_step_end": 299,
+      "policy_step_start": 25,
+      "policy_step_end": 49,
+      "phrase": "closing gripper and lifting empty",
+      "description": "closing gripper and lifting empty"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.9880239520958085,
+      "t_end_sec": 6.287425149700598,
+      "control_step_start": 300,
+      "control_step_end": 314,
+      "policy_step_start": 50,
+      "policy_step_end": 52,
+      "phrase": "hovering arm above table",
+      "description": "hovering arm above table"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 4.810379241516966,
+    "vlm_failure_onset_timestamp": "00:04.8",
+    "vlm_failure_onset_frame": 241,
+    "vlm_failure_onset_step": 241,
+    "vlm_coarse_failure_onset_seconds": 2.5,
+    "vlm_failure_window_start_seconds": 2.0,
+    "vlm_failure_window_end_seconds": 3.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot lowered its end effector and closed its gripper near the mushroom, but failed to secure a grasp on it, lifting away empty.",
+    "vlm_failure_reason_summary": "failed grasp on mushroom; empty gripper",
+    "vlm_recovery_action": "Re-align the end-effector over the mushroom, close the gripper securely around its stem/cap, and carry it onto the green plate.",
+    "vlm_recovery_action_summary": "re-align gripper; grasp mushroom; place on plate",
+    "vlm_prevention_action": "Position the open gripper directly over the mushroom's center before closing the fingers to ensure a secure grip.",
+    "vlm_prevention_action_summary": "center gripper over mushroom; establish firm grasp",
+    "vlm_justification": "Around 2.5 seconds, the robot closes its gripper around the mushroom but fails to grasp it securely. It then lifts its arm at 5.0 seconds without holding the mushroom, leaving the goal unfulfilled.",
+    "vlm_temporal_justification": "At 04:01 (241 seconds into the clip), the robot gripper fully closes its fingers without securing the mushroom, leaving the mushroom behind on the table as the arm begins to lift away empty.",
+    "vlm_video_fps": 50.1,
+    "vlm_video_num_frames": 315,
+    "vlm_video_duration_seconds": 6.267465069860279,
+    "vlm_usage": {
+      "input_tokens": 19615,
+      "output_tokens": 364,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 2.5,\n \"failure_window_start_seconds\": 2.0,\n \"failure_window_end_seconds\": 3.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot lowered its end effector and closed its gripper near the mushroom, but failed to secure a grasp on it, lifting away empty.\",\n \"failure_reason_summary\": \"failed grasp on mushroom; empty gripper\",\n \"recovery_action\": \"Re-align the end-effector over the mushroom, close the gripper securely around its stem/cap, and carry it onto the green plate.\",\n \"recovery_action_summary\": \"re-align gripper; grasp mushroom; place on plate\",\n \"prevention_action\": \"Position the open gripper directly over the mushroom's center before closing the fingers to ensure a secure grip.\",\n \"prevention_action_summary\": \"center gripper over mushroom; establish firm grasp\",\n \"justification\": \"Around 2.5 seconds, the robot closes its gripper around the mushroom but fails to grasp it securely. It then lifts its arm at 5.0 seconds without holding the mushroom, leaving the goal unfulfilled.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 241,
+    "failure_sim_step": 241,
+    "failure_policy_step": 40,
+    "failure_chunk_index": 1,
+    "first_post_failure_policy_step": 41,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-align the end-effector over the mushroom, close the gripper securely around its stem/cap, and carry it onto the green plate."
+  }
+}
\ No newline at end of file
diff --git a/mushroom/ep018/labels.npz b/mushroom/ep018/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..43a07b86b651c1f155851b3cb6fd60cd4fc90abc
--- /dev/null
+++ b/mushroom/ep018/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d95a827d46ce778fd193d683c08eee90591559604609d526c6dc4c4fa19cd358
+size 3772
diff --git a/mushroom/ep019/example.md b/mushroom/ep019/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..32dfea5244d400dcea15dc36173015aee16315b4
--- /dev/null
+++ b/mushroom/ep019/example.md
@@ -0,0 +1,21 @@
+# t001__ep019__mushroom
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the mushroom on the plate
+- **success:** True
+- **control steps:** 340  (6.8s at 50 Hz)
+- **policy inferences:** 59  (replan_steps=6)
+- **video:** `mushroom/ep019/rollout.mp4`
+- **features:** `59` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–149 | 0–24 | reaching for the mushroom |
+| 3.0–6.0 | 150–299 | 25–49 | placing mushroom on the plate |
+| 6.0–6.8 | 300–339 | 50–56 | retracting robot arm |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/mushroom/ep019/labels.json b/mushroom/ep019/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..365bb9b44a15737bba8ee8bee22aa056fc63330d
--- /dev/null
+++ b/mushroom/ep019/labels.json
@@ -0,0 +1,68 @@
+{
+  "rollout_id": "t001__ep019__mushroom",
+  "scene_variant": "",
+  "task_id": 1,
+  "task_file": "",
+  "instruction": "Put the mushroom on the plate",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:01:48+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0,
+      "control_step_start": 0,
+      "control_step_end": 149,
+      "policy_step_start": 0,
+      "policy_step_end": 24,
+      "phrase": "reaching for the mushroom",
+      "description": "reaching for the mushroom"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0,
+      "t_end_sec": 6.0,
+      "control_step_start": 150,
+      "control_step_end": 299,
+      "policy_step_start": 25,
+      "policy_step_end": 49,
+      "phrase": "placing mushroom on the plate",
+      "description": "placing mushroom on the plate"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.0,
+      "t_end_sec": 6.8,
+      "control_step_start": 300,
+      "control_step_end": 339,
+      "policy_step_start": 50,
+      "policy_step_end": 56,
+      "phrase": "retracting robot arm",
+      "description": "retracting robot arm"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/mushroom/ep019/labels.npz b/mushroom/ep019/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..eb40683ae289b514cae8560b7bc3a07faf80e294
--- /dev/null
+++ b/mushroom/ep019/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cb49f227941d6b76f7b701d41d25e05c3d9948fd46928b94aefddb363767a0a6
+size 3073
diff --git a/mushroom/ep020/example.md b/mushroom/ep020/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..450b101374e9074b81081db3411b0aed7410895b
--- /dev/null
+++ b/mushroom/ep020/example.md
@@ -0,0 +1,34 @@
+# t001__ep020__mushroom
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the mushroom on the plate
+- **success:** False
+- **control steps:** 568  (11.2s at 51 Hz)
+- **policy inferences:** 97  (replan_steps=6)
+- **video:** `mushroom/ep020/rollout.mp4`
+- **features:** `97` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching down toward mushroom |
+| 3.0–6.0 | 152–303 | 25–50 | failing grasp on mushroom |
+| 6.0–9.0 | 304–455 | 50–75 | hovering empty over green plate |
+| 9.0–11.2 | 456–567 | 76–94 | lowering empty gripper over plate |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 4.704724409448819 s (00:04.7)
+- **frame / control_step:** 239 (policy_step=39, chunk_index=5)
+- **coarse onset:** 4.0 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the mushroom, but failed to secure a firm hold, causing the mushroom to slip out and knock over onto the table while the arm moved away empty-handed.
+  Summary: failed grasp on mushroom; moving empty gripper
+- **recovery:** Re-align the gripper over the fallen mushroom, close the fingers firmly to pick it up, and place it on the green plate.
+  Summary: re-grasp mushroom; place on green plate
+- **prevention:** Center and align the gripper precisely over the mushroom cap before closing to ensure a secure grip before moving to the target.
+  Summary: align gripper properly; securely pick up mushroom
+- **justification:** At around 4.0 seconds, the robot closes its gripper on the mushroom, but fails to grasp it securely. As the arm lifts and moves toward the plate, the mushroom slips out and remains on the table.
diff --git a/mushroom/ep020/labels.json b/mushroom/ep020/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..201808a85ec1727e2c523badce45759bcbce1dda
--- /dev/null
+++ b/mushroom/ep020/labels.json
@@ -0,0 +1,109 @@
+{
+  "rollout_id": "t001__ep020__mushroom",
+  "scene_variant": "",
+  "task_id": 1,
+  "task_file": "",
+  "instruction": "Put the mushroom on the plate",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:02:21+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9921259842519685,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching down toward mushroom",
+      "description": "reaching down toward mushroom"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9921259842519685,
+      "t_end_sec": 5.984251968503937,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "failing grasp on mushroom",
+      "description": "failing grasp on mushroom"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.984251968503937,
+      "t_end_sec": 8.976377952755906,
+      "control_step_start": 304,
+      "control_step_end": 455,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "hovering empty over green plate",
+      "description": "hovering empty over green plate"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.976377952755906,
+      "t_end_sec": 11.181102362204726,
+      "control_step_start": 456,
+      "control_step_end": 567,
+      "policy_step_start": 76,
+      "policy_step_end": 94,
+      "phrase": "lowering empty gripper over plate",
+      "description": "lowering empty gripper over plate"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 4.704724409448819,
+    "vlm_failure_onset_timestamp": "00:04.7",
+    "vlm_failure_onset_frame": 239,
+    "vlm_failure_onset_step": 239,
+    "vlm_coarse_failure_onset_seconds": 4.0,
+    "vlm_failure_window_start_seconds": 3.5,
+    "vlm_failure_window_end_seconds": 4.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the mushroom, but failed to secure a firm hold, causing the mushroom to slip out and knock over onto the table while the arm moved away empty-handed.",
+    "vlm_failure_reason_summary": "failed grasp on mushroom; moving empty gripper",
+    "vlm_recovery_action": "Re-align the gripper over the fallen mushroom, close the fingers firmly to pick it up, and place it on the green plate.",
+    "vlm_recovery_action_summary": "re-grasp mushroom; place on green plate",
+    "vlm_prevention_action": "Center and align the gripper precisely over the mushroom cap before closing to ensure a secure grip before moving to the target.",
+    "vlm_prevention_action_summary": "align gripper properly; securely pick up mushroom",
+    "vlm_justification": "At around 4.0 seconds, the robot closes its gripper on the mushroom, but fails to grasp it securely. As the arm lifts and moves toward the plate, the mushroom slips out and remains on the table.",
+    "vlm_temporal_justification": "At 03:08, the robot's gripper attempts to close on the mushroom, but fails to secure a proper grip. The fingers push against the mushroom, causing it to tilt and slip out of the gripper onto its side on the table.",
+    "vlm_video_fps": 50.8,
+    "vlm_video_num_frames": 568,
+    "vlm_video_duration_seconds": 11.161417322834646,
+    "vlm_usage": {
+      "input_tokens": 21772,
+      "output_tokens": 381,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 4.0,\n \"failure_window_start_seconds\": 3.5,\n \"failure_window_end_seconds\": 4.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the mushroom, but failed to secure a firm hold, causing the mushroom to slip out and knock over onto the table while the arm moved away empty-handed.\",\n \"failure_reason_summary\": \"failed grasp on mushroom; moving empty gripper\",\n \"recovery_action\": \"Re-align the gripper over the fallen mushroom, close the fingers firmly to pick it up, and place it on the green plate.\",\n \"recovery_action_summary\": \"re-grasp mushroom; place on green plate\",\n \"prevention_action\": \"Center and align the gripper precisely over the mushroom cap before closing to ensure a secure grip before moving to the target.\",\n \"prevention_action_summary\": \"align gripper properly; securely pick up mushroom\",\n \"justification\": \"At around 4.0 seconds, the robot closes its gripper on the mushroom, but fails to grasp it securely. As the arm lifts and moves toward the plate, the mushroom slips out and remains on the table.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 239,
+    "failure_sim_step": 239,
+    "failure_policy_step": 39,
+    "failure_chunk_index": 5,
+    "first_post_failure_policy_step": 40,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-align the gripper over the fallen mushroom, close the fingers firmly to pick it up, and place it on the green plate."
+  }
+}
\ No newline at end of file
diff --git a/mushroom/ep020/labels.npz b/mushroom/ep020/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..cf4f206a85ce465082035425b2aec42fb14719f8
--- /dev/null
+++ b/mushroom/ep020/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9bb6553a6f8e5cd957f28e39ccfbc74e673ca925c8dce830486d6609f5ccc8b1
+size 3850
diff --git a/mushroom/ep021/example.md b/mushroom/ep021/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..5114096596c257d9bf5d49b6dc90eee8afc72054
--- /dev/null
+++ b/mushroom/ep021/example.md
@@ -0,0 +1,33 @@
+# t001__ep021__mushroom
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the mushroom on the plate
+- **success:** False
+- **control steps:** 306  (6.0s at 51 Hz)
+- **policy inferences:** 54  (replan_steps=6)
+- **video:** `mushroom/ep021/rollout.mp4`
+- **features:** `54` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | grasping and lifting mushroom |
+| 3.0–6.0 | 152–303 | 25–50 | releasing mushroom beside plate |
+| 6.0–6.0 | 304–305 | 50–50 | idle above table |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 4.516765285996055 s (00:04.5)
+- **frame / control_step:** 229 (policy_step=38, chunk_index=1)
+- **coarse onset:** 4.0 s (refined=True)
+- **mode:** placement_or_insertion_failure
+- **reason:** The robot moved the mushroom toward the green plate but released its grip prematurely near the edge, causing the mushroom to drop onto the table surface off-target.
+  Summary: releasing mushroom off-target; mushroom falling onto table
+- **recovery:** Re-grasp the mushroom from the table surface, carry it fully over the center of the green plate, and release it onto the plate.
+  Summary: re-grasp mushroom; place centered on green plate
+- **prevention:** Center the gripper fully over the green plate before opening the fingers to ensure the mushroom settles properly on the plate.
+  Summary: center mushroom over green plate; delay release until aligned
+- **justification:** The robot successfully grasped and lifted the mushroom by around 3.0 seconds, but at 4.0 seconds opened its gripper while positioned over the edge of the plate, dropping the mushroom onto the tabletop instead of the plate.
diff --git a/mushroom/ep021/labels.json b/mushroom/ep021/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..fa81aa7fa2882260a0f47c4eefa4192571c07206
--- /dev/null
+++ b/mushroom/ep021/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t001__ep021__mushroom",
+  "scene_variant": "",
+  "task_id": 1,
+  "task_file": "",
+  "instruction": "Put the mushroom on the plate",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:02:57+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9980276134122286,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "grasping and lifting mushroom",
+      "description": "grasping and lifting mushroom"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9980276134122286,
+      "t_end_sec": 5.996055226824457,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "releasing mushroom beside plate",
+      "description": "releasing mushroom beside plate"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.996055226824457,
+      "t_end_sec": 6.035502958579881,
+      "control_step_start": 304,
+      "control_step_end": 305,
+      "policy_step_start": 50,
+      "policy_step_end": 50,
+      "phrase": "idle above table",
+      "description": "idle above table"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "placement_or_insertion_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 4.516765285996055,
+    "vlm_failure_onset_timestamp": "00:04.5",
+    "vlm_failure_onset_frame": 229,
+    "vlm_failure_onset_step": 229,
+    "vlm_coarse_failure_onset_seconds": 4.0,
+    "vlm_failure_window_start_seconds": 3.5,
+    "vlm_failure_window_end_seconds": 4.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot moved the mushroom toward the green plate but released its grip prematurely near the edge, causing the mushroom to drop onto the table surface off-target.",
+    "vlm_failure_reason_summary": "releasing mushroom off-target; mushroom falling onto table",
+    "vlm_recovery_action": "Re-grasp the mushroom from the table surface, carry it fully over the center of the green plate, and release it onto the plate.",
+    "vlm_recovery_action_summary": "re-grasp mushroom; place centered on green plate",
+    "vlm_prevention_action": "Center the gripper fully over the green plate before opening the fingers to ensure the mushroom settles properly on the plate.",
+    "vlm_prevention_action_summary": "center mushroom over green plate; delay release until aligned",
+    "vlm_justification": "The robot successfully grasped and lifted the mushroom by around 3.0 seconds, but at 4.0 seconds opened its gripper while positioned over the edge of the plate, dropping the mushroom onto the tabletop instead of the plate.",
+    "vlm_temporal_justification": "At second 178 (02:58 in the clip), the robot gripper prematurely opens its fingers while positioned near the edge of the plate, causing the mushroom to drop onto the table surface off-target.",
+    "vlm_video_fps": 50.7,
+    "vlm_video_num_frames": 306,
+    "vlm_video_duration_seconds": 6.01577909270217,
+    "vlm_usage": {
+      "input_tokens": 18303,
+      "output_tokens": 380,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"placement_or_insertion_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 4.0,\n \"failure_window_start_seconds\": 3.5,\n \"failure_window_end_seconds\": 4.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot moved the mushroom toward the green plate but released its grip prematurely near the edge, causing the mushroom to drop onto the table surface off-target.\",\n \"failure_reason_summary\": \"releasing mushroom off-target; mushroom falling onto table\",\n \"recovery_action\": \"Re-grasp the mushroom from the table surface, carry it fully over the center of the green plate, and release it onto the plate.\",\n \"recovery_action_summary\": \"re-grasp mushroom; place centered on green plate\",\n \"prevention_action\": \"Center the gripper fully over the green plate before opening the fingers to ensure the mushroom settles properly on the plate.\",\n \"prevention_action_summary\": \"center mushroom over green plate; delay release until aligned\",\n \"justification\": \"The robot successfully grasped and lifted the mushroom by around 3.0 seconds, but at 4.0 seconds opened its gripper while positioned over the edge of the plate, dropping the mushroom onto the tabletop instead of the plate.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 229,
+    "failure_sim_step": 229,
+    "failure_policy_step": 38,
+    "failure_chunk_index": 1,
+    "first_post_failure_policy_step": 39,
+    "failure_type": "placement_or_insertion_failure",
+    "correction_action": "Re-grasp the mushroom from the table surface, carry it fully over the center of the green plate, and release it onto the plate."
+  }
+}
\ No newline at end of file
diff --git a/mushroom/ep021/labels.npz b/mushroom/ep021/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..7077fa8e9f9c46bc7a607ab690e38ee0106ce1d5
--- /dev/null
+++ b/mushroom/ep021/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3e25726ef3bf2d426e5872bcc7116ebd1f066d1700adb92aaed23eac229f5c28
+size 3822
diff --git a/mushroom/ep022/example.md b/mushroom/ep022/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..2efc053652a7982b440a43be7308678b4e348fb5
--- /dev/null
+++ b/mushroom/ep022/example.md
@@ -0,0 +1,21 @@
+# t001__ep022__mushroom
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the mushroom on the plate
+- **success:** True
+- **control steps:** 403  (8.1s at 50 Hz)
+- **policy inferences:** 69  (replan_steps=6)
+- **video:** `mushroom/ep022/rollout.mp4`
+- **features:** `69` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–149 | 0–24 | reaching for mushroom |
+| 3.0–6.0 | 150–299 | 25–49 | moving mushroom to plate |
+| 6.0–8.1 | 300–402 | 50–67 | releasing mushroom and retracting |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/mushroom/ep022/labels.json b/mushroom/ep022/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..bfe7b882c9bda44e336ddc74cf627fa474ce57d6
--- /dev/null
+++ b/mushroom/ep022/labels.json
@@ -0,0 +1,68 @@
+{
+  "rollout_id": "t001__ep022__mushroom",
+  "scene_variant": "",
+  "task_id": 1,
+  "task_file": "",
+  "instruction": "Put the mushroom on the plate",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:03:05+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.006012024048096,
+      "control_step_start": 0,
+      "control_step_end": 149,
+      "policy_step_start": 0,
+      "policy_step_end": 24,
+      "phrase": "reaching for mushroom",
+      "description": "reaching for mushroom"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.006012024048096,
+      "t_end_sec": 6.012024048096192,
+      "control_step_start": 150,
+      "control_step_end": 299,
+      "policy_step_start": 25,
+      "policy_step_end": 49,
+      "phrase": "moving mushroom to plate",
+      "description": "moving mushroom to plate"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.012024048096192,
+      "t_end_sec": 8.076152304609218,
+      "control_step_start": 300,
+      "control_step_end": 402,
+      "policy_step_start": 50,
+      "policy_step_end": 67,
+      "phrase": "releasing mushroom and retracting",
+      "description": "releasing mushroom and retracting"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/mushroom/ep022/labels.npz b/mushroom/ep022/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..f2f0965af3f98eb801fed9a125e629f577187d45
--- /dev/null
+++ b/mushroom/ep022/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5c0bf591f3b0c5306bd7d1344bdaedf95de3bb89e0682ad0a5799d5b55021514
+size 3082
diff --git a/mushroom/ep023/example.md b/mushroom/ep023/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..c15fe2c0544134a0e24571f60ae19cf01e9fadb5
--- /dev/null
+++ b/mushroom/ep023/example.md
@@ -0,0 +1,34 @@
+# t001__ep023__mushroom
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the mushroom on the plate
+- **success:** False
+- **control steps:** 520  (10.1s at 51 Hz)
+- **policy inferences:** 89  (replan_steps=6)
+- **video:** `mushroom/ep023/rollout.mp4`
+- **features:** `89` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–153 | 0–25 | reaching toward mushroom |
+| 3.0–6.0 | 154–307 | 25–51 | closing gripper on mushroom |
+| 6.0–9.0 | 308–461 | 51–76 | lifting empty gripper |
+| 9.0–10.1 | 462–519 | 77–86 | hovering over workspace |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 4.522417153996102 s (00:04.5)
+- **frame / control_step:** 232 (policy_step=38, chunk_index=4)
+- **coarse onset:** 3.5 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot aligned over the mushroom and closed its gripper, but failed to establish a secure hold, causing the mushroom to remain on the table as the arm lifted up.
+  Summary: failed grasp attempt; mushroom left on table
+- **recovery:** Lower the end effector back over the mushroom, align the fingers securely around the object, close the gripper to grasp it, and move it onto the green plate.
+  Summary: realign over mushroom; grasp mushroom securely; place on plate
+- **prevention:** Position the gripper lower and center it properly over the mushroom before closing the fingers to guarantee a secure grasp.
+  Summary: center gripper lower; ensure firm grasp; pick and place object
+- **justification:** Between 3.0s and 4.0s, the robot attempts to grasp the mushroom. At around 3.5s, the fingers close without securing the object, and by 5.0s the arm rises with an empty gripper, leaving the mushroom unmoved.
diff --git a/mushroom/ep023/labels.json b/mushroom/ep023/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..9ca9caaea5f4deacd8639b8d34fedbe9cae204d9
--- /dev/null
+++ b/mushroom/ep023/labels.json
@@ -0,0 +1,109 @@
+{
+  "rollout_id": "t001__ep023__mushroom",
+  "scene_variant": "",
+  "task_id": 1,
+  "task_file": "",
+  "instruction": "Put the mushroom on the plate",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:03:37+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0019493177387915,
+      "control_step_start": 0,
+      "control_step_end": 153,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching toward mushroom",
+      "description": "reaching toward mushroom"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0019493177387915,
+      "t_end_sec": 6.003898635477583,
+      "control_step_start": 154,
+      "control_step_end": 307,
+      "policy_step_start": 25,
+      "policy_step_end": 51,
+      "phrase": "closing gripper on mushroom",
+      "description": "closing gripper on mushroom"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.003898635477583,
+      "t_end_sec": 9.005847953216374,
+      "control_step_start": 308,
+      "control_step_end": 461,
+      "policy_step_start": 51,
+      "policy_step_end": 76,
+      "phrase": "lifting empty gripper",
+      "description": "lifting empty gripper"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.005847953216374,
+      "t_end_sec": 10.1364522417154,
+      "control_step_start": 462,
+      "control_step_end": 519,
+      "policy_step_start": 77,
+      "policy_step_end": 86,
+      "phrase": "hovering over workspace",
+      "description": "hovering over workspace"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 4.522417153996102,
+    "vlm_failure_onset_timestamp": "00:04.5",
+    "vlm_failure_onset_frame": 232,
+    "vlm_failure_onset_step": 232,
+    "vlm_coarse_failure_onset_seconds": 3.5,
+    "vlm_failure_window_start_seconds": 3.0,
+    "vlm_failure_window_end_seconds": 5.0,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot aligned over the mushroom and closed its gripper, but failed to establish a secure hold, causing the mushroom to remain on the table as the arm lifted up.",
+    "vlm_failure_reason_summary": "failed grasp attempt; mushroom left on table",
+    "vlm_recovery_action": "Lower the end effector back over the mushroom, align the fingers securely around the object, close the gripper to grasp it, and move it onto the green plate.",
+    "vlm_recovery_action_summary": "realign over mushroom; grasp mushroom securely; place on plate",
+    "vlm_prevention_action": "Position the gripper lower and center it properly over the mushroom before closing the fingers to guarantee a secure grasp.",
+    "vlm_prevention_action_summary": "center gripper lower; ensure firm grasp; pick and place object",
+    "vlm_justification": "Between 3.0s and 4.0s, the robot attempts to grasp the mushroom. At around 3.5s, the fingers close without securing the object, and by 5.0s the arm rises with an empty gripper, leaving the mushroom unmoved.",
+    "vlm_temporal_justification": "At timestamp 03:26 (second 206), the robot arm begins moving upward after closing its gripper, but fails to pick up the mushroom, leaving it behind on the table.",
+    "vlm_video_fps": 51.3,
+    "vlm_video_num_frames": 520,
+    "vlm_video_duration_seconds": 10.116959064327485,
+    "vlm_usage": {
+      "input_tokens": 21957,
+      "output_tokens": 390,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 3.5,\n \"failure_window_start_seconds\": 3.0,\n \"failure_window_end_seconds\": 5.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot aligned over the mushroom and closed its gripper, but failed to establish a secure hold, causing the mushroom to remain on the table as the arm lifted up.\",\n \"failure_reason_summary\": \"failed grasp attempt; mushroom left on table\",\n \"recovery_action\": \"Lower the end effector back over the mushroom, align the fingers securely around the object, close the gripper to grasp it, and move it onto the green plate.\",\n \"recovery_action_summary\": \"realign over mushroom; grasp mushroom securely; place on plate\",\n \"prevention_action\": \"Position the gripper lower and center it properly over the mushroom before closing the fingers to guarantee a secure grasp.\",\n \"prevention_action_summary\": \"center gripper lower; ensure firm grasp; pick and place object\",\n \"justification\": \"Between 3.0s and 4.0s, the robot attempts to grasp the mushroom. At around 3.5s, the fingers close without securing the object, and by 5.0s the arm rises with an empty gripper, leaving the mushroom unmoved.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 232,
+    "failure_sim_step": 232,
+    "failure_policy_step": 38,
+    "failure_chunk_index": 4,
+    "first_post_failure_policy_step": 39,
+    "failure_type": "grasp_failure",
+    "correction_action": "Lower the end effector back over the mushroom, align the fingers securely around the object, close the gripper to grasp it, and move it onto the green plate."
+  }
+}
\ No newline at end of file
diff --git a/mushroom/ep023/labels.npz b/mushroom/ep023/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..6233c43e6be106a887e614c44496d5491f7988dd
--- /dev/null
+++ b/mushroom/ep023/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c2e41bd64b8ce1d77761463df2efbbd38750fa30d2fa7c4ae0ab3fa5eff1927f
+size 3898
diff --git a/mushroom/ep024/example.md b/mushroom/ep024/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..3cc4fcb8c73a4d9aa105dd00363edcbda107a6ab
--- /dev/null
+++ b/mushroom/ep024/example.md
@@ -0,0 +1,33 @@
+# t001__ep024__mushroom
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the mushroom on the plate
+- **success:** False
+- **control steps:** 336  (6.7s at 50 Hz)
+- **policy inferences:** 59  (replan_steps=6)
+- **video:** `mushroom/ep024/rollout.mp4`
+- **features:** `59` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–149 | 0–24 | descending and grasping mushroom |
+| 3.0–6.0 | 150–299 | 25–49 | lifting and dropping mushroom |
+| 6.0–6.7 | 300–335 | 50–55 | hovering idle above table |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 5.070140280561122 s (00:05.1)
+- **frame / control_step:** 253 (policy_step=42, chunk_index=1)
+- **coarse onset:** 4.8 s (refined=True)
+- **mode:** object_displacement
+- **reason:** The robot successfully grasped the mushroom but prematurely opened its gripper while moving towards the plate, dropping the mushroom onto the table.
+  Summary: premature release; dropping mushroom onto table
+- **recovery:** Re-grasp the mushroom from the tabletop, carry it over to the green plate, and lower it onto the plate before releasing.
+  Summary: re-grasp mushroom; place on plate
+- **prevention:** Maintain a firm grip on the mushroom during transport and only release the gripper once positioned directly above the green plate.
+  Summary: maintain grasp; release only over plate
+- **justification:** Between timestamps 00:04 and 00:05 (around 4.8s), the robot opens its gripper prematurely while transporting the mushroom, causing the mushroom to fall onto the table surface rather than being safely transferred to the green plate.
diff --git a/mushroom/ep024/labels.json b/mushroom/ep024/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..ce0d1545f49fa7b438903157f8400c63c36a1ef8
--- /dev/null
+++ b/mushroom/ep024/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t001__ep024__mushroom",
+  "scene_variant": "",
+  "task_id": 1,
+  "task_file": "",
+  "instruction": "Put the mushroom on the plate",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:04:06+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.006012024048096,
+      "control_step_start": 0,
+      "control_step_end": 149,
+      "policy_step_start": 0,
+      "policy_step_end": 24,
+      "phrase": "descending and grasping mushroom",
+      "description": "descending and grasping mushroom"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.006012024048096,
+      "t_end_sec": 6.012024048096192,
+      "control_step_start": 150,
+      "control_step_end": 299,
+      "policy_step_start": 25,
+      "policy_step_end": 49,
+      "phrase": "lifting and dropping mushroom",
+      "description": "lifting and dropping mushroom"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.012024048096192,
+      "t_end_sec": 6.733466933867736,
+      "control_step_start": 300,
+      "control_step_end": 335,
+      "policy_step_start": 50,
+      "policy_step_end": 55,
+      "phrase": "hovering idle above table",
+      "description": "hovering idle above table"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "object_displacement",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 5.070140280561122,
+    "vlm_failure_onset_timestamp": "00:05.1",
+    "vlm_failure_onset_frame": 253,
+    "vlm_failure_onset_step": 253,
+    "vlm_coarse_failure_onset_seconds": 4.8,
+    "vlm_failure_window_start_seconds": 4.3,
+    "vlm_failure_window_end_seconds": 5.3,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot successfully grasped the mushroom but prematurely opened its gripper while moving towards the plate, dropping the mushroom onto the table.",
+    "vlm_failure_reason_summary": "premature release; dropping mushroom onto table",
+    "vlm_recovery_action": "Re-grasp the mushroom from the tabletop, carry it over to the green plate, and lower it onto the plate before releasing.",
+    "vlm_recovery_action_summary": "re-grasp mushroom; place on plate",
+    "vlm_prevention_action": "Maintain a firm grip on the mushroom during transport and only release the gripper once positioned directly above the green plate.",
+    "vlm_prevention_action_summary": "maintain grasp; release only over plate",
+    "vlm_justification": "Between timestamps 00:04 and 00:05 (around 4.8s), the robot opens its gripper prematurely while transporting the mushroom, causing the mushroom to fall onto the table surface rather than being safely transferred to the green plate.",
+    "vlm_temporal_justification": "At 02:43 (163s), the robot's gripper fingers begin opening prematurely while moving towards the green plate, releasing the mushroom and causing it to drop onto the table surface.",
+    "vlm_video_fps": 49.9,
+    "vlm_video_num_frames": 336,
+    "vlm_video_duration_seconds": 6.713426853707415,
+    "vlm_usage": {
+      "input_tokens": 17788,
+      "output_tokens": 363,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"object_displacement\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 4.8,\n \"failure_window_start_seconds\": 4.3,\n \"failure_window_end_seconds\": 5.3,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot successfully grasped the mushroom but prematurely opened its gripper while moving towards the plate, dropping the mushroom onto the table.\",\n \"failure_reason_summary\": \"premature release; dropping mushroom onto table\",\n \"recovery_action\": \"Re-grasp the mushroom from the tabletop, carry it over to the green plate, and lower it onto the plate before releasing.\",\n \"recovery_action_summary\": \"re-grasp mushroom; place on plate\",\n \"prevention_action\": \"Maintain a firm grip on the mushroom during transport and only release the gripper once positioned directly above the green plate.\",\n \"prevention_action_summary\": \"maintain grasp; release only over plate\",\n \"justification\": \"Between timestamps 00:04 and 00:05 (around 4.8s), the robot opens its gripper prematurely while transporting the mushroom, causing the mushroom to fall onto the table surface rather than being safely transferred to the green plate.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 253,
+    "failure_sim_step": 253,
+    "failure_policy_step": 42,
+    "failure_chunk_index": 1,
+    "first_post_failure_policy_step": 43,
+    "failure_type": "object_displacement",
+    "correction_action": "Re-grasp the mushroom from the tabletop, carry it over to the green plate, and lower it onto the plate before releasing."
+  }
+}
\ No newline at end of file
diff --git a/mushroom/ep024/labels.npz b/mushroom/ep024/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..0d4dd1dd3d41760e472696b274eef02c48f552d8
--- /dev/null
+++ b/mushroom/ep024/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:75b873fcf8822d9b190a1ed3eb538dc973bb9c164095d7f04619c69538ccfe8a
+size 3758
diff --git a/mushroom/ep025/example.md b/mushroom/ep025/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..6ac4282bc148dffff44fab4b0f899c62ba9b06cf
--- /dev/null
+++ b/mushroom/ep025/example.md
@@ -0,0 +1,33 @@
+# t001__ep025__mushroom
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the mushroom on the plate
+- **success:** False
+- **control steps:** 370  (7.3s at 51 Hz)
+- **policy inferences:** 65  (replan_steps=5)
+- **video:** `mushroom/ep025/rollout.mp4`
+- **features:** `65` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–30 | lowering arm toward mushroom |
+| 3.0–6.0 | 152–303 | 30–60 | attempting grasp and lifting arm |
+| 6.0–7.3 | 304–369 | 60–73 | moving empty gripper over plate |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 5.936883629191321 s (00:05.9)
+- **frame / control_step:** 301 (policy_step=60, chunk_index=1)
+- **coarse onset:** 4.5 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the mushroom, but failed to secure a proper grip before moving its arm away towards the plate, leaving the mushroom on the table.
+  Summary: failed grasp on mushroom; moving away empty-handed
+- **recovery:** Reposition the gripper directly over the mushroom, execute a firm grasp, lift the mushroom, and place it onto the green plate.
+  Summary: re-grasp mushroom; place on plate
+- **prevention:** Align the end-effector accurately over the mushroom and ensure a secure grasp is established before beginning the transfer motion toward the plate.
+  Summary: align gripper precisely; secure firm grasp before transport
+- **justification:** Around 4.5 seconds (between 00:04 and 00:05), the gripper closes on the mushroom but fails to acquire it securely. The robot arm then lifts and moves over the green plate without carrying the mushroom, leading to task failure.
diff --git a/mushroom/ep025/labels.json b/mushroom/ep025/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..b7e87411b90180b047755d33ab8a589e3d537587
--- /dev/null
+++ b/mushroom/ep025/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t001__ep025__mushroom",
+  "scene_variant": "",
+  "task_id": 1,
+  "task_file": "",
+  "instruction": "Put the mushroom on the plate",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:04:37+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9980276134122286,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 30,
+      "phrase": "lowering arm toward mushroom",
+      "description": "lowering arm toward mushroom"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9980276134122286,
+      "t_end_sec": 5.996055226824457,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 30,
+      "policy_step_end": 60,
+      "phrase": "attempting grasp and lifting arm",
+      "description": "attempting grasp and lifting arm"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.996055226824457,
+      "t_end_sec": 7.297830374753452,
+      "control_step_start": 304,
+      "control_step_end": 369,
+      "policy_step_start": 60,
+      "policy_step_end": 73,
+      "phrase": "moving empty gripper over plate",
+      "description": "moving empty gripper over plate"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 5.936883629191321,
+    "vlm_failure_onset_timestamp": "00:05.9",
+    "vlm_failure_onset_frame": 301,
+    "vlm_failure_onset_step": 301,
+    "vlm_coarse_failure_onset_seconds": 4.5,
+    "vlm_failure_window_start_seconds": 4.0,
+    "vlm_failure_window_end_seconds": 5.0,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the mushroom, but failed to secure a proper grip before moving its arm away towards the plate, leaving the mushroom on the table.",
+    "vlm_failure_reason_summary": "failed grasp on mushroom; moving away empty-handed",
+    "vlm_recovery_action": "Reposition the gripper directly over the mushroom, execute a firm grasp, lift the mushroom, and place it onto the green plate.",
+    "vlm_recovery_action_summary": "re-grasp mushroom; place on plate",
+    "vlm_prevention_action": "Align the end-effector accurately over the mushroom and ensure a secure grasp is established before beginning the transfer motion toward the plate.",
+    "vlm_prevention_action_summary": "align gripper precisely; secure firm grasp before transport",
+    "vlm_justification": "Around 4.5 seconds (between 00:04 and 00:05), the gripper closes on the mushroom but fails to acquire it securely. The robot arm then lifts and moves over the green plate without carrying the mushroom, leading to task failure.",
+    "vlm_temporal_justification": "At clip second 225 (03:45), the robot finishes closing its gripper without securing the mushroom and begins moving its arm upward and toward the green plate, leaving the mushroom behind on the table.",
+    "vlm_video_fps": 50.7,
+    "vlm_video_num_frames": 370,
+    "vlm_video_duration_seconds": 7.278106508875739,
+    "vlm_usage": {
+      "input_tokens": 20819,
+      "output_tokens": 382,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 4.5,\n \"failure_window_start_seconds\": 4.0,\n \"failure_window_end_seconds\": 5.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the mushroom, but failed to secure a proper grip before moving its arm away towards the plate, leaving the mushroom on the table.\",\n \"failure_reason_summary\": \"failed grasp on mushroom; moving away empty-handed\",\n \"recovery_action\": \"Reposition the gripper directly over the mushroom, execute a firm grasp, lift the mushroom, and place it onto the green plate.\",\n \"recovery_action_summary\": \"re-grasp mushroom; place on plate\",\n \"prevention_action\": \"Align the end-effector accurately over the mushroom and ensure a secure grasp is established before beginning the transfer motion toward the plate.\",\n \"prevention_action_summary\": \"align gripper precisely; secure firm grasp before transport\",\n \"justification\": \"Around 4.5 seconds (between 00:04 and 00:05), the gripper closes on the mushroom but fails to acquire it securely. The robot arm then lifts and moves over the green plate without carrying the mushroom, leading to task failure.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 301,
+    "failure_sim_step": 301,
+    "failure_policy_step": 60,
+    "failure_chunk_index": 1,
+    "first_post_failure_policy_step": 61,
+    "failure_type": "grasp_failure",
+    "correction_action": "Reposition the gripper directly over the mushroom, execute a firm grasp, lift the mushroom, and place it onto the green plate."
+  }
+}
\ No newline at end of file
diff --git a/mushroom/ep025/labels.npz b/mushroom/ep025/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..4a66b02e10ce71d199de9242026a42c6ad2de194
--- /dev/null
+++ b/mushroom/ep025/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ec26990d4bf530644cd55f482942f701d3a75e0498d82e4ccd2056a10cdc4755
+size 3832
diff --git a/mushroom/ep026/example.md b/mushroom/ep026/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..86140ba1e2b92771f81a3254fdc238edcc1ce3bd
--- /dev/null
+++ b/mushroom/ep026/example.md
@@ -0,0 +1,33 @@
+# t001__ep026__mushroom
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the mushroom on the plate
+- **success:** False
+- **control steps:** 360  (7.2s at 50 Hz)
+- **policy inferences:** 62  (replan_steps=6)
+- **video:** `mushroom/ep026/rollout.mp4`
+- **features:** `62` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | reaching down toward mushroom |
+| 3.0–6.0 | 151–301 | 25–50 | closing gripper and lifting empty |
+| 6.0–7.2 | 302–359 | 50–59 | hovering empty over green plate |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 5.526838966202783 s (00:05.5)
+- **frame / control_step:** 278 (policy_step=46, chunk_index=2)
+- **coarse onset:** 4.5 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the mushroom on the table, but its gripper failed to secure a grip, leaving the mushroom behind while moving empty-handed toward the plate.
+  Summary: failed grasp on mushroom; moving away empty-handed
+- **recovery:** Re-align the gripper over the mushroom, close the fingers firmly around it to secure it, then lift and place it onto the green plate.
+  Summary: re-align gripper; grasp mushroom; place on plate
+- **prevention:** Lower the end-effector further and center the fingers precisely over the mushroom body before closing to achieve a firm grasp.
+  Summary: center gripper; descend fully; grasp mushroom securely
+- **justification:** At around 4.5 seconds, the robot attempts to grasp the mushroom, but the fingers fail to clamp securely around it. The robot then lifts up and moves over the green plate empty-handed, leaving the task incomplete.
diff --git a/mushroom/ep026/labels.json b/mushroom/ep026/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..998b540a1abdd0ae22a7d4cbc48aa6f4f188a4b4
--- /dev/null
+++ b/mushroom/ep026/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t001__ep026__mushroom",
+  "scene_variant": "",
+  "task_id": 1,
+  "task_file": "",
+  "instruction": "Put the mushroom on the plate",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:05:10+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.001988071570577,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching down toward mushroom",
+      "description": "reaching down toward mushroom"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.001988071570577,
+      "t_end_sec": 6.003976143141154,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "closing gripper and lifting empty",
+      "description": "closing gripper and lifting empty"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.003976143141154,
+      "t_end_sec": 7.1570576540755475,
+      "control_step_start": 302,
+      "control_step_end": 359,
+      "policy_step_start": 50,
+      "policy_step_end": 59,
+      "phrase": "hovering empty over green plate",
+      "description": "hovering empty over green plate"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 5.526838966202783,
+    "vlm_failure_onset_timestamp": "00:05.5",
+    "vlm_failure_onset_frame": 278,
+    "vlm_failure_onset_step": 278,
+    "vlm_coarse_failure_onset_seconds": 4.5,
+    "vlm_failure_window_start_seconds": 3.5,
+    "vlm_failure_window_end_seconds": 5.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the mushroom on the table, but its gripper failed to secure a grip, leaving the mushroom behind while moving empty-handed toward the plate.",
+    "vlm_failure_reason_summary": "failed grasp on mushroom; moving away empty-handed",
+    "vlm_recovery_action": "Re-align the gripper over the mushroom, close the fingers firmly around it to secure it, then lift and place it onto the green plate.",
+    "vlm_recovery_action_summary": "re-align gripper; grasp mushroom; place on plate",
+    "vlm_prevention_action": "Lower the end-effector further and center the fingers precisely over the mushroom body before closing to achieve a firm grasp.",
+    "vlm_prevention_action_summary": "center gripper; descend fully; grasp mushroom securely",
+    "vlm_justification": "At around 4.5 seconds, the robot attempts to grasp the mushroom, but the fingers fail to clamp securely around it. The robot then lifts up and moves over the green plate empty-handed, leaving the task incomplete.",
+    "vlm_temporal_justification": "At clip second 203 (03:23), the robot arm begins lifting upward empty-handed after attempting to close its gripper, leaving the mushroom behind on the table.",
+    "vlm_video_fps": 50.3,
+    "vlm_video_num_frames": 360,
+    "vlm_video_duration_seconds": 7.137176938369782,
+    "vlm_usage": {
+      "input_tokens": 20254,
+      "output_tokens": 372,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 4.5,\n \"failure_window_start_seconds\": 3.5,\n \"failure_window_end_seconds\": 5.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the mushroom on the table, but its gripper failed to secure a grip, leaving the mushroom behind while moving empty-handed toward the plate.\",\n \"failure_reason_summary\": \"failed grasp on mushroom; moving away empty-handed\",\n \"recovery_action\": \"Re-align the gripper over the mushroom, close the fingers firmly around it to secure it, then lift and place it onto the green plate.\",\n \"recovery_action_summary\": \"re-align gripper; grasp mushroom; place on plate\",\n \"prevention_action\": \"Lower the end-effector further and center the fingers precisely over the mushroom body before closing to achieve a firm grasp.\",\n \"prevention_action_summary\": \"center gripper; descend fully; grasp mushroom securely\",\n \"justification\": \"At around 4.5 seconds, the robot attempts to grasp the mushroom, but the fingers fail to clamp securely around it. The robot then lifts up and moves over the green plate empty-handed, leaving the task incomplete.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 278,
+    "failure_sim_step": 278,
+    "failure_policy_step": 46,
+    "failure_chunk_index": 2,
+    "first_post_failure_policy_step": 47,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-align the gripper over the mushroom, close the fingers firmly around it to secure it, then lift and place it onto the green plate."
+  }
+}
\ No newline at end of file
diff --git a/mushroom/ep026/labels.npz b/mushroom/ep026/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..810cd2d4029f3a434f01d5f7ef8f457d2b0310e7
--- /dev/null
+++ b/mushroom/ep026/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f4140fc5d860574d951145321dd87fbe8c8027d449ca5c48718161944ac08ccc
+size 3846
diff --git a/mushroom/ep027/example.md b/mushroom/ep027/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..1b5ddb01cf8f08de1c5e9c32ebee411757418fa1
--- /dev/null
+++ b/mushroom/ep027/example.md
@@ -0,0 +1,34 @@
+# t001__ep027__mushroom
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the mushroom on the plate
+- **success:** False
+- **control steps:** 582  (11.6s at 50 Hz)
+- **policy inferences:** 100  (replan_steps=6)
+- **video:** `mushroom/ep027/rollout.mp4`
+- **features:** `100` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | lowering gripper toward mushroom |
+| 3.0–6.0 | 151–301 | 25–50 | closing gripper on mushroom |
+| 6.0–9.0 | 302–452 | 50–75 | attempting second grasp on mushroom |
+| 9.0–11.6 | 453–581 | 75–96 | attempting third grasp on mushroom |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 6.274900398406374 s (00:06.3)
+- **frame / control_step:** 315 (policy_step=52, chunk_index=3)
+- **coarse onset:** 4.0 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot aligned its end-effector over the mushroom and closed its gripper, but failed to establish a secure hold, leaving the mushroom resting on the table.
+  Summary: failed to grasp mushroom; improper gripper alignment
+- **recovery:** Re-align the gripper directly over the mushroom, close the fingers firmly around it, lift it, and place it on the green plate.
+  Summary: realign gripper over mushroom; grasp and place on plate
+- **prevention:** Position the gripper precisely around the mushroom before closing the fingers to ensure a secure grip on the object.
+  Summary: center gripper over mushroom; establish secure grip
+- **justification:** None
diff --git a/mushroom/ep027/labels.json b/mushroom/ep027/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..c7d84ba20c06ebda91b419461ef8e294d32d324b
--- /dev/null
+++ b/mushroom/ep027/labels.json
@@ -0,0 +1,109 @@
+{
+  "rollout_id": "t001__ep027__mushroom",
+  "scene_variant": "",
+  "task_id": 1,
+  "task_file": "",
+  "instruction": "Put the mushroom on the plate",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:05:46+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00796812749004,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "lowering gripper toward mushroom",
+      "description": "lowering gripper toward mushroom"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00796812749004,
+      "t_end_sec": 6.01593625498008,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "closing gripper on mushroom",
+      "description": "closing gripper on mushroom"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01593625498008,
+      "t_end_sec": 9.02390438247012,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "attempting second grasp on mushroom",
+      "description": "attempting second grasp on mushroom"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.02390438247012,
+      "t_end_sec": 11.593625498007967,
+      "control_step_start": 453,
+      "control_step_end": 581,
+      "policy_step_start": 75,
+      "policy_step_end": 96,
+      "phrase": "attempting third grasp on mushroom",
+      "description": "attempting third grasp on mushroom"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 6.274900398406374,
+    "vlm_failure_onset_timestamp": "00:06.3",
+    "vlm_failure_onset_frame": 315,
+    "vlm_failure_onset_step": 315,
+    "vlm_coarse_failure_onset_seconds": 4.0,
+    "vlm_failure_window_start_seconds": 4.0,
+    "vlm_failure_window_end_seconds": 5.0,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot aligned its end-effector over the mushroom and closed its gripper, but failed to establish a secure hold, leaving the mushroom resting on the table.",
+    "vlm_failure_reason_summary": "failed to grasp mushroom; improper gripper alignment",
+    "vlm_recovery_action": "Re-align the gripper directly over the mushroom, close the fingers firmly around it, lift it, and place it on the green plate.",
+    "vlm_recovery_action_summary": "realign gripper over mushroom; grasp and place on plate",
+    "vlm_prevention_action": "Position the gripper precisely around the mushroom before closing the fingers to ensure a secure grip on the object.",
+    "vlm_prevention_action_summary": "center gripper over mushroom; establish secure grip",
+    "vlm_justification": null,
+    "vlm_temporal_justification": "At 04:25 (265 seconds into the clip), the robot gripper finishes closing its fingers over the mushroom but fails to pinch or establish a secure grip, leaving the object resting on the table.",
+    "vlm_video_fps": 50.2,
+    "vlm_video_num_frames": 582,
+    "vlm_video_duration_seconds": 11.573705179282868,
+    "vlm_usage": {
+      "input_tokens": 21703,
+      "output_tokens": 319,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 4.0,\n \"failure_window_start_seconds\": 4.0,\n \"failure_window_end_seconds\": 5.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot aligned its end-effector over the mushroom and closed its gripper, but failed to establish a secure hold, leaving the mushroom resting on the table.\",\n \"failure_reason_summary\": \"failed to grasp mushroom; improper gripper alignment\",\n \"recovery_action\": \"Re-align the gripper directly over the mushroom, close the fingers firmly around it, lift it, and place it on the green plate.\",\n \"recovery_action_summary\": \"realign gripper over mushroom; grasp and place on plate\",\n \"prevention_action\": \"Position the gripper precisely around the mushroom before closing the fingers to ensure a secure grip on the object.\",\n \"prevention_action_summary\": \"center gripper over mushroom; establish secure grip\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 315,
+    "failure_sim_step": 315,
+    "failure_policy_step": 52,
+    "failure_chunk_index": 3,
+    "first_post_failure_policy_step": 53,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-align the gripper directly over the mushroom, close the fingers firmly around it, lift it, and place it on the green plate."
+  }
+}
\ No newline at end of file
diff --git a/mushroom/ep027/labels.npz b/mushroom/ep027/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..d1368a3a063ca3b55b0777e3b2b46463c83f052d
--- /dev/null
+++ b/mushroom/ep027/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:345e34eb7e3689ba41037ae786507df9b938982c970eb10d73f8cd3a77a575ee
+size 3812
diff --git a/mushroom/ep028/example.md b/mushroom/ep028/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..72cf80926f26a88c6dc6db8ad4d69536c727730d
--- /dev/null
+++ b/mushroom/ep028/example.md
@@ -0,0 +1,21 @@
+# t001__ep028__mushroom
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the mushroom on the plate
+- **success:** True
+- **control steps:** 360  (7.2s at 50 Hz)
+- **policy inferences:** 63  (replan_steps=6)
+- **video:** `mushroom/ep028/rollout.mp4`
+- **features:** `63` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–149 | 0–24 | reaching and grasping mushroom |
+| 3.0–6.0 | 150–299 | 25–49 | placing mushroom on plate |
+| 6.0–7.2 | 300–359 | 50–59 | releasing mushroom and retracting |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/mushroom/ep028/labels.json b/mushroom/ep028/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..7a33c7ecb59c689e421f9ce08c218d088171cb1c
--- /dev/null
+++ b/mushroom/ep028/labels.json
@@ -0,0 +1,68 @@
+{
+  "rollout_id": "t001__ep028__mushroom",
+  "scene_variant": "",
+  "task_id": 1,
+  "task_file": "",
+  "instruction": "Put the mushroom on the plate",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:05:54+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.006012024048096,
+      "control_step_start": 0,
+      "control_step_end": 149,
+      "policy_step_start": 0,
+      "policy_step_end": 24,
+      "phrase": "reaching and grasping mushroom",
+      "description": "reaching and grasping mushroom"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.006012024048096,
+      "t_end_sec": 6.012024048096192,
+      "control_step_start": 150,
+      "control_step_end": 299,
+      "policy_step_start": 25,
+      "policy_step_end": 49,
+      "phrase": "placing mushroom on plate",
+      "description": "placing mushroom on plate"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.012024048096192,
+      "t_end_sec": 7.214428857715431,
+      "control_step_start": 300,
+      "control_step_end": 359,
+      "policy_step_start": 50,
+      "policy_step_end": 59,
+      "phrase": "releasing mushroom and retracting",
+      "description": "releasing mushroom and retracting"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/mushroom/ep028/labels.npz b/mushroom/ep028/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..923b6afa51b333a6d094d93917cee1672a59b5b1
--- /dev/null
+++ b/mushroom/ep028/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a9c9555d8fedfc4329ff5fa9ea1a43cd6a2cdeac660999e429ff28f3a56f63f6
+size 3080
diff --git a/mushroom/ep029/example.md b/mushroom/ep029/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..4b59afbbcf7a0bff62c19776305391d2222aeabe
--- /dev/null
+++ b/mushroom/ep029/example.md
@@ -0,0 +1,21 @@
+# t001__ep029__mushroom
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the mushroom on the plate
+- **success:** True
+- **control steps:** 382  (7.5s at 51 Hz)
+- **policy inferences:** 65  (replan_steps=6)
+- **video:** `mushroom/ep029/rollout.mp4`
+- **features:** `65` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching for and grasping mushroom |
+| 3.0–6.0 | 152–303 | 25–50 | placing mushroom on green plate |
+| 6.0–7.5 | 304–381 | 50–63 | releasing mushroom and retracting arm |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/mushroom/ep029/labels.json b/mushroom/ep029/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..56afae9b36cf4012aca0feabc44f71f7a37aeb4f
--- /dev/null
+++ b/mushroom/ep029/labels.json
@@ -0,0 +1,68 @@
+{
+  "rollout_id": "t001__ep029__mushroom",
+  "scene_variant": "",
+  "task_id": 1,
+  "task_file": "",
+  "instruction": "Put the mushroom on the plate",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:06:04+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0039525691699605,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for and grasping mushroom",
+      "description": "reaching for and grasping mushroom"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0039525691699605,
+      "t_end_sec": 6.007905138339921,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "placing mushroom on green plate",
+      "description": "placing mushroom on green plate"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.007905138339921,
+      "t_end_sec": 7.549407114624506,
+      "control_step_start": 304,
+      "control_step_end": 381,
+      "policy_step_start": 50,
+      "policy_step_end": 63,
+      "phrase": "releasing mushroom and retracting arm",
+      "description": "releasing mushroom and retracting arm"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/mushroom/ep029/labels.npz b/mushroom/ep029/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..4e54cf6dbbc2804602be3eb7f23ad38c2191febd
--- /dev/null
+++ b/mushroom/ep029/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:451ab18efa59a15e290fcd93c72ff66dcfce59e46d2d2ce0ec03cc74cd020fca
+size 3092
diff --git a/mushroom/ep030/example.md b/mushroom/ep030/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..66cd0ad85dc6772a87e7b436f8724fe09deae028
--- /dev/null
+++ b/mushroom/ep030/example.md
@@ -0,0 +1,21 @@
+# t001__ep030__mushroom
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the mushroom on the plate
+- **success:** True
+- **control steps:** 413  (8.2s at 51 Hz)
+- **policy inferences:** 72  (replan_steps=6)
+- **video:** `mushroom/ep030/rollout.mp4`
+- **features:** `72` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching and grasping mushroom |
+| 3.0–6.0 | 152–303 | 25–50 | moving mushroom to plate |
+| 6.0–8.2 | 304–412 | 50–68 | releasing mushroom and retracting |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/mushroom/ep030/labels.json b/mushroom/ep030/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..00c85c83490d09254d23d250f3203289383fbfa9
--- /dev/null
+++ b/mushroom/ep030/labels.json
@@ -0,0 +1,68 @@
+{
+  "rollout_id": "t001__ep030__mushroom",
+  "scene_variant": "",
+  "task_id": 1,
+  "task_file": "",
+  "instruction": "Put the mushroom on the plate",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:06:10+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0039525691699605,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching and grasping mushroom",
+      "description": "reaching and grasping mushroom"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0039525691699605,
+      "t_end_sec": 6.007905138339921,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "moving mushroom to plate",
+      "description": "moving mushroom to plate"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.007905138339921,
+      "t_end_sec": 8.162055335968379,
+      "control_step_start": 304,
+      "control_step_end": 412,
+      "policy_step_start": 50,
+      "policy_step_end": 68,
+      "phrase": "releasing mushroom and retracting",
+      "description": "releasing mushroom and retracting"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/mushroom/ep030/labels.npz b/mushroom/ep030/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..d237d4e8fde6c6d8095c1cf72b64d54e61c2b2ab
--- /dev/null
+++ b/mushroom/ep030/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:602675804d12aada38bc3198125ddc7499d28f213850d8b6aa34cbe4ad116ac9
+size 3086
diff --git a/mushroom/ep031/example.md b/mushroom/ep031/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..18cdddc82c9f8e0db611e085c23f06c6e73eade6
--- /dev/null
+++ b/mushroom/ep031/example.md
@@ -0,0 +1,22 @@
+# t001__ep031__mushroom
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the mushroom on the plate
+- **success:** True
+- **control steps:** 581  (11.4s at 51 Hz)
+- **policy inferences:** 100  (replan_steps=6)
+- **video:** `mushroom/ep031/rollout.mp4`
+- **features:** `100` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–152 | 0–25 | reaching for mushroom |
+| 3.0–6.0 | 153–305 | 25–50 | grasping mushroom |
+| 6.0–9.0 | 306–458 | 51–76 | moving mushroom to plate |
+| 9.0–11.4 | 459–580 | 76–96 | releasing mushroom and retracting |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/mushroom/ep031/labels.json b/mushroom/ep031/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..891883a7e0e7c6e065f0c0c8e827058774ddc652
--- /dev/null
+++ b/mushroom/ep031/labels.json
@@ -0,0 +1,79 @@
+{
+  "rollout_id": "t001__ep031__mushroom",
+  "scene_variant": "",
+  "task_id": 1,
+  "task_file": "",
+  "instruction": "Put the mushroom on the plate",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:06:18+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0,
+      "control_step_start": 0,
+      "control_step_end": 152,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for mushroom",
+      "description": "reaching for mushroom"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0,
+      "t_end_sec": 6.0,
+      "control_step_start": 153,
+      "control_step_end": 305,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "grasping mushroom",
+      "description": "grasping mushroom"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.0,
+      "t_end_sec": 9.0,
+      "control_step_start": 306,
+      "control_step_end": 458,
+      "policy_step_start": 51,
+      "policy_step_end": 76,
+      "phrase": "moving mushroom to plate",
+      "description": "moving mushroom to plate"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.0,
+      "t_end_sec": 11.392156862745098,
+      "control_step_start": 459,
+      "control_step_end": 580,
+      "policy_step_start": 76,
+      "policy_step_end": 96,
+      "phrase": "releasing mushroom and retracting",
+      "description": "releasing mushroom and retracting"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/mushroom/ep031/labels.npz b/mushroom/ep031/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..d8cb7d399b6444650b9207f61018c430dae14c7d
--- /dev/null
+++ b/mushroom/ep031/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5b362c8d3a7bb7f27593652cc6285e4f2f5083b91fe607d4df5501f267b577de
+size 3107
diff --git a/mushroom/ep032/example.md b/mushroom/ep032/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..6f0e1cc415bbf62a83636f6e2e89c88061d17dc6
--- /dev/null
+++ b/mushroom/ep032/example.md
@@ -0,0 +1,33 @@
+# t001__ep032__mushroom
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the mushroom on the plate
+- **success:** False
+- **control steps:** 380  (7.5s at 51 Hz)
+- **policy inferences:** 66  (replan_steps=6)
+- **video:** `mushroom/ep032/rollout.mp4`
+- **features:** `66` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching for mushroom |
+| 3.0–6.0 | 152–303 | 25–50 | attempting grasp and moving empty |
+| 6.0–7.5 | 304–379 | 50–63 | hovering over green plate |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 3.799212598425197 s (00:03.8)
+- **frame / control_step:** 193 (policy_step=32, chunk_index=1)
+- **coarse onset:** 3.8 s (refined=False)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the mushroom, but failed to secure a firm grip on it, lifting the gripper away while leaving the mushroom behind on the table.
+  Summary: failed grasp; lifting empty gripper
+- **recovery:** Lower the gripper back over the mushroom, firmly grasp its stem, and carry it onto the green plate.
+  Summary: re-align gripper; grasp mushroom; place on plate
+- **prevention:** Align the gripper properly around the mushroom before closing and ensure a firm grip before initiating the upward lifting movement.
+  Summary: align gripper centered; secure firm grasp
+- **justification:** At around 3.8 seconds, the robot arm lifts after closing the gripper, but the mushroom slips out and remains on the tabletop. The robot continues moving the empty gripper toward the green plate, failing the task.
diff --git a/mushroom/ep032/labels.json b/mushroom/ep032/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..7983727d151d8c775d62a987bb7a85fbef951d1c
--- /dev/null
+++ b/mushroom/ep032/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t001__ep032__mushroom",
+  "scene_variant": "",
+  "task_id": 1,
+  "task_file": "",
+  "instruction": "Put the mushroom on the plate",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:06:51+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9921259842519685,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for mushroom",
+      "description": "reaching for mushroom"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9921259842519685,
+      "t_end_sec": 5.984251968503937,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "attempting grasp and moving empty",
+      "description": "attempting grasp and moving empty"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.984251968503937,
+      "t_end_sec": 7.480314960629921,
+      "control_step_start": 304,
+      "control_step_end": 379,
+      "policy_step_start": 50,
+      "policy_step_end": 63,
+      "phrase": "hovering over green plate",
+      "description": "hovering over green plate"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 3.799212598425197,
+    "vlm_failure_onset_timestamp": "00:03.8",
+    "vlm_failure_onset_frame": 193,
+    "vlm_failure_onset_step": 193,
+    "vlm_coarse_failure_onset_seconds": 3.8,
+    "vlm_failure_window_start_seconds": 3.3,
+    "vlm_failure_window_end_seconds": 4.5,
+    "vlm_temporal_refined": false,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the mushroom, but failed to secure a firm grip on it, lifting the gripper away while leaving the mushroom behind on the table.",
+    "vlm_failure_reason_summary": "failed grasp; lifting empty gripper",
+    "vlm_recovery_action": "Lower the gripper back over the mushroom, firmly grasp its stem, and carry it onto the green plate.",
+    "vlm_recovery_action_summary": "re-align gripper; grasp mushroom; place on plate",
+    "vlm_prevention_action": "Align the gripper properly around the mushroom before closing and ensure a firm grip before initiating the upward lifting movement.",
+    "vlm_prevention_action_summary": "align gripper centered; secure firm grasp",
+    "vlm_justification": "At around 3.8 seconds, the robot arm lifts after closing the gripper, but the mushroom slips out and remains on the tabletop. The robot continues moving the empty gripper toward the green plate, failing the task.",
+    "vlm_temporal_justification": "At 03:58, the robot gripper opens and begins lifting upward away from the mushroom without securing it, leaving the mushroom behind on the table.",
+    "vlm_video_fps": 50.8,
+    "vlm_video_num_frames": 380,
+    "vlm_video_duration_seconds": 7.460629921259843,
+    "vlm_usage": {
+      "input_tokens": 21512,
+      "output_tokens": 348,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 3.8,\n \"failure_window_start_seconds\": 3.3,\n \"failure_window_end_seconds\": 4.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the mushroom, but failed to secure a firm grip on it, lifting the gripper away while leaving the mushroom behind on the table.\",\n \"failure_reason_summary\": \"failed grasp; lifting empty gripper\",\n \"recovery_action\": \"Lower the gripper back over the mushroom, firmly grasp its stem, and carry it onto the green plate.\",\n \"recovery_action_summary\": \"re-align gripper; grasp mushroom; place on plate\",\n \"prevention_action\": \"Align the gripper properly around the mushroom before closing and ensure a firm grip before initiating the upward lifting movement.\",\n \"prevention_action_summary\": \"align gripper centered; secure firm grasp\",\n \"justification\": \"At around 3.8 seconds, the robot arm lifts after closing the gripper, but the mushroom slips out and remains on the tabletop. The robot continues moving the empty gripper toward the green plate, failing the task.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 193,
+    "failure_sim_step": 193,
+    "failure_policy_step": 32,
+    "failure_chunk_index": 1,
+    "first_post_failure_policy_step": 33,
+    "failure_type": "grasp_failure",
+    "correction_action": "Lower the gripper back over the mushroom, firmly grasp its stem, and carry it onto the green plate."
+  }
+}
\ No newline at end of file
diff --git a/mushroom/ep032/labels.npz b/mushroom/ep032/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..330b0de53912357556c7f2d20eee1c148851bf42
--- /dev/null
+++ b/mushroom/ep032/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:76b9d6893dee35952ba04a4cca9bf86ffe1aa32dbcb6eb6bb22bcbbf9191be51
+size 3752
diff --git a/mushroom/ep033/example.md b/mushroom/ep033/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..a5e8c001ebc639187915b4610f6957c3595d26e6
--- /dev/null
+++ b/mushroom/ep033/example.md
@@ -0,0 +1,35 @@
+# t001__ep033__mushroom
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the mushroom on the plate
+- **success:** False
+- **control steps:** 614  (12.3s at 50 Hz)
+- **policy inferences:** 106  (replan_steps=6)
+- **video:** `mushroom/ep033/rollout.mp4`
+- **features:** `106` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–149 | 0–24 | approaching mushroom |
+| 3.0–6.0 | 150–299 | 25–49 | grasping and lifting mushroom |
+| 6.0–9.0 | 300–449 | 50–74 | moving mushroom toward plate |
+| 9.0–12.0 | 450–599 | 75–99 | dropping mushroom onto table |
+| 12.0–12.3 | 600–613 | 100–102 | hovering above table |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 10.039920159680639 s (00:10.0)
+- **frame / control_step:** 503 (policy_step=83, chunk_index=5)
+- **coarse onset:** 9.2 s (refined=True)
+- **mode:** object_displacement
+- **reason:** The robot carried the mushroom toward the green plate but dropped it near the edge, causing it to fall onto the table surface outside the plate.
+  Summary: dropping grasped object; object falling onto table
+- **recovery:** Re-grasp the mushroom from the table surface, carry it over the green plate, and lower it into the center of the plate before releasing.
+  Summary: re-grasp mushroom; place in plate
+- **prevention:** Move the gripper further over the center of the green plate and lower it completely before releasing the mushroom.
+  Summary: center gripper over plate; release gently
+- **justification:** At approximately 9.2 seconds, the robot opens its gripper to release the mushroom, but it is not sufficiently centered over the plate, resulting in the mushroom falling onto the table in front of the plate.
diff --git a/mushroom/ep033/labels.json b/mushroom/ep033/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..f184213877c3295df2a50d91188f0a7485db55e0
--- /dev/null
+++ b/mushroom/ep033/labels.json
@@ -0,0 +1,120 @@
+{
+  "rollout_id": "t001__ep033__mushroom",
+  "scene_variant": "",
+  "task_id": 1,
+  "task_file": "",
+  "instruction": "Put the mushroom on the plate",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:07:22+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9940119760479043,
+      "control_step_start": 0,
+      "control_step_end": 149,
+      "policy_step_start": 0,
+      "policy_step_end": 24,
+      "phrase": "approaching mushroom",
+      "description": "approaching mushroom"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9940119760479043,
+      "t_end_sec": 5.9880239520958085,
+      "control_step_start": 150,
+      "control_step_end": 299,
+      "policy_step_start": 25,
+      "policy_step_end": 49,
+      "phrase": "grasping and lifting mushroom",
+      "description": "grasping and lifting mushroom"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.9880239520958085,
+      "t_end_sec": 8.982035928143713,
+      "control_step_start": 300,
+      "control_step_end": 449,
+      "policy_step_start": 50,
+      "policy_step_end": 74,
+      "phrase": "moving mushroom toward plate",
+      "description": "moving mushroom toward plate"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.982035928143713,
+      "t_end_sec": 11.976047904191617,
+      "control_step_start": 450,
+      "control_step_end": 599,
+      "policy_step_start": 75,
+      "policy_step_end": 99,
+      "phrase": "dropping mushroom onto table",
+      "description": "dropping mushroom onto table"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 11.976047904191617,
+      "t_end_sec": 12.255489021956087,
+      "control_step_start": 600,
+      "control_step_end": 613,
+      "policy_step_start": 100,
+      "policy_step_end": 102,
+      "phrase": "hovering above table",
+      "description": "hovering above table"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "object_displacement",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 10.039920159680639,
+    "vlm_failure_onset_timestamp": "00:10.0",
+    "vlm_failure_onset_frame": 503,
+    "vlm_failure_onset_step": 503,
+    "vlm_coarse_failure_onset_seconds": 9.2,
+    "vlm_failure_window_start_seconds": 8.7,
+    "vlm_failure_window_end_seconds": 9.7,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot carried the mushroom toward the green plate but dropped it near the edge, causing it to fall onto the table surface outside the plate.",
+    "vlm_failure_reason_summary": "dropping grasped object; object falling onto table",
+    "vlm_recovery_action": "Re-grasp the mushroom from the table surface, carry it over the green plate, and lower it into the center of the plate before releasing.",
+    "vlm_recovery_action_summary": "re-grasp mushroom; place in plate",
+    "vlm_prevention_action": "Move the gripper further over the center of the green plate and lower it completely before releasing the mushroom.",
+    "vlm_prevention_action_summary": "center gripper over plate; release gently",
+    "vlm_justification": "At approximately 9.2 seconds, the robot opens its gripper to release the mushroom, but it is not sufficiently centered over the plate, resulting in the mushroom falling onto the table in front of the plate.",
+    "vlm_temporal_justification": "At 03:12 (192 seconds), the robot opens its gripper to release the mushroom, but it is positioned over the table surface outside the green plate rather than properly inside it, causing the mushroom to fall onto the table.",
+    "vlm_video_fps": 50.1,
+    "vlm_video_num_frames": 614,
+    "vlm_video_duration_seconds": 12.235528942115769,
+    "vlm_usage": {
+      "input_tokens": 21577,
+      "output_tokens": 365,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"object_displacement\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 9.2,\n \"failure_window_start_seconds\": 8.7,\n \"failure_window_end_seconds\": 9.7,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot carried the mushroom toward the green plate but dropped it near the edge, causing it to fall onto the table surface outside the plate.\",\n \"failure_reason_summary\": \"dropping grasped object; object falling onto table\",\n \"recovery_action\": \"Re-grasp the mushroom from the table surface, carry it over the green plate, and lower it into the center of the plate before releasing.\",\n \"recovery_action_summary\": \"re-grasp mushroom; place in plate\",\n \"prevention_action\": \"Move the gripper further over the center of the green plate and lower it completely before releasing the mushroom.\",\n \"prevention_action_summary\": \"center gripper over plate; release gently\",\n \"justification\": \"At approximately 9.2 seconds, the robot opens its gripper to release the mushroom, but it is not sufficiently centered over the plate, resulting in the mushroom falling onto the table in front of the plate.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 503,
+    "failure_sim_step": 503,
+    "failure_policy_step": 83,
+    "failure_chunk_index": 5,
+    "first_post_failure_policy_step": 84,
+    "failure_type": "object_displacement",
+    "correction_action": "Re-grasp the mushroom from the table surface, carry it over the green plate, and lower it into the center of the plate before releasing."
+  }
+}
\ No newline at end of file
diff --git a/mushroom/ep033/labels.npz b/mushroom/ep033/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..705c10023a77e400d1360165aa90d499c43e4d29
--- /dev/null
+++ b/mushroom/ep033/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3b284024b75c2b8678147c862fef1530a1b2386fe22e82449c9e916cb26e2b4b
+size 3825
diff --git a/mushroom/ep034/example.md b/mushroom/ep034/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..305e4ba5f602a78c3e9ef3c557c868ff067702b8
--- /dev/null
+++ b/mushroom/ep034/example.md
@@ -0,0 +1,33 @@
+# t001__ep034__mushroom
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the mushroom on the plate
+- **success:** False
+- **control steps:** 334  (6.7s at 50 Hz)
+- **policy inferences:** 59  (replan_steps=5)
+- **video:** `mushroom/ep034/rollout.mp4`
+- **features:** `59` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–149 | 0–29 | approaching and grasping mushroom |
+| 3.0–6.0 | 150–299 | 30–59 | dropping mushroom on table |
+| 6.0–6.7 | 300–333 | 60–66 | retracting arm upward |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 5.169660678642715 s (00:05.2)
+- **frame / control_step:** 259 (policy_step=51, chunk_index=4)
+- **coarse onset:** 4.5 s (refined=True)
+- **mode:** object_displacement
+- **reason:** The robot successfully grasped and lifted the mushroom, but dropped it back onto the table before transporting it to the green plate.
+  Summary: dropped mushroom; premature release before plate
+- **recovery:** Re-align the gripper above the dropped mushroom, grasp it firmly, and carry it over to place it onto the green plate.
+  Summary: re-grasp mushroom; place onto green plate
+- **prevention:** Maintain a firm grip on the mushroom while moving toward the plate, releasing it only when positioned directly above the plate.
+  Summary: maintain grasp; release only over plate
+- **justification:** Around 4.5 seconds, after picking up the mushroom, the robot gripper opens prematurely and drops the mushroom onto the table surface. The arm then retracts empty, leaving the task incomplete.
diff --git a/mushroom/ep034/labels.json b/mushroom/ep034/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..b6b58d77fa9ecff6c7c733079a7e450dbedab1dc
--- /dev/null
+++ b/mushroom/ep034/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t001__ep034__mushroom",
+  "scene_variant": "",
+  "task_id": 1,
+  "task_file": "",
+  "instruction": "Put the mushroom on the plate",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:07:54+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9940119760479043,
+      "control_step_start": 0,
+      "control_step_end": 149,
+      "policy_step_start": 0,
+      "policy_step_end": 29,
+      "phrase": "approaching and grasping mushroom",
+      "description": "approaching and grasping mushroom"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9940119760479043,
+      "t_end_sec": 5.9880239520958085,
+      "control_step_start": 150,
+      "control_step_end": 299,
+      "policy_step_start": 30,
+      "policy_step_end": 59,
+      "phrase": "dropping mushroom on table",
+      "description": "dropping mushroom on table"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.9880239520958085,
+      "t_end_sec": 6.666666666666666,
+      "control_step_start": 300,
+      "control_step_end": 333,
+      "policy_step_start": 60,
+      "policy_step_end": 66,
+      "phrase": "retracting arm upward",
+      "description": "retracting arm upward"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "object_displacement",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 5.169660678642715,
+    "vlm_failure_onset_timestamp": "00:05.2",
+    "vlm_failure_onset_frame": 259,
+    "vlm_failure_onset_step": 259,
+    "vlm_coarse_failure_onset_seconds": 4.5,
+    "vlm_failure_window_start_seconds": 4.0,
+    "vlm_failure_window_end_seconds": 5.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot successfully grasped and lifted the mushroom, but dropped it back onto the table before transporting it to the green plate.",
+    "vlm_failure_reason_summary": "dropped mushroom; premature release before plate",
+    "vlm_recovery_action": "Re-align the gripper above the dropped mushroom, grasp it firmly, and carry it over to place it onto the green plate.",
+    "vlm_recovery_action_summary": "re-grasp mushroom; place onto green plate",
+    "vlm_prevention_action": "Maintain a firm grip on the mushroom while moving toward the plate, releasing it only when positioned directly above the plate.",
+    "vlm_prevention_action_summary": "maintain grasp; release only over plate",
+    "vlm_justification": "Around 4.5 seconds, after picking up the mushroom, the robot gripper opens prematurely and drops the mushroom onto the table surface. The arm then retracts empty, leaving the task incomplete.",
+    "vlm_temporal_justification": "At timestamp 03:04 (184s in the video clip), the robot gripper unexpectedly releases the mushroom, causing it to slip out of the fingers and fall back onto the table.",
+    "vlm_video_fps": 50.1,
+    "vlm_video_num_frames": 334,
+    "vlm_video_duration_seconds": 6.6467065868263475,
+    "vlm_usage": {
+      "input_tokens": 18607,
+      "output_tokens": 349,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"object_displacement\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 4.5,\n \"failure_window_start_seconds\": 4.0,\n \"failure_window_end_seconds\": 5.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot successfully grasped and lifted the mushroom, but dropped it back onto the table before transporting it to the green plate.\",\n \"failure_reason_summary\": \"dropped mushroom; premature release before plate\",\n \"recovery_action\": \"Re-align the gripper above the dropped mushroom, grasp it firmly, and carry it over to place it onto the green plate.\",\n \"recovery_action_summary\": \"re-grasp mushroom; place onto green plate\",\n \"prevention_action\": \"Maintain a firm grip on the mushroom while moving toward the plate, releasing it only when positioned directly above the plate.\",\n \"prevention_action_summary\": \"maintain grasp; release only over plate\",\n \"justification\": \"Around 4.5 seconds, after picking up the mushroom, the robot gripper opens prematurely and drops the mushroom onto the table surface. The arm then retracts empty, leaving the task incomplete.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 259,
+    "failure_sim_step": 259,
+    "failure_policy_step": 51,
+    "failure_chunk_index": 4,
+    "first_post_failure_policy_step": 52,
+    "failure_type": "object_displacement",
+    "correction_action": "Re-align the gripper above the dropped mushroom, grasp it firmly, and carry it over to place it onto the green plate."
+  }
+}
\ No newline at end of file
diff --git a/mushroom/ep034/labels.npz b/mushroom/ep034/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..9475c3a4f02c3384b1d7dae48b234db2b1713e19
--- /dev/null
+++ b/mushroom/ep034/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1118c6f172754ae2a1f07b9e1effa23b7785de363c2e5a97682508fae6bc48fd
+size 3760
diff --git a/mushroom/ep035/example.md b/mushroom/ep035/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..8dfd43feb644dbe8607b5f8545e176c2d7958858
--- /dev/null
+++ b/mushroom/ep035/example.md
@@ -0,0 +1,32 @@
+# t001__ep035__mushroom
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the mushroom on the plate
+- **success:** False
+- **control steps:** 295  (5.9s at 50 Hz)
+- **policy inferences:** 51  (replan_steps=6)
+- **video:** `mushroom/ep035/rollout.mp4`
+- **features:** `51` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | reaching and closing gripper on mushroom |
+| 3.0–5.9 | 151–294 | 25–49 | moving empty gripper over plate |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 3.538767395626243 s (00:03.5)
+- **frame / control_step:** 178 (policy_step=29, chunk_index=4)
+- **coarse onset:** 2.5 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to pick up the mushroom, but the gripper closed without securely grasping it, causing the robot to lift up empty and execute the placing motion without the object.
+  Summary: failed grasp on mushroom; moving empty gripper to plate
+- **recovery:** Re-align the gripper over the mushroom on the tabletop, close the fingers securely around it, and then carry it to the green plate.
+  Summary: re-align gripper; grasp mushroom; place on plate
+- **prevention:** Position the end-effector accurately over the center of the mushroom before closing the gripper to ensure a firm grasp before lifting.
+  Summary: center gripper over mushroom; establish secure grip before lifting
+- **justification:**  around 2.5 seconds, the robot closes its gripper over the mushroom but fails to secure it. By 3.0 seconds, the arm lifts up empty while the mushroom stays on the table, leading directly to task failure.
diff --git a/mushroom/ep035/labels.json b/mushroom/ep035/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..ec09a5f1bf11fd216e311eb20d840398067b1e07
--- /dev/null
+++ b/mushroom/ep035/labels.json
@@ -0,0 +1,87 @@
+{
+  "rollout_id": "t001__ep035__mushroom",
+  "scene_variant": "",
+  "task_id": 1,
+  "task_file": "",
+  "instruction": "Put the mushroom on the plate",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:08:26+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.001988071570577,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching and closing gripper on mushroom",
+      "description": "reaching and closing gripper on mushroom"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.001988071570577,
+      "t_end_sec": 5.864811133200796,
+      "control_step_start": 151,
+      "control_step_end": 294,
+      "policy_step_start": 25,
+      "policy_step_end": 49,
+      "phrase": "moving empty gripper over plate",
+      "description": "moving empty gripper over plate"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 3.538767395626243,
+    "vlm_failure_onset_timestamp": "00:03.5",
+    "vlm_failure_onset_frame": 178,
+    "vlm_failure_onset_step": 178,
+    "vlm_coarse_failure_onset_seconds": 2.5,
+    "vlm_failure_window_start_seconds": 2.0,
+    "vlm_failure_window_end_seconds": 3.0,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to pick up the mushroom, but the gripper closed without securely grasping it, causing the robot to lift up empty and execute the placing motion without the object.",
+    "vlm_failure_reason_summary": "failed grasp on mushroom; moving empty gripper to plate",
+    "vlm_recovery_action": "Re-align the gripper over the mushroom on the tabletop, close the fingers securely around it, and then carry it to the green plate.",
+    "vlm_recovery_action_summary": "re-align gripper; grasp mushroom; place on plate",
+    "vlm_prevention_action": "Position the end-effector accurately over the center of the mushroom before closing the gripper to ensure a firm grasp before lifting.",
+    "vlm_prevention_action_summary": "center gripper over mushroom; establish secure grip before lifting",
+    "vlm_justification": " around 2.5 seconds, the robot closes its gripper over the mushroom but fails to secure it. By 3.0 seconds, the arm lifts up empty while the mushroom stays on the table, leading directly to task failure.",
+    "vlm_temporal_justification": "The robot lowers its gripper around the mushroom and closes its fingers. By timestamp 02:58 (178 seconds), the gripper fingers have closed completely without grasping the mushroom, leaving the mushroom behind on the table as the robot arm prepares to lift.",
+    "vlm_video_fps": 50.3,
+    "vlm_video_num_frames": 295,
+    "vlm_video_duration_seconds": 5.84493041749503,
+    "vlm_usage": {
+      "input_tokens": 19748,
+      "output_tokens": 390,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 2.5,\n \"failure_window_start_seconds\": 2.0,\n \"failure_window_end_seconds\": 3.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to pick up the mushroom, but the gripper closed without securely grasping it, causing the robot to lift up empty and execute the placing motion without the object.\",\n \"failure_reason_summary\": \"failed grasp on mushroom; moving empty gripper to plate\",\n \"recovery_action\": \"Re-align the gripper over the mushroom on the tabletop, close the fingers securely around it, and then carry it to the green plate.\",\n \"recovery_action_summary\": \"re-align gripper; grasp mushroom; place on plate\",\n \"prevention_action\": \"Position the end-effector accurately over the center of the mushroom before closing the gripper to ensure a firm grasp before lifting.\",\n \"prevention_action_summary\": \"center gripper over mushroom; establish secure grip before lifting\",\n \"justification\": \" around 2.5 seconds, the robot closes its gripper over the mushroom but fails to secure it. By 3.0 seconds, the arm lifts up empty while the mushroom stays on the table, leading directly to task failure.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 178,
+    "failure_sim_step": 178,
+    "failure_policy_step": 29,
+    "failure_chunk_index": 4,
+    "first_post_failure_policy_step": 30,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-align the gripper over the mushroom on the tabletop, close the fingers securely around it, and then carry it to the green plate."
+  }
+}
\ No newline at end of file
diff --git a/mushroom/ep035/labels.npz b/mushroom/ep035/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..9424f30640d0407c3d586119f2784c3ce8e41225
--- /dev/null
+++ b/mushroom/ep035/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c9662d030da0b03945931bc8947cb46cd0641ffff8888a4a5926297c1b106b67
+size 3823
diff --git a/mushroom/ep036/example.md b/mushroom/ep036/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..bfd5092698f10e41838202e048dc665c4b07931d
--- /dev/null
+++ b/mushroom/ep036/example.md
@@ -0,0 +1,33 @@
+# t001__ep036__mushroom
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the mushroom on the plate
+- **success:** False
+- **control steps:** 379  (7.5s at 51 Hz)
+- **policy inferences:** 65  (replan_steps=6)
+- **video:** `mushroom/ep036/rollout.mp4`
+- **features:** `65` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching down toward mushroom |
+| 3.0–6.0 | 152–303 | 25–50 | closing gripper around mushroom |
+| 6.0–7.5 | 304–378 | 50–63 | holding empty gripper closed |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 5.216535433070867 s (00:05.2)
+- **frame / control_step:** 265 (policy_step=44, chunk_index=1)
+- **coarse onset:** 4.8 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempts to grasp the mushroom, but the gripper fingers close improperly around its stem/cap, causing the mushroom to slip out of the grasp.
+  Summary: improper grasp position; mushroom slipping from gripper
+- **recovery:** Re-center the gripper over the displaced mushroom, open the gripper fingers, execute a firm grasp, and then transfer it onto the green plate.
+  Summary: re-align gripper; regrasp mushroom; place on plate
+- **prevention:** Position the end-effector lower and centered over the mushroom's stem before closing the fingers to ensure a secure hold.
+  Summary: position gripper lower; secure grasp before lifting
+- **justification:** None
diff --git a/mushroom/ep036/labels.json b/mushroom/ep036/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..aa057e4a440c43a70d5ec4bfcefa4807f67459e3
--- /dev/null
+++ b/mushroom/ep036/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t001__ep036__mushroom",
+  "scene_variant": "",
+  "task_id": 1,
+  "task_file": "",
+  "instruction": "Put the mushroom on the plate",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:09:02+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9921259842519685,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching down toward mushroom",
+      "description": "reaching down toward mushroom"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9921259842519685,
+      "t_end_sec": 5.984251968503937,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "closing gripper around mushroom",
+      "description": "closing gripper around mushroom"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.984251968503937,
+      "t_end_sec": 7.460629921259843,
+      "control_step_start": 304,
+      "control_step_end": 378,
+      "policy_step_start": 50,
+      "policy_step_end": 63,
+      "phrase": "holding empty gripper closed",
+      "description": "holding empty gripper closed"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 5.216535433070867,
+    "vlm_failure_onset_timestamp": "00:05.2",
+    "vlm_failure_onset_frame": 265,
+    "vlm_failure_onset_step": 265,
+    "vlm_coarse_failure_onset_seconds": 4.8,
+    "vlm_failure_window_start_seconds": 4.0,
+    "vlm_failure_window_end_seconds": 5.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempts to grasp the mushroom, but the gripper fingers close improperly around its stem/cap, causing the mushroom to slip out of the grasp.",
+    "vlm_failure_reason_summary": "improper grasp position; mushroom slipping from gripper",
+    "vlm_recovery_action": "Re-center the gripper over the displaced mushroom, open the gripper fingers, execute a firm grasp, and then transfer it onto the green plate.",
+    "vlm_recovery_action_summary": "re-align gripper; regrasp mushroom; place on plate",
+    "vlm_prevention_action": "Position the end-effector lower and centered over the mushroom's stem before closing the fingers to ensure a secure hold.",
+    "vlm_prevention_action_summary": "position gripper lower; secure grasp before lifting",
+    "vlm_justification": null,
+    "vlm_temporal_justification": "At 02:53 (173 seconds into the clip), the robot gripper closes on the mushroom's cap, but improper finger alignment causes the mushroom to visibly tilt and slip out of the grasp.",
+    "vlm_video_fps": 50.8,
+    "vlm_video_num_frames": 379,
+    "vlm_video_duration_seconds": 7.440944881889764,
+    "vlm_usage": {
+      "input_tokens": 20377,
+      "output_tokens": 325,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 4.8,\n \"failure_window_start_seconds\": 4.0,\n \"failure_window_end_seconds\": 5.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempts to grasp the mushroom, but the gripper fingers close improperly around its stem/cap, causing the mushroom to slip out of the grasp.\",\n \"failure_reason_summary\": \"improper grasp position; mushroom slipping from gripper\",\n \"recovery_action\": \"Re-center the gripper over the displaced mushroom, open the gripper fingers, execute a firm grasp, and then transfer it onto the green plate.\",\n \"recovery_action_summary\": \"re-align gripper; regrasp mushroom; place on plate\",\n \"prevention_action\": \"Position the end-effector lower and centered over the mushroom's stem before closing the fingers to ensure a secure hold.\",\n \"prevention_action_summary\": \"position gripper lower; secure grasp before lifting\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 265,
+    "failure_sim_step": 265,
+    "failure_policy_step": 44,
+    "failure_chunk_index": 1,
+    "first_post_failure_policy_step": 45,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-center the gripper over the displaced mushroom, open the gripper fingers, execute a firm grasp, and then transfer it onto the green plate."
+  }
+}
\ No newline at end of file
diff --git a/mushroom/ep036/labels.npz b/mushroom/ep036/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..c87af321a877cdc7ae0bfa3910839969eea6aad5
--- /dev/null
+++ b/mushroom/ep036/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f4ab1f8905880a85813a52f35aba9f6fa73d3be6d38bf0db947d151885186a57
+size 3792
diff --git a/mushroom/ep037/example.md b/mushroom/ep037/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..41085db292c47477e685e102fb86d6c78ac1b145
--- /dev/null
+++ b/mushroom/ep037/example.md
@@ -0,0 +1,35 @@
+# t001__ep037__mushroom
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the mushroom on the plate
+- **success:** False
+- **control steps:** 725  (14.4s at 50 Hz)
+- **policy inferences:** 126  (replan_steps=6)
+- **video:** `mushroom/ep037/rollout.mp4`
+- **features:** `126` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching for mushroom |
+| 3.0–6.0 | 152–303 | 25–50 | closing gripper on mushroom |
+| 6.0–9.0 | 304–455 | 50–75 | attempting grasp on mushroom |
+| 9.0–12.0 | 456–607 | 76–101 | pressing gripper onto table |
+| 12.0–14.4 | 608–724 | 101–120 | holding position over mushroom |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 3.6633663366336635 s (00:03.7)
+- **frame / control_step:** 185 (policy_step=30, chunk_index=5)
+- **coarse onset:** 5.0 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot arm positioned its end-effector over the mushroom and closed its gripper, but failed to establish a secure hold, allowing the mushroom to slip out and remain on the table surface.
+  Summary: failed grasp on mushroom; gripper closed without securing object
+- **recovery:** Re-open the gripper, realign the fingers centered over the mushroom at a lower grasp height, establish a firm grip, and carry it onto the green plate.
+  Summary: realign gripper over mushroom; establish secure grasp; place on plate
+- **prevention:** Ensure precise alignment and height calibration over the mushroom's center of mass before closing the gripper fingers to prevent slippage.
+  Summary: center gripper over mushroom; execute stable grasp
+- **justification:** At approximately 5.0 seconds, the robot closes its gripper onto the mushroom, but the object slips out instead of being grasped. The robot continues trying to close on the mushroom without successfully picking it up, failing the task goal.
diff --git a/mushroom/ep037/labels.json b/mushroom/ep037/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..1af702901e470a43da1e41796d15604aabb027d0
--- /dev/null
+++ b/mushroom/ep037/labels.json
@@ -0,0 +1,120 @@
+{
+  "rollout_id": "t001__ep037__mushroom",
+  "scene_variant": "",
+  "task_id": 1,
+  "task_file": "",
+  "instruction": "Put the mushroom on the plate",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:09:34+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00990099009901,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for mushroom",
+      "description": "reaching for mushroom"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00990099009901,
+      "t_end_sec": 6.01980198019802,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "closing gripper on mushroom",
+      "description": "closing gripper on mushroom"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01980198019802,
+      "t_end_sec": 9.029702970297029,
+      "control_step_start": 304,
+      "control_step_end": 455,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "attempting grasp on mushroom",
+      "description": "attempting grasp on mushroom"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.029702970297029,
+      "t_end_sec": 12.03960396039604,
+      "control_step_start": 456,
+      "control_step_end": 607,
+      "policy_step_start": 76,
+      "policy_step_end": 101,
+      "phrase": "pressing gripper onto table",
+      "description": "pressing gripper onto table"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 12.03960396039604,
+      "t_end_sec": 14.356435643564357,
+      "control_step_start": 608,
+      "control_step_end": 724,
+      "policy_step_start": 101,
+      "policy_step_end": 120,
+      "phrase": "holding position over mushroom",
+      "description": "holding position over mushroom"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 3.6633663366336635,
+    "vlm_failure_onset_timestamp": "00:03.7",
+    "vlm_failure_onset_frame": 185,
+    "vlm_failure_onset_step": 185,
+    "vlm_coarse_failure_onset_seconds": 5.0,
+    "vlm_failure_window_start_seconds": 4.5,
+    "vlm_failure_window_end_seconds": 5.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot arm positioned its end-effector over the mushroom and closed its gripper, but failed to establish a secure hold, allowing the mushroom to slip out and remain on the table surface.",
+    "vlm_failure_reason_summary": "failed grasp on mushroom; gripper closed without securing object",
+    "vlm_recovery_action": "Re-open the gripper, realign the fingers centered over the mushroom at a lower grasp height, establish a firm grip, and carry it onto the green plate.",
+    "vlm_recovery_action_summary": "realign gripper over mushroom; establish secure grasp; place on plate",
+    "vlm_prevention_action": "Ensure precise alignment and height calibration over the mushroom's center of mass before closing the gripper fingers to prevent slippage.",
+    "vlm_prevention_action_summary": "center gripper over mushroom; execute stable grasp",
+    "vlm_justification": "At approximately 5.0 seconds, the robot closes its gripper onto the mushroom, but the object slips out instead of being grasped. The robot continues trying to close on the mushroom without successfully picking it up, failing the task goal.",
+    "vlm_temporal_justification": "At timestamp 01:25 (second 85), the robot gripper finishes closing over the mushroom without establishing a secure grasp, leaving the mushroom sitting ungrasped on the table as the arm starts lifting away.",
+    "vlm_video_fps": 50.5,
+    "vlm_video_num_frames": 725,
+    "vlm_video_duration_seconds": 14.336633663366337,
+    "vlm_usage": {
+      "input_tokens": 21964,
+      "output_tokens": 389,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 5.0,\n \"failure_window_start_seconds\": 4.5,\n \"failure_window_end_seconds\": 5.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot arm positioned its end-effector over the mushroom and closed its gripper, but failed to establish a secure hold, allowing the mushroom to slip out and remain on the table surface.\",\n \"failure_reason_summary\": \"failed grasp on mushroom; gripper closed without securing object\",\n \"recovery_action\": \"Re-open the gripper, realign the fingers centered over the mushroom at a lower grasp height, establish a firm grip, and carry it onto the green plate.\",\n \"recovery_action_summary\": \"realign gripper over mushroom; establish secure grasp; place on plate\",\n \"prevention_action\": \"Ensure precise alignment and height calibration over the mushroom's center of mass before closing the gripper fingers to prevent slippage.\",\n \"prevention_action_summary\": \"center gripper over mushroom; execute stable grasp\",\n \"justification\": \"At approximately 5.0 seconds, the robot closes its gripper onto the mushroom, but the object slips out instead of being grasped. The robot continues trying to close on the mushroom without successfully picking it up, failing the task goal.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 185,
+    "failure_sim_step": 185,
+    "failure_policy_step": 30,
+    "failure_chunk_index": 5,
+    "first_post_failure_policy_step": 31,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-open the gripper, realign the fingers centered over the mushroom at a lower grasp height, establish a firm grip, and carry it onto the green plate."
+  }
+}
\ No newline at end of file
diff --git a/mushroom/ep037/labels.npz b/mushroom/ep037/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..939df3ec74efd67f98bba7688a76d742f58b7c63
--- /dev/null
+++ b/mushroom/ep037/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:757a3a0827217a1d6e0839363dd24be250b3b92a1c6550cd8feec2409c5e7973
+size 3983
diff --git a/mushroom/ep038/example.md b/mushroom/ep038/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..9e61a520ba25ece5dae871f6692cea678b94c414
--- /dev/null
+++ b/mushroom/ep038/example.md
@@ -0,0 +1,34 @@
+# t001__ep038__mushroom
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the mushroom on the plate
+- **success:** False
+- **control steps:** 581  (11.5s at 50 Hz)
+- **policy inferences:** 100  (replan_steps=6)
+- **video:** `mushroom/ep038/rollout.mp4`
+- **features:** `100` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | approaching mushroom |
+| 3.0–6.0 | 152–303 | 25–50 | closing gripper on mushroom |
+| 6.0–9.0 | 304–455 | 50–75 | moving empty gripper toward plate |
+| 9.0–11.5 | 456–580 | 76–96 | opening empty gripper over plate |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 8.257425742574258 s (00:08.3)
+- **frame / control_step:** 417 (policy_step=69, chunk_index=3)
+- **coarse onset:** 5.5 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot closed its gripper on the mushroom but failed to form a secure grip, leaving the mushroom on the table while moving the empty gripper toward the green plate.
+  Summary: failed grasp attempt; robot moving empty gripper to plate
+- **recovery:** Re-align the end-effector over the mushroom on the table, lower the gripper to clamp securely around the mushroom, lift it, and place it onto the green plate.
+  Summary: re-align gripper; regrasp mushroom; place on green plate
+- **prevention:** Lower the end-effector closer to the mushroom and align the fingers firmly around its stem before closing the gripper and lifting.
+  Summary: position gripper lower; grasp mushroom firmly before moving
+- **justification:** None
diff --git a/mushroom/ep038/labels.json b/mushroom/ep038/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..5ecd5f1110afa34253230784699ccda41eed7773
--- /dev/null
+++ b/mushroom/ep038/labels.json
@@ -0,0 +1,109 @@
+{
+  "rollout_id": "t001__ep038__mushroom",
+  "scene_variant": "",
+  "task_id": 1,
+  "task_file": "",
+  "instruction": "Put the mushroom on the plate",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:10:09+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00990099009901,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "approaching mushroom",
+      "description": "approaching mushroom"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00990099009901,
+      "t_end_sec": 6.01980198019802,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "closing gripper on mushroom",
+      "description": "closing gripper on mushroom"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01980198019802,
+      "t_end_sec": 9.029702970297029,
+      "control_step_start": 304,
+      "control_step_end": 455,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "moving empty gripper toward plate",
+      "description": "moving empty gripper toward plate"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.029702970297029,
+      "t_end_sec": 11.504950495049505,
+      "control_step_start": 456,
+      "control_step_end": 580,
+      "policy_step_start": 76,
+      "policy_step_end": 96,
+      "phrase": "opening empty gripper over plate",
+      "description": "opening empty gripper over plate"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 8.257425742574258,
+    "vlm_failure_onset_timestamp": "00:08.3",
+    "vlm_failure_onset_frame": 417,
+    "vlm_failure_onset_step": 417,
+    "vlm_coarse_failure_onset_seconds": 5.5,
+    "vlm_failure_window_start_seconds": 5.0,
+    "vlm_failure_window_end_seconds": 6.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot closed its gripper on the mushroom but failed to form a secure grip, leaving the mushroom on the table while moving the empty gripper toward the green plate.",
+    "vlm_failure_reason_summary": "failed grasp attempt; robot moving empty gripper to plate",
+    "vlm_recovery_action": "Re-align the end-effector over the mushroom on the table, lower the gripper to clamp securely around the mushroom, lift it, and place it onto the green plate.",
+    "vlm_recovery_action_summary": "re-align gripper; regrasp mushroom; place on green plate",
+    "vlm_prevention_action": "Lower the end-effector closer to the mushroom and align the fingers firmly around its stem before closing the gripper and lifting.",
+    "vlm_prevention_action_summary": "position gripper lower; grasp mushroom firmly before moving",
+    "vlm_justification": null,
+    "vlm_temporal_justification": "At timestamp 04:51 (291 seconds into the clip), the robot arm begins moving upward, but the mushroom remains on the table surface, making it visibly apparent that the grasp attempt failed to secure the object.",
+    "vlm_video_fps": 50.5,
+    "vlm_video_num_frames": 581,
+    "vlm_video_duration_seconds": 11.485148514851485,
+    "vlm_usage": {
+      "input_tokens": 21833,
+      "output_tokens": 337,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 5.5,\n \"failure_window_start_seconds\": 5.0,\n \"failure_window_end_seconds\": 6.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot closed its gripper on the mushroom but failed to form a secure grip, leaving the mushroom on the table while moving the empty gripper toward the green plate.\",\n \"failure_reason_summary\": \"failed grasp attempt; robot moving empty gripper to plate\",\n \"recovery_action\": \"Re-align the end-effector over the mushroom on the table, lower the gripper to clamp securely around the mushroom, lift it, and place it onto the green plate.\",\n \"recovery_action_summary\": \"re-align gripper; regrasp mushroom; place on green plate\",\n \"prevention_action\": \"Lower the end-effector closer to the mushroom and align the fingers firmly around its stem before closing the gripper and lifting.\",\n \"prevention_action_summary\": \"position gripper lower; grasp mushroom firmly before moving\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 417,
+    "failure_sim_step": 417,
+    "failure_policy_step": 69,
+    "failure_chunk_index": 3,
+    "first_post_failure_policy_step": 70,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-align the end-effector over the mushroom on the table, lower the gripper to clamp securely around the mushroom, lift it, and place it onto the green plate."
+  }
+}
\ No newline at end of file
diff --git a/mushroom/ep038/labels.npz b/mushroom/ep038/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..59043c75c44a7f7a377481fe5f56500a4ca22392
--- /dev/null
+++ b/mushroom/ep038/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0fba6ae78c8e17157b6c6055b9fc400eb57ac5060a6dff3e71804c768e600296
+size 3888
diff --git a/mushroom/ep039/example.md b/mushroom/ep039/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..2bd8d3f527b7401499f38dae4e0bd640eccd3968
--- /dev/null
+++ b/mushroom/ep039/example.md
@@ -0,0 +1,33 @@
+# t001__ep039__mushroom
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the mushroom on the plate
+- **success:** False
+- **control steps:** 303  (6.0s at 50 Hz)
+- **policy inferences:** 52  (replan_steps=6)
+- **video:** `mushroom/ep039/rollout.mp4`
+- **features:** `52` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–149 | 0–24 | lowering arm toward mushroom |
+| 3.0–6.0 | 150–299 | 25–49 | closing gripper around mushroom |
+| 6.0–6.0 | 300–302 | 50–50 | remaining stationary above mushroom |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 4.35129740518962 s (00:04.4)
+- **frame / control_step:** 218 (policy_step=36, chunk_index=2)
+- **coarse onset:** 4.2 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot descended to grasp the mushroom but failed to close its gripper securely around it, leaving the mushroom on the table.
+  Summary: failed grasp attempt; object not secured
+- **recovery:** Re-open the gripper, reposition it directly around the mushroom, close firmly to grasp it, and place it on the plate.
+  Summary: re-open gripper; grasp mushroom securely; place on plate
+- **prevention:** Align the gripper precisely over the center of the mushroom before closing the fingers to ensure a stable grasp.
+  Summary: align gripper precisely; secure grasp on mushroom
+- **justification:** At approximately 4.2 seconds, the robot's gripper attempts to close on the mushroom but fails to secure a grip, missing the object entirely as it stays on the table surface.
diff --git a/mushroom/ep039/labels.json b/mushroom/ep039/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..09fc6d424329fca0130383ee9705a3a85edd71fb
--- /dev/null
+++ b/mushroom/ep039/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t001__ep039__mushroom",
+  "scene_variant": "",
+  "task_id": 1,
+  "task_file": "",
+  "instruction": "Put the mushroom on the plate",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:10:39+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9940119760479043,
+      "control_step_start": 0,
+      "control_step_end": 149,
+      "policy_step_start": 0,
+      "policy_step_end": 24,
+      "phrase": "lowering arm toward mushroom",
+      "description": "lowering arm toward mushroom"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9940119760479043,
+      "t_end_sec": 5.9880239520958085,
+      "control_step_start": 150,
+      "control_step_end": 299,
+      "policy_step_start": 25,
+      "policy_step_end": 49,
+      "phrase": "closing gripper around mushroom",
+      "description": "closing gripper around mushroom"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.9880239520958085,
+      "t_end_sec": 6.047904191616766,
+      "control_step_start": 300,
+      "control_step_end": 302,
+      "policy_step_start": 50,
+      "policy_step_end": 50,
+      "phrase": "remaining stationary above mushroom",
+      "description": "remaining stationary above mushroom"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 4.35129740518962,
+    "vlm_failure_onset_timestamp": "00:04.4",
+    "vlm_failure_onset_frame": 218,
+    "vlm_failure_onset_step": 218,
+    "vlm_coarse_failure_onset_seconds": 4.2,
+    "vlm_failure_window_start_seconds": 3.7,
+    "vlm_failure_window_end_seconds": 4.7,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot descended to grasp the mushroom but failed to close its gripper securely around it, leaving the mushroom on the table.",
+    "vlm_failure_reason_summary": "failed grasp attempt; object not secured",
+    "vlm_recovery_action": "Re-open the gripper, reposition it directly around the mushroom, close firmly to grasp it, and place it on the plate.",
+    "vlm_recovery_action_summary": "re-open gripper; grasp mushroom securely; place on plate",
+    "vlm_prevention_action": "Align the gripper precisely over the center of the mushroom before closing the fingers to ensure a stable grasp.",
+    "vlm_prevention_action_summary": "align gripper precisely; secure grasp on mushroom",
+    "vlm_justification": "At approximately 4.2 seconds, the robot's gripper attempts to close on the mushroom but fails to secure a grip, missing the object entirely as it stays on the table surface.",
+    "vlm_temporal_justification": "At 02:38 (second 158), the robot gripper finishes closing at the target location but fails to securely grasp the mushroom, leaving it on the table.",
+    "vlm_video_fps": 50.1,
+    "vlm_video_num_frames": 303,
+    "vlm_video_duration_seconds": 6.0279441117764465,
+    "vlm_usage": {
+      "input_tokens": 17536,
+      "output_tokens": 345,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 4.2,\n \"failure_window_start_seconds\": 3.7,\n \"failure_window_end_seconds\": 4.7,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot descended to grasp the mushroom but failed to close its gripper securely around it, leaving the mushroom on the table.\",\n \"failure_reason_summary\": \"failed grasp attempt; object not secured\",\n \"recovery_action\": \"Re-open the gripper, reposition it directly around the mushroom, close firmly to grasp it, and place it on the plate.\",\n \"recovery_action_summary\": \"re-open gripper; grasp mushroom securely; place on plate\",\n \"prevention_action\": \"Align the gripper precisely over the center of the mushroom before closing the fingers to ensure a stable grasp.\",\n \"prevention_action_summary\": \"align gripper precisely; secure grasp on mushroom\",\n \"justification\": \"At approximately 4.2 seconds, the robot's gripper attempts to close on the mushroom but fails to secure a grip, missing the object entirely as it stays on the table surface.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 218,
+    "failure_sim_step": 218,
+    "failure_policy_step": 36,
+    "failure_chunk_index": 2,
+    "first_post_failure_policy_step": 37,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-open the gripper, reposition it directly around the mushroom, close firmly to grasp it, and place it on the plate."
+  }
+}
\ No newline at end of file
diff --git a/mushroom/ep039/labels.npz b/mushroom/ep039/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..cfd509e268b89a5f7532c2075cd2dfbaea08291e
--- /dev/null
+++ b/mushroom/ep039/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:913894a71424f8121b893a72f71dd1b4bd57da56065ab8151242d28e2eaba69a
+size 3755
diff --git a/mushroom/ep040/example.md b/mushroom/ep040/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..a3f7731b5270957ae28393da2235943e331f2b31
--- /dev/null
+++ b/mushroom/ep040/example.md
@@ -0,0 +1,21 @@
+# t001__ep040__mushroom
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the mushroom on the plate
+- **success:** True
+- **control steps:** 398  (7.9s at 51 Hz)
+- **policy inferences:** 69  (replan_steps=6)
+- **video:** `mushroom/ep040/rollout.mp4`
+- **features:** `69` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching and grasping mushroom |
+| 3.0–6.0 | 152–303 | 25–50 | placing mushroom on plate |
+| 6.0–7.9 | 304–397 | 50–66 | retracting robot arm |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/mushroom/ep040/labels.json b/mushroom/ep040/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..18fdfdf9be4b29e49d5af71fe2332d7843f2f5b6
--- /dev/null
+++ b/mushroom/ep040/labels.json
@@ -0,0 +1,68 @@
+{
+  "rollout_id": "t001__ep040__mushroom",
+  "scene_variant": "",
+  "task_id": 1,
+  "task_file": "",
+  "instruction": "Put the mushroom on the plate",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:10:46+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0039525691699605,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching and grasping mushroom",
+      "description": "reaching and grasping mushroom"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0039525691699605,
+      "t_end_sec": 6.007905138339921,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "placing mushroom on plate",
+      "description": "placing mushroom on plate"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.007905138339921,
+      "t_end_sec": 7.865612648221344,
+      "control_step_start": 304,
+      "control_step_end": 397,
+      "policy_step_start": 50,
+      "policy_step_end": 66,
+      "phrase": "retracting robot arm",
+      "description": "retracting robot arm"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/mushroom/ep040/labels.npz b/mushroom/ep040/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..317a73a487db0df8a5dbb0db980b0199b39581a8
--- /dev/null
+++ b/mushroom/ep040/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:87671b65d5acc75a7e762044d5d03de1de9553b0feed9d23d37914f5d7fe9902
+size 3078
diff --git a/mushroom/ep041/example.md b/mushroom/ep041/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..c237df3d6c0a821efeeb4ef4e6417d64437d5491
--- /dev/null
+++ b/mushroom/ep041/example.md
@@ -0,0 +1,33 @@
+# t001__ep041__mushroom
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the mushroom on the plate
+- **success:** False
+- **control steps:** 308  (6.1s at 50 Hz)
+- **policy inferences:** 54  (replan_steps=5)
+- **video:** `mushroom/ep041/rollout.mp4`
+- **features:** `54` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–30 | reaching and grasping at mushroom |
+| 3.0–6.0 | 151–301 | 30–60 | moving empty gripper to plate |
+| 6.0–6.1 | 302–307 | 60–61 | hovering over green plate |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 3.8492063492063493 s (00:03.8)
+- **frame / control_step:** 194 (policy_step=38, chunk_index=4)
+- **coarse onset:** 2.2 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the mushroom but failed to establish a secure grip, leaving the mushroom behind on the table while moving the empty end-effector to the plate.
+  Summary: failed grasp on mushroom; moving empty gripper
+- **recovery:** Return the end-effector to the mushroom, align the gripper precisely to form a secure hold, pick it up, and place it on the green plate.
+  Summary: re-align gripper; grasp mushroom securely; place on plate
+- **prevention:** Ensure accurate gripper alignment and complete closure around the mushroom to secure the grasp before lifting and moving towards the target plate.
+  Summary: align gripper precisely; verify secure grasp before moving
+- **justification:** Around 2.2 seconds into the rollout, the robot closes its gripper on the mushroom but fails to hold it. The arm then lifts and moves over the green plate with an empty gripper, leaving the task incomplete.
diff --git a/mushroom/ep041/labels.json b/mushroom/ep041/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..d9b32bf9c6d690c575e35ac4ede84bc14a7a2f0f
--- /dev/null
+++ b/mushroom/ep041/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t001__ep041__mushroom",
+  "scene_variant": "",
+  "task_id": 1,
+  "task_file": "",
+  "instruction": "Put the mushroom on the plate",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:11:20+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.996031746031746,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 30,
+      "phrase": "reaching and grasping at mushroom",
+      "description": "reaching and grasping at mushroom"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.996031746031746,
+      "t_end_sec": 5.992063492063492,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 30,
+      "policy_step_end": 60,
+      "phrase": "moving empty gripper to plate",
+      "description": "moving empty gripper to plate"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.992063492063492,
+      "t_end_sec": 6.111111111111112,
+      "control_step_start": 302,
+      "control_step_end": 307,
+      "policy_step_start": 60,
+      "policy_step_end": 61,
+      "phrase": "hovering over green plate",
+      "description": "hovering over green plate"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 3.8492063492063493,
+    "vlm_failure_onset_timestamp": "00:03.8",
+    "vlm_failure_onset_frame": 194,
+    "vlm_failure_onset_step": 194,
+    "vlm_coarse_failure_onset_seconds": 2.2,
+    "vlm_failure_window_start_seconds": 1.7,
+    "vlm_failure_window_end_seconds": 2.7,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the mushroom but failed to establish a secure grip, leaving the mushroom behind on the table while moving the empty end-effector to the plate.",
+    "vlm_failure_reason_summary": "failed grasp on mushroom; moving empty gripper",
+    "vlm_recovery_action": "Return the end-effector to the mushroom, align the gripper precisely to form a secure hold, pick it up, and place it on the green plate.",
+    "vlm_recovery_action_summary": "re-align gripper; grasp mushroom securely; place on plate",
+    "vlm_prevention_action": "Ensure accurate gripper alignment and complete closure around the mushroom to secure the grasp before lifting and moving towards the target plate.",
+    "vlm_prevention_action_summary": "align gripper precisely; verify secure grasp before moving",
+    "vlm_justification": "Around 2.2 seconds into the rollout, the robot closes its gripper on the mushroom but fails to hold it. The arm then lifts and moves over the green plate with an empty gripper, leaving the task incomplete.",
+    "vlm_temporal_justification": "At timestamp 03:14 (frame 194), the robot arm begins lifting its end-effector after closing the gripper, but fails to secure the mushroom, leaving it behind on the table surface.",
+    "vlm_video_fps": 50.4,
+    "vlm_video_num_frames": 308,
+    "vlm_video_duration_seconds": 6.091269841269842,
+    "vlm_usage": {
+      "input_tokens": 18802,
+      "output_tokens": 377,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 2.2,\n \"failure_window_start_seconds\": 1.7,\n \"failure_window_end_seconds\": 2.7,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the mushroom but failed to establish a secure grip, leaving the mushroom behind on the table while moving the empty end-effector to the plate.\",\n \"failure_reason_summary\": \"failed grasp on mushroom; moving empty gripper\",\n \"recovery_action\": \"Return the end-effector to the mushroom, align the gripper precisely to form a secure hold, pick it up, and place it on the green plate.\",\n \"recovery_action_summary\": \"re-align gripper; grasp mushroom securely; place on plate\",\n \"prevention_action\": \"Ensure accurate gripper alignment and complete closure around the mushroom to secure the grasp before lifting and moving towards the target plate.\",\n \"prevention_action_summary\": \"align gripper precisely; verify secure grasp before moving\",\n \"justification\": \"Around 2.2 seconds into the rollout, the robot closes its gripper on the mushroom but fails to hold it. The arm then lifts and moves over the green plate with an empty gripper, leaving the task incomplete.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 194,
+    "failure_sim_step": 194,
+    "failure_policy_step": 38,
+    "failure_chunk_index": 4,
+    "first_post_failure_policy_step": 39,
+    "failure_type": "grasp_failure",
+    "correction_action": "Return the end-effector to the mushroom, align the gripper precisely to form a secure hold, pick it up, and place it on the green plate."
+  }
+}
\ No newline at end of file
diff --git a/mushroom/ep041/labels.npz b/mushroom/ep041/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..945e141ea695950fee92b1606959dcf4a3f8c667
--- /dev/null
+++ b/mushroom/ep041/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:80f0be15b02994631fef5ba0376d43caef0aaeb315e4d6cb9df972ca64d4ab85
+size 3859
diff --git a/mushroom/ep042/example.md b/mushroom/ep042/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..194934ea380a7bd85d0e64e023aadabe6e2fa570
--- /dev/null
+++ b/mushroom/ep042/example.md
@@ -0,0 +1,33 @@
+# t001__ep042__mushroom
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the mushroom on the plate
+- **success:** False
+- **control steps:** 407  (8.0s at 51 Hz)
+- **policy inferences:** 70  (replan_steps=6)
+- **video:** `mushroom/ep042/rollout.mp4`
+- **features:** `70` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching for mushroom; closing gripper |
+| 3.0–6.0 | 152–303 | 25–50 | lifting mushroom; dropping near plate |
+| 6.0–8.0 | 304–406 | 50–67 | retracting empty gripper |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 6.370808678500985 s (00:06.4)
+- **frame / control_step:** 323 (policy_step=53, chunk_index=5)
+- **coarse onset:** 5.2 s (refined=True)
+- **mode:** object_displacement
+- **reason:** The robot successfully picked up the mushroom and transported it toward the green plate, but prematurely opened its gripper, dropping the mushroom onto the table surface next to the plate.
+  Summary: dropping mushroom before plate; premature release
+- **recovery:** Lower the end effector back to the table, re-grasp the dropped mushroom, and place it onto the center of the green plate.
+  Summary: re-grasp mushroom; place on green plate
+- **prevention:** Maintain a secure hold on the mushroom during transport and only open the gripper after positioning the end effector directly above the green plate.
+  Summary: maintain grip over plate; center release position
+- **justification:** At approximately 5.2 seconds, as the robot arm reaches toward the green plate, it opens its gripper too early, causing the held mushroom to fall onto the table right beside the plate instead of landing on it.
diff --git a/mushroom/ep042/labels.json b/mushroom/ep042/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..3f0562344d34294e09f7390fb3e1824fecc28ccb
--- /dev/null
+++ b/mushroom/ep042/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t001__ep042__mushroom",
+  "scene_variant": "",
+  "task_id": 1,
+  "task_file": "",
+  "instruction": "Put the mushroom on the plate",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:11:53+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9980276134122286,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for mushroom; closing gripper",
+      "description": "reaching for mushroom; closing gripper"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9980276134122286,
+      "t_end_sec": 5.996055226824457,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "lifting mushroom; dropping near plate",
+      "description": "lifting mushroom; dropping near plate"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.996055226824457,
+      "t_end_sec": 8.027613412228796,
+      "control_step_start": 304,
+      "control_step_end": 406,
+      "policy_step_start": 50,
+      "policy_step_end": 67,
+      "phrase": "retracting empty gripper",
+      "description": "retracting empty gripper"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "object_displacement",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 6.370808678500985,
+    "vlm_failure_onset_timestamp": "00:06.4",
+    "vlm_failure_onset_frame": 323,
+    "vlm_failure_onset_step": 323,
+    "vlm_coarse_failure_onset_seconds": 5.2,
+    "vlm_failure_window_start_seconds": 4.7,
+    "vlm_failure_window_end_seconds": 5.7,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot successfully picked up the mushroom and transported it toward the green plate, but prematurely opened its gripper, dropping the mushroom onto the table surface next to the plate.",
+    "vlm_failure_reason_summary": "dropping mushroom before plate; premature release",
+    "vlm_recovery_action": "Lower the end effector back to the table, re-grasp the dropped mushroom, and place it onto the center of the green plate.",
+    "vlm_recovery_action_summary": "re-grasp mushroom; place on green plate",
+    "vlm_prevention_action": "Maintain a secure hold on the mushroom during transport and only open the gripper after positioning the end effector directly above the green plate.",
+    "vlm_prevention_action_summary": "maintain grip over plate; center release position",
+    "vlm_justification": "At approximately 5.2 seconds, as the robot arm reaches toward the green plate, it opens its gripper too early, causing the held mushroom to fall onto the table right beside the plate instead of landing on it.",
+    "vlm_temporal_justification": "At timestamp 03:31, the robot arm prematurely opens its gripper while positioned over the table instead of over the green plate, causing the mushroom to fall onto the table surface.",
+    "vlm_video_fps": 50.7,
+    "vlm_video_num_frames": 407,
+    "vlm_video_duration_seconds": 8.007889546351084,
+    "vlm_usage": {
+      "input_tokens": 20950,
+      "output_tokens": 364,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"object_displacement\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 5.2,\n \"failure_window_start_seconds\": 4.7,\n \"failure_window_end_seconds\": 5.7,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot successfully picked up the mushroom and transported it toward the green plate, but prematurely opened its gripper, dropping the mushroom onto the table surface next to the plate.\",\n \"failure_reason_summary\": \"dropping mushroom before plate; premature release\",\n \"recovery_action\": \"Lower the end effector back to the table, re-grasp the dropped mushroom, and place it onto the center of the green plate.\",\n \"recovery_action_summary\": \"re-grasp mushroom; place on green plate\",\n \"prevention_action\": \"Maintain a secure hold on the mushroom during transport and only open the gripper after positioning the end effector directly above the green plate.\",\n \"prevention_action_summary\": \"maintain grip over plate; center release position\",\n \"justification\": \"At approximately 5.2 seconds, as the robot arm reaches toward the green plate, it opens its gripper too early, causing the held mushroom to fall onto the table right beside the plate instead of landing on it.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 323,
+    "failure_sim_step": 323,
+    "failure_policy_step": 53,
+    "failure_chunk_index": 5,
+    "first_post_failure_policy_step": 54,
+    "failure_type": "object_displacement",
+    "correction_action": "Lower the end effector back to the table, re-grasp the dropped mushroom, and place it onto the center of the green plate."
+  }
+}
\ No newline at end of file
diff --git a/mushroom/ep042/labels.npz b/mushroom/ep042/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..0d1088de5a9123156ce4f295660c4e27b84365f5
--- /dev/null
+++ b/mushroom/ep042/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6a2f89a80e8ee74b61c8a59ff318fc85050aa20ecbbeb2deba0507b1f0e0d3ed
+size 3841
diff --git a/mushroom/ep043/example.md b/mushroom/ep043/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..115bb42ca3e878a113cfc7b374a4f2ac5765b185
--- /dev/null
+++ b/mushroom/ep043/example.md
@@ -0,0 +1,34 @@
+# t001__ep043__mushroom
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the mushroom on the plate
+- **success:** False
+- **control steps:** 563  (11.1s at 50 Hz)
+- **policy inferences:** 97  (replan_steps=6)
+- **video:** `mushroom/ep043/rollout.mp4`
+- **features:** `97` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching for mushroom |
+| 3.0–6.0 | 152–303 | 25–50 | grasping and releasing mushroom prematurely |
+| 6.0–9.0 | 304–455 | 50–75 | hovering near green plate |
+| 9.0–11.1 | 456–562 | 76–93 | idle above table |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 4.554455445544554 s (00:04.6)
+- **frame / control_step:** 230 (policy_step=38, chunk_index=2)
+- **coarse onset:** 5.5 s (refined=True)
+- **mode:** placement_or_insertion_failure
+- **reason:** The robot moved the mushroom toward the green plate but opened its gripper prematurely at the edge, causing the mushroom to fall onto the table outside the plate.
+  Summary: premature release; mushroom placed off-target
+- **recovery:** Re-grasp the mushroom from the table, position the end-effector centered directly over the green plate, and release the object onto the plate surface.
+  Summary: regrasp mushroom; place inside plate
+- **prevention:** Carry the mushroom fully over the center of the green plate before opening the gripper to release it.
+  Summary: center over plate; release mushroom
+- **justification:** None
diff --git a/mushroom/ep043/labels.json b/mushroom/ep043/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..e455db8739ee54d2c1b6b52650e5d0b96eda50cb
--- /dev/null
+++ b/mushroom/ep043/labels.json
@@ -0,0 +1,109 @@
+{
+  "rollout_id": "t001__ep043__mushroom",
+  "scene_variant": "",
+  "task_id": 1,
+  "task_file": "",
+  "instruction": "Put the mushroom on the plate",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:12:28+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00990099009901,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for mushroom",
+      "description": "reaching for mushroom"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00990099009901,
+      "t_end_sec": 6.01980198019802,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "grasping and releasing mushroom prematurely",
+      "description": "grasping and releasing mushroom prematurely"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01980198019802,
+      "t_end_sec": 9.029702970297029,
+      "control_step_start": 304,
+      "control_step_end": 455,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "hovering near green plate",
+      "description": "hovering near green plate"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.029702970297029,
+      "t_end_sec": 11.148514851485148,
+      "control_step_start": 456,
+      "control_step_end": 562,
+      "policy_step_start": 76,
+      "policy_step_end": 93,
+      "phrase": "idle above table",
+      "description": "idle above table"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "placement_or_insertion_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 4.554455445544554,
+    "vlm_failure_onset_timestamp": "00:04.6",
+    "vlm_failure_onset_frame": 230,
+    "vlm_failure_onset_step": 230,
+    "vlm_coarse_failure_onset_seconds": 5.5,
+    "vlm_failure_window_start_seconds": 5.0,
+    "vlm_failure_window_end_seconds": 6.0,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot moved the mushroom toward the green plate but opened its gripper prematurely at the edge, causing the mushroom to fall onto the table outside the plate.",
+    "vlm_failure_reason_summary": "premature release; mushroom placed off-target",
+    "vlm_recovery_action": "Re-grasp the mushroom from the table, position the end-effector centered directly over the green plate, and release the object onto the plate surface.",
+    "vlm_recovery_action_summary": "regrasp mushroom; place inside plate",
+    "vlm_prevention_action": "Carry the mushroom fully over the center of the green plate before opening the gripper to release it.",
+    "vlm_prevention_action_summary": "center over plate; release mushroom",
+    "vlm_justification": null,
+    "vlm_temporal_justification": "At timestamp 01:44 (second 104), the robot gripper begins to open prematurely while positioned over the table surface outside the green plate, causing the mushroom to drop onto the table.",
+    "vlm_video_fps": 50.5,
+    "vlm_video_num_frames": 563,
+    "vlm_video_duration_seconds": 11.128712871287128,
+    "vlm_usage": {
+      "input_tokens": 21772,
+      "output_tokens": 314,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"placement_or_insertion_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 5.5,\n \"failure_window_start_seconds\": 5.0,\n \"failure_window_end_seconds\": 6.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot moved the mushroom toward the green plate but opened its gripper prematurely at the edge, causing the mushroom to fall onto the table outside the plate.\",\n \"failure_reason_summary\": \"premature release; mushroom placed off-target\",\n \"recovery_action\": \"Re-grasp the mushroom from the table, position the end-effector centered directly over the green plate, and release the object onto the plate surface.\",\n \"recovery_action_summary\": \"regrasp mushroom; place inside plate\",\n \"prevention_action\": \"Carry the mushroom fully over the center of the green plate before opening the gripper to release it.\",\n \"prevention_action_summary\": \"center over plate; release mushroom\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 230,
+    "failure_sim_step": 230,
+    "failure_policy_step": 38,
+    "failure_chunk_index": 2,
+    "first_post_failure_policy_step": 39,
+    "failure_type": "placement_or_insertion_failure",
+    "correction_action": "Re-grasp the mushroom from the table, position the end-effector centered directly over the green plate, and release the object onto the plate surface."
+  }
+}
\ No newline at end of file
diff --git a/mushroom/ep043/labels.npz b/mushroom/ep043/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..a626a28eedbddeff84730857fa816b69b9756464
--- /dev/null
+++ b/mushroom/ep043/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e6bb208e2dc3dcb55a8b12c7e36ac47c1fb1898530141f986ab1cfee5100f253
+size 3845
diff --git a/mushroom/ep044/example.md b/mushroom/ep044/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..4de583c7a66eef2c283d09dcec4ea53ec726787c
--- /dev/null
+++ b/mushroom/ep044/example.md
@@ -0,0 +1,21 @@
+# t001__ep044__mushroom
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the mushroom on the plate
+- **success:** True
+- **control steps:** 413  (8.2s at 50 Hz)
+- **policy inferences:** 71  (replan_steps=5)
+- **video:** `mushroom/ep044/rollout.mp4`
+- **features:** `71` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–149 | 0–29 | reaching for and grasping mushroom |
+| 3.0–6.0 | 150–299 | 30–59 | placing mushroom onto green plate |
+| 6.0–8.2 | 300–412 | 60–82 | releasing mushroom and retracting arm |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/mushroom/ep044/labels.json b/mushroom/ep044/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..92c4540ae1dc6dd80b48c598f9761abe10710b54
--- /dev/null
+++ b/mushroom/ep044/labels.json
@@ -0,0 +1,68 @@
+{
+  "rollout_id": "t001__ep044__mushroom",
+  "scene_variant": "",
+  "task_id": 1,
+  "task_file": "",
+  "instruction": "Put the mushroom on the plate",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:12:35+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9940119760479043,
+      "control_step_start": 0,
+      "control_step_end": 149,
+      "policy_step_start": 0,
+      "policy_step_end": 29,
+      "phrase": "reaching for and grasping mushroom",
+      "description": "reaching for and grasping mushroom"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9940119760479043,
+      "t_end_sec": 5.9880239520958085,
+      "control_step_start": 150,
+      "control_step_end": 299,
+      "policy_step_start": 30,
+      "policy_step_end": 59,
+      "phrase": "placing mushroom onto green plate",
+      "description": "placing mushroom onto green plate"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.9880239520958085,
+      "t_end_sec": 8.243512974051896,
+      "control_step_start": 300,
+      "control_step_end": 412,
+      "policy_step_start": 60,
+      "policy_step_end": 82,
+      "phrase": "releasing mushroom and retracting arm",
+      "description": "releasing mushroom and retracting arm"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/mushroom/ep044/labels.npz b/mushroom/ep044/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..d11abbee846135091861e0e977dc72965580939d
--- /dev/null
+++ b/mushroom/ep044/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:943e56dd17cb0faf53721ef3acb428dc1a1666a64da7ffcd55e5a2f4e9cc1687
+size 3095
diff --git a/mushroom/ep045/example.md b/mushroom/ep045/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..57ed3b06c648a3229db1586bdf9d3dbd58014dc0
--- /dev/null
+++ b/mushroom/ep045/example.md
@@ -0,0 +1,33 @@
+# t001__ep045__mushroom
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the mushroom on the plate
+- **success:** False
+- **control steps:** 338  (6.8s at 50 Hz)
+- **policy inferences:** 59  (replan_steps=6)
+- **video:** `mushroom/ep045/rollout.mp4`
+- **features:** `59` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–149 | 0–24 | lowering gripper toward mushroom |
+| 3.0–6.0 | 150–299 | 25–49 | grasping mushroom and lifting empty gripper |
+| 6.0–6.8 | 300–337 | 50–56 | hovering empty gripper over plate |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 1.9639278557114228 s (00:02.0)
+- **frame / control_step:** 98 (policy_step=16, chunk_index=2)
+- **coarse onset:** 3.2 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the mushroom but failed to establish a secure grip, causing the mushroom to tip over on the table when the end-effector lifted.
+  Summary: unsecure grasp; mushroom dropped on table
+- **recovery:** Re-align the gripper over the fallen mushroom, grasp it securely by the stem, and move it onto the green plate.
+  Summary: re-align gripper; grasp mushroom; place on plate
+- **prevention:** Lower the gripper further and center it precisely around the mushroom stem before closing the fingers to ensure a firm hold before lifting.
+  Summary: center gripper over stem; secure grasp before lifting
+- **justification:** At around 3.2 seconds, the gripper closes around the mushroom but fails to hold it securely. As the arm lifts at 4.0 seconds, the mushroom slips out and falls onto its side on the table while the empty gripper continues toward the green plate.
diff --git a/mushroom/ep045/labels.json b/mushroom/ep045/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..f7acba7134ad0d595c12e57a508dac0e36ffcde5
--- /dev/null
+++ b/mushroom/ep045/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t001__ep045__mushroom",
+  "scene_variant": "",
+  "task_id": 1,
+  "task_file": "",
+  "instruction": "Put the mushroom on the plate",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:13:10+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.006012024048096,
+      "control_step_start": 0,
+      "control_step_end": 149,
+      "policy_step_start": 0,
+      "policy_step_end": 24,
+      "phrase": "lowering gripper toward mushroom",
+      "description": "lowering gripper toward mushroom"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.006012024048096,
+      "t_end_sec": 6.012024048096192,
+      "control_step_start": 150,
+      "control_step_end": 299,
+      "policy_step_start": 25,
+      "policy_step_end": 49,
+      "phrase": "grasping mushroom and lifting empty gripper",
+      "description": "grasping mushroom and lifting empty gripper"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.012024048096192,
+      "t_end_sec": 6.773547094188377,
+      "control_step_start": 300,
+      "control_step_end": 337,
+      "policy_step_start": 50,
+      "policy_step_end": 56,
+      "phrase": "hovering empty gripper over plate",
+      "description": "hovering empty gripper over plate"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 1.9639278557114228,
+    "vlm_failure_onset_timestamp": "00:02.0",
+    "vlm_failure_onset_frame": 98,
+    "vlm_failure_onset_step": 98,
+    "vlm_coarse_failure_onset_seconds": 3.2,
+    "vlm_failure_window_start_seconds": 3.0,
+    "vlm_failure_window_end_seconds": 4.0,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the mushroom but failed to establish a secure grip, causing the mushroom to tip over on the table when the end-effector lifted.",
+    "vlm_failure_reason_summary": "unsecure grasp; mushroom dropped on table",
+    "vlm_recovery_action": "Re-align the gripper over the fallen mushroom, grasp it securely by the stem, and move it onto the green plate.",
+    "vlm_recovery_action_summary": "re-align gripper; grasp mushroom; place on plate",
+    "vlm_prevention_action": "Lower the gripper further and center it precisely around the mushroom stem before closing the fingers to ensure a firm hold before lifting.",
+    "vlm_prevention_action_summary": "center gripper over stem; secure grasp before lifting",
+    "vlm_justification": "At around 3.2 seconds, the gripper closes around the mushroom but fails to hold it securely. As the arm lifts at 4.0 seconds, the mushroom slips out and falls onto its side on the table while the empty gripper continues toward the green plate.",
+    "vlm_temporal_justification": "At 01:28 into the clip (88 seconds), the gripper closes around the mushroom improperly. Instead of establishing a secure hold, the fingers pinch the side/top, causing the mushroom to visibly tilt sideways and fail to be grasped properly.",
+    "vlm_video_fps": 49.9,
+    "vlm_video_num_frames": 338,
+    "vlm_video_duration_seconds": 6.753507014028056,
+    "vlm_usage": {
+      "input_tokens": 21260,
+      "output_tokens": 401,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n  \"failure_mode\": \"grasp_failure\",\n  \"onset_type\": \"obvious_mistake\",\n  \"failure_onset_seconds\": 3.2,\n  \"failure_window_start_seconds\": 3.0,\n  \"failure_window_end_seconds\": 4.0,\n  \"confidence\": \"high\",\n  \"failure_reason\": \"The robot attempted to grasp the mushroom but failed to establish a secure grip, causing the mushroom to tip over on the table when the end-effector lifted.\",\n  \"failure_reason_summary\": \"unsecure grasp; mushroom dropped on table\",\n  \"recovery_action\": \"Re-align the gripper over the fallen mushroom, grasp it securely by the stem, and move it onto the green plate.\",\n  \"recovery_action_summary\": \"re-align gripper; grasp mushroom; place on plate\",\n  \"prevention_action\": \"Lower the gripper further and center it precisely around the mushroom stem before closing the fingers to ensure a firm hold before lifting.\",\n  \"prevention_action_summary\": \"center gripper over stem; secure grasp before lifting\",\n  \"justification\": \"At around 3.2 seconds, the gripper closes around the mushroom but fails to hold it securely. As the arm lifts at 4.0 seconds, the mushroom slips out and falls onto its side on the table while the empty gripper continues toward the green plate.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 98,
+    "failure_sim_step": 98,
+    "failure_policy_step": 16,
+    "failure_chunk_index": 2,
+    "first_post_failure_policy_step": 17,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-align the gripper over the fallen mushroom, grasp it securely by the stem, and move it onto the green plate."
+  }
+}
\ No newline at end of file
diff --git a/mushroom/ep045/labels.npz b/mushroom/ep045/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..f97c05ebbc2786821f02d830754479d4e048a0f7
--- /dev/null
+++ b/mushroom/ep045/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:59d266efa8418e234008fbd414d69befd545f85dff7d3b47e95cae23ae5fae8f
+size 3794
diff --git a/mushroom/ep046/example.md b/mushroom/ep046/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..1b61e5e619644f0ee52cf8a0d1257fd317cbede8
--- /dev/null
+++ b/mushroom/ep046/example.md
@@ -0,0 +1,34 @@
+# t001__ep046__mushroom
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the mushroom on the plate
+- **success:** False
+- **control steps:** 567  (11.3s at 50 Hz)
+- **policy inferences:** 98  (replan_steps=6)
+- **video:** `mushroom/ep046/rollout.mp4`
+- **features:** `98` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | lowering gripper toward mushroom |
+| 3.0–6.0 | 151–301 | 25–50 | attempting grasp on mushroom |
+| 6.0–9.0 | 302–452 | 50–75 | pressing gripper against mushroom |
+| 9.0–11.3 | 453–566 | 75–94 | remaining stationary near mushroom |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 6.832669322709163 s (00:06.8)
+- **frame / control_step:** 343 (policy_step=57, chunk_index=1)
+- **coarse onset:** 4.5 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the mushroom, but failed to secure a proper grip on it, leaving the mushroom behind on the table when moving.
+  Summary: failed grasp attempt; mushroom not secured
+- **recovery:** Re-open the end-effector fingers, reposition directly above the mushroom, execute a secure grasp, and place it onto the green plate.
+  Summary: open gripper; regrasp mushroom; place on plate
+- **prevention:** Align the end-effector accurately with the mushroom's center of mass before closing the gripper to ensure a firm hold.
+  Summary: center gripper over mushroom; secure grasp; place on plate
+- **justification:** At around 4.5 seconds (frames 00:04 to 00:05), the robot closes its gripper on the mushroom but fails to hold it, attempting to continue the motion without the object in hand.
diff --git a/mushroom/ep046/labels.json b/mushroom/ep046/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..2b29d81d7641d809ab157dc2a9c4742d16c917ce
--- /dev/null
+++ b/mushroom/ep046/labels.json
@@ -0,0 +1,109 @@
+{
+  "rollout_id": "t001__ep046__mushroom",
+  "scene_variant": "",
+  "task_id": 1,
+  "task_file": "",
+  "instruction": "Put the mushroom on the plate",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:13:53+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00796812749004,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "lowering gripper toward mushroom",
+      "description": "lowering gripper toward mushroom"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00796812749004,
+      "t_end_sec": 6.01593625498008,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "attempting grasp on mushroom",
+      "description": "attempting grasp on mushroom"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01593625498008,
+      "t_end_sec": 9.02390438247012,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "pressing gripper against mushroom",
+      "description": "pressing gripper against mushroom"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.02390438247012,
+      "t_end_sec": 11.294820717131474,
+      "control_step_start": 453,
+      "control_step_end": 566,
+      "policy_step_start": 75,
+      "policy_step_end": 94,
+      "phrase": "remaining stationary near mushroom",
+      "description": "remaining stationary near mushroom"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 6.832669322709163,
+    "vlm_failure_onset_timestamp": "00:06.8",
+    "vlm_failure_onset_frame": 343,
+    "vlm_failure_onset_step": 343,
+    "vlm_coarse_failure_onset_seconds": 4.5,
+    "vlm_failure_window_start_seconds": 3.5,
+    "vlm_failure_window_end_seconds": 5.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the mushroom, but failed to secure a proper grip on it, leaving the mushroom behind on the table when moving.",
+    "vlm_failure_reason_summary": "failed grasp attempt; mushroom not secured",
+    "vlm_recovery_action": "Re-open the end-effector fingers, reposition directly above the mushroom, execute a secure grasp, and place it onto the green plate.",
+    "vlm_recovery_action_summary": "open gripper; regrasp mushroom; place on plate",
+    "vlm_prevention_action": "Align the end-effector accurately with the mushroom's center of mass before closing the gripper to ensure a firm hold.",
+    "vlm_prevention_action_summary": "center gripper over mushroom; secure grasp; place on plate",
+    "vlm_justification": "At around 4.5 seconds (frames 00:04 to 00:05), the robot closes its gripper on the mushroom but fails to hold it, attempting to continue the motion without the object in hand.",
+    "vlm_temporal_justification": "At clip time 04:28 (268 seconds), the robot's gripper fingers fully close together without securing the mushroom, leaving the mushroom behind on the table.",
+    "vlm_video_fps": 50.2,
+    "vlm_video_num_frames": 567,
+    "vlm_video_duration_seconds": 11.274900398406373,
+    "vlm_usage": {
+      "input_tokens": 21637,
+      "output_tokens": 366,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 4.5,\n \"failure_window_start_seconds\": 3.5,\n \"failure_window_end_seconds\": 5.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the mushroom, but failed to secure a proper grip on it, leaving the mushroom behind on the table when moving.\",\n \"failure_reason_summary\": \"failed grasp attempt; mushroom not secured\",\n \"recovery_action\": \"Re-open the end-effector fingers, reposition directly above the mushroom, execute a secure grasp, and place it onto the green plate.\",\n \"recovery_action_summary\": \"open gripper; regrasp mushroom; place on plate\",\n \"prevention_action\": \"Align the end-effector accurately with the mushroom's center of mass before closing the gripper to ensure a firm hold.\",\n \"prevention_action_summary\": \"center gripper over mushroom; secure grasp; place on plate\",\n \"justification\": \"At around 4.5 seconds (frames 00:04 to 00:05), the robot closes its gripper on the mushroom but fails to hold it, attempting to continue the motion without the object in hand.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 343,
+    "failure_sim_step": 343,
+    "failure_policy_step": 57,
+    "failure_chunk_index": 1,
+    "first_post_failure_policy_step": 58,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-open the end-effector fingers, reposition directly above the mushroom, execute a secure grasp, and place it onto the green plate."
+  }
+}
\ No newline at end of file
diff --git a/mushroom/ep046/labels.npz b/mushroom/ep046/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..a11c61b456663c88ae84b6919699ace3e0f2e157
--- /dev/null
+++ b/mushroom/ep046/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bd55c0e6850040c0793f3aed622e1512db987cda87f01df6a9c8c3c72d891e09
+size 3831
diff --git a/mushroom/ep047/example.md b/mushroom/ep047/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..bc1777767751eac8215dca879879c53d6d2cdcd9
--- /dev/null
+++ b/mushroom/ep047/example.md
@@ -0,0 +1,22 @@
+# t001__ep047__mushroom
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the mushroom on the plate
+- **success:** True
+- **control steps:** 507  (10.1s at 50 Hz)
+- **policy inferences:** 88  (replan_steps=6)
+- **video:** `mushroom/ep047/rollout.mp4`
+- **features:** `88` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | reaching for the mushroom |
+| 3.0–6.0 | 151–301 | 25–50 | grasping the mushroom |
+| 6.0–9.0 | 302–452 | 50–75 | placing mushroom on plate |
+| 9.0–10.1 | 453–506 | 75–84 | retracting arm |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/mushroom/ep047/labels.json b/mushroom/ep047/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..d39a0eff29fea18122efb6eaad61ec4a87c9e6f4
--- /dev/null
+++ b/mushroom/ep047/labels.json
@@ -0,0 +1,79 @@
+{
+  "rollout_id": "t001__ep047__mushroom",
+  "scene_variant": "",
+  "task_id": 1,
+  "task_file": "",
+  "instruction": "Put the mushroom on the plate",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:14:01+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00796812749004,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for the mushroom",
+      "description": "reaching for the mushroom"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00796812749004,
+      "t_end_sec": 6.01593625498008,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "grasping the mushroom",
+      "description": "grasping the mushroom"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01593625498008,
+      "t_end_sec": 9.02390438247012,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "placing mushroom on plate",
+      "description": "placing mushroom on plate"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.02390438247012,
+      "t_end_sec": 10.099601593625497,
+      "control_step_start": 453,
+      "control_step_end": 506,
+      "policy_step_start": 75,
+      "policy_step_end": 84,
+      "phrase": "retracting arm",
+      "description": "retracting arm"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/mushroom/ep047/labels.npz b/mushroom/ep047/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..d820a3ff9f67854ab7857b1165124b2289f7f315
--- /dev/null
+++ b/mushroom/ep047/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f861e13e0090b3cdeb2068022b069b167e276c50a04f3120a5956f82f0f3d288
+size 3105
diff --git a/mushroom/ep048/example.md b/mushroom/ep048/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..e64867db952d56742aed2fe500674795bcff0724
--- /dev/null
+++ b/mushroom/ep048/example.md
@@ -0,0 +1,22 @@
+# t001__ep048__mushroom
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the mushroom on the plate
+- **success:** True
+- **control steps:** 493  (9.9s at 50 Hz)
+- **policy inferences:** 86  (replan_steps=6)
+- **video:** `mushroom/ep048/rollout.mp4`
+- **features:** `86` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–149 | 0–24 | reaching for mushroom |
+| 3.0–6.0 | 150–299 | 25–49 | grasping and lifting mushroom |
+| 6.0–9.0 | 300–449 | 50–74 | moving mushroom to plate |
+| 9.0–9.9 | 450–492 | 75–82 | releasing mushroom and retracting |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/mushroom/ep048/labels.json b/mushroom/ep048/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..01279d84a256ee0508689e72f8097660688ab0e8
--- /dev/null
+++ b/mushroom/ep048/labels.json
@@ -0,0 +1,79 @@
+{
+  "rollout_id": "t001__ep048__mushroom",
+  "scene_variant": "",
+  "task_id": 1,
+  "task_file": "",
+  "instruction": "Put the mushroom on the plate",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:14:09+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0,
+      "control_step_start": 0,
+      "control_step_end": 149,
+      "policy_step_start": 0,
+      "policy_step_end": 24,
+      "phrase": "reaching for mushroom",
+      "description": "reaching for mushroom"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0,
+      "t_end_sec": 6.0,
+      "control_step_start": 150,
+      "control_step_end": 299,
+      "policy_step_start": 25,
+      "policy_step_end": 49,
+      "phrase": "grasping and lifting mushroom",
+      "description": "grasping and lifting mushroom"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.0,
+      "t_end_sec": 9.0,
+      "control_step_start": 300,
+      "control_step_end": 449,
+      "policy_step_start": 50,
+      "policy_step_end": 74,
+      "phrase": "moving mushroom to plate",
+      "description": "moving mushroom to plate"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.0,
+      "t_end_sec": 9.86,
+      "control_step_start": 450,
+      "control_step_end": 492,
+      "policy_step_start": 75,
+      "policy_step_end": 82,
+      "phrase": "releasing mushroom and retracting",
+      "description": "releasing mushroom and retracting"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/mushroom/ep048/labels.npz b/mushroom/ep048/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..ef8c8515f19cd96ad69776aebfd0534ef9e4cfcf
--- /dev/null
+++ b/mushroom/ep048/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b6b23e5d5b6e9effdc67563690fa95406d3a4b5f83f35c2f4e2c704dbb97ff12
+size 3115
diff --git a/pineapple_out_of_drawer/ep000/example.md b/pineapple_out_of_drawer/ep000/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..ef48fc65c1031693ee0e0b05fcbab43d55aabc5f
--- /dev/null
+++ b/pineapple_out_of_drawer/ep000/example.md
@@ -0,0 +1,34 @@
+# t007__ep000__pineapple_out_of_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the pineapple out of the drawer
+- **success:** False
+- **control steps:** 573  (11.2s at 51 Hz)
+- **policy inferences:** 98  (replan_steps=6)
+- **video:** `pineapple_out_of_drawer/ep000/rollout.mp4`
+- **features:** `98` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–152 | 0–25 | lowering arm near base |
+| 3.0–6.0 | 153–305 | 25–50 | hovering near central base |
+| 6.0–9.0 | 306–458 | 51–76 | oscillating in place near base |
+| 9.0–11.2 | 459–572 | 76–95 | remaining stationary near base |
+
+## Failure
+
+- **yes**
+- **onset type:** operator_intervention
+- **onset:** 2.6666666666666665 s (00:02.7)
+- **frame / control_step:** 136 (policy_step=22, chunk_index=4)
+- **coarse onset:** 4.0 s (refined=True)
+- **mode:** stuck_or_no_progress
+- **reason:** The robot arm gets stuck moving down and oscillating around its base instead of extending rightward toward the drawer to pick up the pineapple.
+  Summary: robot hovering near base; no progress toward drawer
+- **recovery:** Re-orient the end-effector and command the arm to reach toward the cyan drawer box on the right.
+  Summary: reach toward drawer; extract pineapple
+- **prevention:** Plan a direct motion path from the initial pose toward the drawer on the right without hovering at the base.
+  Summary: move directly to drawer; avoid idling
+- **justification:** None
diff --git a/pineapple_out_of_drawer/ep000/labels.json b/pineapple_out_of_drawer/ep000/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..67a4005f632d40b63d5de959fc3ade33f1880532
--- /dev/null
+++ b/pineapple_out_of_drawer/ep000/labels.json
@@ -0,0 +1,109 @@
+{
+  "rollout_id": "t007__ep000__pineapple_out_of_drawer",
+  "scene_variant": "",
+  "task_id": 7,
+  "task_file": "",
+  "instruction": "Take the pineapple out of the drawer",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:14:40+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0,
+      "control_step_start": 0,
+      "control_step_end": 152,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "lowering arm near base",
+      "description": "lowering arm near base"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0,
+      "t_end_sec": 6.0,
+      "control_step_start": 153,
+      "control_step_end": 305,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "hovering near central base",
+      "description": "hovering near central base"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.0,
+      "t_end_sec": 9.0,
+      "control_step_start": 306,
+      "control_step_end": 458,
+      "policy_step_start": 51,
+      "policy_step_end": 76,
+      "phrase": "oscillating in place near base",
+      "description": "oscillating in place near base"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.0,
+      "t_end_sec": 11.235294117647058,
+      "control_step_start": 459,
+      "control_step_end": 572,
+      "policy_step_start": 76,
+      "policy_step_end": 95,
+      "phrase": "remaining stationary near base",
+      "description": "remaining stationary near base"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "stuck_or_no_progress",
+    "vlm_failure_onset_type": "operator_intervention",
+    "vlm_failure_onset_seconds": 2.6666666666666665,
+    "vlm_failure_onset_timestamp": "00:02.7",
+    "vlm_failure_onset_frame": 136,
+    "vlm_failure_onset_step": 136,
+    "vlm_coarse_failure_onset_seconds": 4.0,
+    "vlm_failure_window_start_seconds": 3.5,
+    "vlm_failure_window_end_seconds": 4.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot arm gets stuck moving down and oscillating around its base instead of extending rightward toward the drawer to pick up the pineapple.",
+    "vlm_failure_reason_summary": "robot hovering near base; no progress toward drawer",
+    "vlm_recovery_action": "Re-orient the end-effector and command the arm to reach toward the cyan drawer box on the right.",
+    "vlm_recovery_action_summary": "reach toward drawer; extract pineapple",
+    "vlm_prevention_action": "Plan a direct motion path from the initial pose toward the drawer on the right without hovering at the base.",
+    "vlm_prevention_action_summary": "move directly to drawer; avoid idling",
+    "vlm_justification": null,
+    "vlm_temporal_justification": "At 01:25 (85 seconds into the clip), the robot arm begins moving downward and folding in toward its own base instead of extending rightward toward the drawer containing the pineapple, marking the onset of the failure trajectory.",
+    "vlm_video_fps": 51.0,
+    "vlm_video_num_frames": 573,
+    "vlm_video_duration_seconds": 11.215686274509803,
+    "vlm_usage": {
+      "input_tokens": 21894,
+      "output_tokens": 308,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"stuck_or_no_progress\",\n \"onset_type\": \"operator_intervention\",\n \"failure_onset_seconds\": 4.0,\n \"failure_window_start_seconds\": 3.5,\n \"failure_window_end_seconds\": 4.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot arm gets stuck moving down and oscillating around its base instead of extending rightward toward the drawer to pick up the pineapple.\",\n \"failure_reason_summary\": \"robot hovering near base; no progress toward drawer\",\n \"recovery_action\": \"Re-orient the end-effector and command the arm to reach toward the cyan drawer box on the right.\",\n \"recovery_action_summary\": \"reach toward drawer; extract pineapple\",\n \"prevention_action\": \"Plan a direct motion path from the initial pose toward the drawer on the right without hovering at the base.\",\n \"prevention_action_summary\": \"move directly to drawer; avoid idling\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 136,
+    "failure_sim_step": 136,
+    "failure_policy_step": 22,
+    "failure_chunk_index": 4,
+    "first_post_failure_policy_step": 23,
+    "failure_type": "stuck_or_no_progress",
+    "correction_action": "Re-orient the end-effector and command the arm to reach toward the cyan drawer box on the right."
+  }
+}
\ No newline at end of file
diff --git a/pineapple_out_of_drawer/ep000/labels.npz b/pineapple_out_of_drawer/ep000/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..fb80a541333b837191a78c7e273cf730beded986
--- /dev/null
+++ b/pineapple_out_of_drawer/ep000/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f9a3873e840a41a432003d717800a965ec23c2cb7de0428635bd931bf343af1a
+size 3764
diff --git a/pineapple_out_of_drawer/ep001/example.md b/pineapple_out_of_drawer/ep001/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..2a091000a3fc0e30d98ccf97561010571aa31a76
--- /dev/null
+++ b/pineapple_out_of_drawer/ep001/example.md
@@ -0,0 +1,37 @@
+# t007__ep001__pineapple_out_of_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the pineapple out of the drawer
+- **success:** False
+- **control steps:** 975  (19.3s at 50 Hz)
+- **policy inferences:** 167  (replan_steps=6)
+- **video:** `pineapple_out_of_drawer/ep001/rollout.mp4`
+- **features:** `167` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching toward pineapple in drawer |
+| 3.0–6.0 | 152–303 | 25–50 | closing gripper on pineapple leaves |
+| 6.0–9.0 | 304–455 | 50–75 | hovering above drawer empty handed |
+| 9.0–12.0 | 456–607 | 76–101 | remaining stationary above drawer |
+| 12.0–15.0 | 608–759 | 101–126 | idle hovering above drawer |
+| 15.0–18.1 | 760–911 | 126–151 | idle hovering above drawer |
+| 18.1–19.3 | 912–974 | 152–162 | remaining stationary above drawer |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 1.504950495049505 s (00:01.5)
+- **frame / control_step:** 76 (policy_step=12, chunk_index=4)
+- **coarse onset:** 4.0 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the pineapple by its leaves, but the gripper closed without securing the object, leaving the pineapple inside the container.
+  Summary: grasp attempt missed object; pineapple remained in drawer
+- **recovery:** Realign the end-effector over the pineapple inside the drawer, lower to surround its main body, close the gripper securely, and lift it out.
+  Summary: realign gripper; grasp pineapple firmly; lift out of drawer
+- **prevention:** Position the gripper lower and align it directly with the pineapple's body rather than pinching near the top leaves before closing.
+  Summary: lower grasp pose; align gripper with pineapple body; close gripper firmly
+- **justification:** Around 00:04, the robot closes its gripper on the pineapple leaves but fails to establish a secure hold. It then lifts away empty and remains stationary above the drawer for the rest of the episode.
diff --git a/pineapple_out_of_drawer/ep001/labels.json b/pineapple_out_of_drawer/ep001/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..22bf420b45fede465377b84a294d99d457cb012f
--- /dev/null
+++ b/pineapple_out_of_drawer/ep001/labels.json
@@ -0,0 +1,142 @@
+{
+  "rollout_id": "t007__ep001__pineapple_out_of_drawer",
+  "scene_variant": "",
+  "task_id": 7,
+  "task_file": "",
+  "instruction": "Take the pineapple out of the drawer",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:15:12+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00990099009901,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching toward pineapple in drawer",
+      "description": "reaching toward pineapple in drawer"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00990099009901,
+      "t_end_sec": 6.01980198019802,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "closing gripper on pineapple leaves",
+      "description": "closing gripper on pineapple leaves"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01980198019802,
+      "t_end_sec": 9.029702970297029,
+      "control_step_start": 304,
+      "control_step_end": 455,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "hovering above drawer empty handed",
+      "description": "hovering above drawer empty handed"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.029702970297029,
+      "t_end_sec": 12.03960396039604,
+      "control_step_start": 456,
+      "control_step_end": 607,
+      "policy_step_start": 76,
+      "policy_step_end": 101,
+      "phrase": "remaining stationary above drawer",
+      "description": "remaining stationary above drawer"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 12.03960396039604,
+      "t_end_sec": 15.049504950495049,
+      "control_step_start": 608,
+      "control_step_end": 759,
+      "policy_step_start": 101,
+      "policy_step_end": 126,
+      "phrase": "idle hovering above drawer",
+      "description": "idle hovering above drawer"
+    },
+    {
+      "segment_index": 5,
+      "t_start_sec": 15.049504950495049,
+      "t_end_sec": 18.059405940594058,
+      "control_step_start": 760,
+      "control_step_end": 911,
+      "policy_step_start": 126,
+      "policy_step_end": 151,
+      "phrase": "idle hovering above drawer",
+      "description": "idle hovering above drawer"
+    },
+    {
+      "segment_index": 6,
+      "t_start_sec": 18.059405940594058,
+      "t_end_sec": 19.306930693069308,
+      "control_step_start": 912,
+      "control_step_end": 974,
+      "policy_step_start": 152,
+      "policy_step_end": 162,
+      "phrase": "remaining stationary above drawer",
+      "description": "remaining stationary above drawer"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 1.504950495049505,
+    "vlm_failure_onset_timestamp": "00:01.5",
+    "vlm_failure_onset_frame": 76,
+    "vlm_failure_onset_step": 76,
+    "vlm_coarse_failure_onset_seconds": 4.0,
+    "vlm_failure_window_start_seconds": 3.5,
+    "vlm_failure_window_end_seconds": 5.0,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the pineapple by its leaves, but the gripper closed without securing the object, leaving the pineapple inside the container.",
+    "vlm_failure_reason_summary": "grasp attempt missed object; pineapple remained in drawer",
+    "vlm_recovery_action": "Realign the end-effector over the pineapple inside the drawer, lower to surround its main body, close the gripper securely, and lift it out.",
+    "vlm_recovery_action_summary": "realign gripper; grasp pineapple firmly; lift out of drawer",
+    "vlm_prevention_action": "Position the gripper lower and align it directly with the pineapple's body rather than pinching near the top leaves before closing.",
+    "vlm_prevention_action_summary": "lower grasp pose; align gripper with pineapple body; close gripper firmly",
+    "vlm_justification": "Around 00:04, the robot closes its gripper on the pineapple leaves but fails to establish a secure hold. It then lifts away empty and remains stationary above the drawer for the rest of the episode.",
+    "vlm_temporal_justification": "At second 26 of this clip, the gripper closes around the tip of the pineapple's leaves without securing a firm grip, causing the pineapple to tip over inside the container rather than being successfully grasped.",
+    "vlm_video_fps": 50.5,
+    "vlm_video_num_frames": 975,
+    "vlm_video_duration_seconds": 19.287128712871286,
+    "vlm_usage": {
+      "input_tokens": 22271,
+      "output_tokens": 372,
+      "cached_tokens": 1293
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 4.0,\n \"failure_window_start_seconds\": 3.5,\n \"failure_window_end_seconds\": 5.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the pineapple by its leaves, but the gripper closed without securing the object, leaving the pineapple inside the container.\",\n \"failure_reason_summary\": \"grasp attempt missed object; pineapple remained in drawer\",\n \"recovery_action\": \"Realign the end-effector over the pineapple inside the drawer, lower to surround its main body, close the gripper securely, and lift it out.\",\n \"recovery_action_summary\": \"realign gripper; grasp pineapple firmly; lift out of drawer\",\n \"prevention_action\": \"Position the gripper lower and align it directly with the pineapple's body rather than pinching near the top leaves before closing.\",\n \"prevention_action_summary\": \"lower grasp pose; align gripper with pineapple body; close gripper firmly\",\n \"justification\": \"Around 00:04, the robot closes its gripper on the pineapple leaves but fails to establish a secure hold. It then lifts away empty and remains stationary above the drawer for the rest of the episode.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 76,
+    "failure_sim_step": 76,
+    "failure_policy_step": 12,
+    "failure_chunk_index": 4,
+    "first_post_failure_policy_step": 13,
+    "failure_type": "grasp_failure",
+    "correction_action": "Realign the end-effector over the pineapple inside the drawer, lower to surround its main body, close the gripper securely, and lift it out."
+  }
+}
\ No newline at end of file
diff --git a/pineapple_out_of_drawer/ep001/labels.npz b/pineapple_out_of_drawer/ep001/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..c9a3874c278bc20500f1e1cc441cd7e9a55a0028
--- /dev/null
+++ b/pineapple_out_of_drawer/ep001/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9cf0ab84aed049d2c1722dae46bb52ec9829cfc214e5d69ed7332c95a677985e
+size 4003
diff --git a/pineapple_out_of_drawer/ep002/example.md b/pineapple_out_of_drawer/ep002/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..03e442eadd23d01f06ded12dae6cb4cb39cf2723
--- /dev/null
+++ b/pineapple_out_of_drawer/ep002/example.md
@@ -0,0 +1,34 @@
+# t007__ep002__pineapple_out_of_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the pineapple out of the drawer
+- **success:** False
+- **control steps:** 502  (10.0s at 50 Hz)
+- **policy inferences:** 87  (replan_steps=6)
+- **video:** `pineapple_out_of_drawer/ep002/rollout.mp4`
+- **features:** `87` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | approaching pineapple in drawer |
+| 3.0–6.0 | 151–301 | 25–50 | attempting grasp on pineapple |
+| 6.0–9.0 | 302–452 | 50–75 | hovering empty gripper |
+| 9.0–10.0 | 453–501 | 75–83 | idle above drawer |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 4.234592445328032 s (00:04.2)
+- **frame / control_step:** 213 (policy_step=35, chunk_index=3)
+- **coarse onset:** 4.5 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot gripper closed around the top edge of the pineapple without making firm contact, failing to secure it before lifting the arm empty.
+  Summary: failed grasp on pineapple; lifting empty gripper
+- **recovery:** Lower the end effector back into the open drawer, position the gripper around the middle of the pineapple, close the fingers firmly, and lift it clear of the drawer.
+  Summary: re-align gripper; grasp pineapple securely; lift out of drawer
+- **prevention:** Descend deeper into the drawer before closing the gripper fingers to ensure a reliable grip on the pineapple.
+  Summary: descend deeper into drawer; grasp pineapple securely
+- **justification:** At around 4.5 seconds, the robot attempts to grasp the pineapple in the drawer, but the gripper closes without acquiring the object and lifts away empty, causing the task to fail.
diff --git a/pineapple_out_of_drawer/ep002/labels.json b/pineapple_out_of_drawer/ep002/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..44382e5e139a365af9de7906ceaa9ab0c6c1954b
--- /dev/null
+++ b/pineapple_out_of_drawer/ep002/labels.json
@@ -0,0 +1,109 @@
+{
+  "rollout_id": "t007__ep002__pineapple_out_of_drawer",
+  "scene_variant": "",
+  "task_id": 7,
+  "task_file": "",
+  "instruction": "Take the pineapple out of the drawer",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:15:46+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.001988071570577,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "approaching pineapple in drawer",
+      "description": "approaching pineapple in drawer"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.001988071570577,
+      "t_end_sec": 6.003976143141154,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "attempting grasp on pineapple",
+      "description": "attempting grasp on pineapple"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.003976143141154,
+      "t_end_sec": 9.00596421471173,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "hovering empty gripper",
+      "description": "hovering empty gripper"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.00596421471173,
+      "t_end_sec": 9.980119284294235,
+      "control_step_start": 453,
+      "control_step_end": 501,
+      "policy_step_start": 75,
+      "policy_step_end": 83,
+      "phrase": "idle above drawer",
+      "description": "idle above drawer"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 4.234592445328032,
+    "vlm_failure_onset_timestamp": "00:04.2",
+    "vlm_failure_onset_frame": 213,
+    "vlm_failure_onset_step": 213,
+    "vlm_coarse_failure_onset_seconds": 4.5,
+    "vlm_failure_window_start_seconds": 4.0,
+    "vlm_failure_window_end_seconds": 5.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot gripper closed around the top edge of the pineapple without making firm contact, failing to secure it before lifting the arm empty.",
+    "vlm_failure_reason_summary": "failed grasp on pineapple; lifting empty gripper",
+    "vlm_recovery_action": "Lower the end effector back into the open drawer, position the gripper around the middle of the pineapple, close the fingers firmly, and lift it clear of the drawer.",
+    "vlm_recovery_action_summary": "re-align gripper; grasp pineapple securely; lift out of drawer",
+    "vlm_prevention_action": "Descend deeper into the drawer before closing the gripper fingers to ensure a reliable grip on the pineapple.",
+    "vlm_prevention_action_summary": "descend deeper into drawer; grasp pineapple securely",
+    "vlm_justification": "At around 4.5 seconds, the robot attempts to grasp the pineapple in the drawer, but the gripper closes without acquiring the object and lifts away empty, causing the task to fail.",
+    "vlm_temporal_justification": "At 02:18 (clip second 138), the robot gripper closes around only the top edge/leaves of the pineapple without grasping its body, failing to secure the object before the arm lifts away empty.",
+    "vlm_video_fps": 50.3,
+    "vlm_video_num_frames": 502,
+    "vlm_video_duration_seconds": 9.96023856858847,
+    "vlm_usage": {
+      "input_tokens": 21574,
+      "output_tokens": 366,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 4.5,\n \"failure_window_start_seconds\": 4.0,\n \"failure_window_end_seconds\": 5.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot gripper closed around the top edge of the pineapple without making firm contact, failing to secure it before lifting the arm empty.\",\n \"failure_reason_summary\": \"failed grasp on pineapple; lifting empty gripper\",\n \"recovery_action\": \"Lower the end effector back into the open drawer, position the gripper around the middle of the pineapple, close the fingers firmly, and lift it clear of the drawer.\",\n \"recovery_action_summary\": \"re-align gripper; grasp pineapple securely; lift out of drawer\",\n \"prevention_action\": \"Descend deeper into the drawer before closing the gripper fingers to ensure a reliable grip on the pineapple.\",\n \"prevention_action_summary\": \"descend deeper into drawer; grasp pineapple securely\",\n \"justification\": \"At around 4.5 seconds, the robot attempts to grasp the pineapple in the drawer, but the gripper closes without acquiring the object and lifts away empty, causing the task to fail.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 213,
+    "failure_sim_step": 213,
+    "failure_policy_step": 35,
+    "failure_chunk_index": 3,
+    "first_post_failure_policy_step": 36,
+    "failure_type": "grasp_failure",
+    "correction_action": "Lower the end effector back into the open drawer, position the gripper around the middle of the pineapple, close the fingers firmly, and lift it clear of the drawer."
+  }
+}
\ No newline at end of file
diff --git a/pineapple_out_of_drawer/ep002/labels.npz b/pineapple_out_of_drawer/ep002/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..cba51b4e82d01bf8dd0d3de52d3e71c034de7361
--- /dev/null
+++ b/pineapple_out_of_drawer/ep002/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c0b5b309b6714bd57032cc6fcec9b11c62a29aed5d03792da47af517985af8b0
+size 3853
diff --git a/pineapple_out_of_drawer/ep003/example.md b/pineapple_out_of_drawer/ep003/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..802deb283b6391344d7e4efe7e06b2070284a693
--- /dev/null
+++ b/pineapple_out_of_drawer/ep003/example.md
@@ -0,0 +1,35 @@
+# t007__ep003__pineapple_out_of_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the pineapple out of the drawer
+- **success:** False
+- **control steps:** 626  (12.3s at 51 Hz)
+- **policy inferences:** 108  (replan_steps=6)
+- **video:** `pineapple_out_of_drawer/ep003/rollout.mp4`
+- **features:** `108` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–152 | 0–25 | reaching toward pineapple in drawer |
+| 3.0–6.0 | 153–305 | 25–50 | attempting grasp on pineapple |
+| 6.0–9.0 | 306–458 | 51–76 | lifting empty gripper up |
+| 9.0–12.0 | 459–611 | 76–101 | hovering empty above drawer |
+| 12.0–12.3 | 612–625 | 102–104 | remaining stationary above table |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 5.627450980392157 s (00:05.6)
+- **frame / control_step:** 287 (policy_step=47, chunk_index=5)
+- **coarse onset:** 5.5 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the pineapple in the drawer, but the gripper closed improperly without securing the object, causing the arm to lift away empty.
+  Summary: failed grasp on pineapple; lifting empty gripper
+- **recovery:** Lower the gripper back onto the pineapple, align the fingers securely around its main body, close the gripper tightly, and lift the pineapple out of the drawer.
+  Summary: re-align gripper; grasp pineapple securely; lift out of drawer
+- **prevention:** Descend deeper over the pineapple and center the gripper around the pineapple's main body before closing the fingers to establish a secure hold.
+  Summary: center gripper on pineapple body; grasp firmly before lifting
+- **justification:** Around 5.5 seconds into the rollout, the gripper attempts to close on the pineapple inside the drawer box but fails to acquire a grip. It lifts up empty at 6.0 seconds and fails to retrieve the object.
diff --git a/pineapple_out_of_drawer/ep003/labels.json b/pineapple_out_of_drawer/ep003/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..bfe8e922116a5e298c47fabe71db2df245053ff1
--- /dev/null
+++ b/pineapple_out_of_drawer/ep003/labels.json
@@ -0,0 +1,120 @@
+{
+  "rollout_id": "t007__ep003__pineapple_out_of_drawer",
+  "scene_variant": "",
+  "task_id": 7,
+  "task_file": "",
+  "instruction": "Take the pineapple out of the drawer",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:16:25+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0,
+      "control_step_start": 0,
+      "control_step_end": 152,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching toward pineapple in drawer",
+      "description": "reaching toward pineapple in drawer"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0,
+      "t_end_sec": 6.0,
+      "control_step_start": 153,
+      "control_step_end": 305,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "attempting grasp on pineapple",
+      "description": "attempting grasp on pineapple"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.0,
+      "t_end_sec": 9.0,
+      "control_step_start": 306,
+      "control_step_end": 458,
+      "policy_step_start": 51,
+      "policy_step_end": 76,
+      "phrase": "lifting empty gripper up",
+      "description": "lifting empty gripper up"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.0,
+      "t_end_sec": 12.0,
+      "control_step_start": 459,
+      "control_step_end": 611,
+      "policy_step_start": 76,
+      "policy_step_end": 101,
+      "phrase": "hovering empty above drawer",
+      "description": "hovering empty above drawer"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 12.0,
+      "t_end_sec": 12.27450980392157,
+      "control_step_start": 612,
+      "control_step_end": 625,
+      "policy_step_start": 102,
+      "policy_step_end": 104,
+      "phrase": "remaining stationary above table",
+      "description": "remaining stationary above table"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 5.627450980392157,
+    "vlm_failure_onset_timestamp": "00:05.6",
+    "vlm_failure_onset_frame": 287,
+    "vlm_failure_onset_step": 287,
+    "vlm_coarse_failure_onset_seconds": 5.5,
+    "vlm_failure_window_start_seconds": 4.5,
+    "vlm_failure_window_end_seconds": 6.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the pineapple in the drawer, but the gripper closed improperly without securing the object, causing the arm to lift away empty.",
+    "vlm_failure_reason_summary": "failed grasp on pineapple; lifting empty gripper",
+    "vlm_recovery_action": "Lower the gripper back onto the pineapple, align the fingers securely around its main body, close the gripper tightly, and lift the pineapple out of the drawer.",
+    "vlm_recovery_action_summary": "re-align gripper; grasp pineapple securely; lift out of drawer",
+    "vlm_prevention_action": "Descend deeper over the pineapple and center the gripper around the pineapple's main body before closing the fingers to establish a secure hold.",
+    "vlm_prevention_action_summary": "center gripper on pineapple body; grasp firmly before lifting",
+    "vlm_justification": "Around 5.5 seconds into the rollout, the gripper attempts to close on the pineapple inside the drawer box but fails to acquire a grip. It lifts up empty at 6.0 seconds and fails to retrieve the object.",
+    "vlm_temporal_justification": "At clip second 02:40 (frame 287), the robot gripper completes its closure on the tip of the pineapple's leaves rather than securing the main body or base of the object, resulting in an improper grasp that fails to lift the pineapple.",
+    "vlm_video_fps": 51.0,
+    "vlm_video_num_frames": 626,
+    "vlm_video_duration_seconds": 12.254901960784315,
+    "vlm_usage": {
+      "input_tokens": 21960,
+      "output_tokens": 392,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 5.5,\n \"failure_window_start_seconds\": 4.5,\n \"failure_window_end_seconds\": 6.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the pineapple in the drawer, but the gripper closed improperly without securing the object, causing the arm to lift away empty.\",\n \"failure_reason_summary\": \"failed grasp on pineapple; lifting empty gripper\",\n \"recovery_action\": \"Lower the gripper back onto the pineapple, align the fingers securely around its main body, close the gripper tightly, and lift the pineapple out of the drawer.\",\n \"recovery_action_summary\": \"re-align gripper; grasp pineapple securely; lift out of drawer\",\n \"prevention_action\": \"Descend deeper over the pineapple and center the gripper around the pineapple's main body before closing the fingers to establish a secure hold.\",\n \"prevention_action_summary\": \"center gripper on pineapple body; grasp firmly before lifting\",\n \"justification\": \"Around 5.5 seconds into the rollout, the gripper attempts to close on the pineapple inside the drawer box but fails to acquire a grip. It lifts up empty at 6.0 seconds and fails to retrieve the object.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 287,
+    "failure_sim_step": 287,
+    "failure_policy_step": 47,
+    "failure_chunk_index": 5,
+    "first_post_failure_policy_step": 48,
+    "failure_type": "grasp_failure",
+    "correction_action": "Lower the gripper back onto the pineapple, align the fingers securely around its main body, close the gripper tightly, and lift the pineapple out of the drawer."
+  }
+}
\ No newline at end of file
diff --git a/pineapple_out_of_drawer/ep003/labels.npz b/pineapple_out_of_drawer/ep003/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..619d2f550e3a9e7f7f42faf10c37aedf7de3c103
--- /dev/null
+++ b/pineapple_out_of_drawer/ep003/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9f91de4b2c7087bdfb65783ab581d91d3457b0fd8a3086e974068e7c833ea276
+size 3957
diff --git a/pineapple_out_of_drawer/ep004/example.md b/pineapple_out_of_drawer/ep004/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..22bd4cde158b0f492785da4fa9cf36d8e5ca327a
--- /dev/null
+++ b/pineapple_out_of_drawer/ep004/example.md
@@ -0,0 +1,35 @@
+# t007__ep004__pineapple_out_of_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the pineapple out of the drawer
+- **success:** False
+- **control steps:** 744  (14.8s at 50 Hz)
+- **policy inferences:** 128  (replan_steps=6)
+- **video:** `pineapple_out_of_drawer/ep004/rollout.mp4`
+- **features:** `128` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | approaching pineapple in drawer |
+| 3.0–6.0 | 151–301 | 25–50 | attempting grasp on pineapple |
+| 6.0–9.0 | 302–452 | 50–75 | lifting empty gripper above drawer |
+| 9.0–12.0 | 453–603 | 75–100 | hovering over pineapple |
+| 12.0–14.8 | 604–743 | 100–123 | remaining stationary above drawer |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 4.285714285714286 s (00:04.3)
+- **frame / control_step:** 216 (policy_step=36, chunk_index=0)
+- **coarse onset:** 5.0 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the pineapple inside the open drawer, but the gripper closed without securing a firm hold, causing the end-effector to lift up empty.
+  Summary: failed grasp on pineapple; lifting empty gripper
+- **recovery:** Re-align the end-effector directly over the pineapple inside the drawer, descend further to enclose the object, close the gripper firmly, and lift it out.
+  Summary: realign gripper; grasp pineapple firmly; lift out
+- **prevention:** Position the gripper lower and more centered over the pineapple's main body before closing the fingers to ensure a reliable grasp.
+  Summary: position gripper lower; secure grip on pineapple
+- **justification:** At timestamp 00:05, the robot closes its gripper on the pineapple, but upon lifting at 00:06, the pineapple remains in the drawer untouched while the empty gripper retreats upwards.
diff --git a/pineapple_out_of_drawer/ep004/labels.json b/pineapple_out_of_drawer/ep004/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..3776e12e04608c34e4626d104d3d952032af31f3
--- /dev/null
+++ b/pineapple_out_of_drawer/ep004/labels.json
@@ -0,0 +1,120 @@
+{
+  "rollout_id": "t007__ep004__pineapple_out_of_drawer",
+  "scene_variant": "",
+  "task_id": 7,
+  "task_file": "",
+  "instruction": "Take the pineapple out of the drawer",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:17:01+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.996031746031746,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "approaching pineapple in drawer",
+      "description": "approaching pineapple in drawer"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.996031746031746,
+      "t_end_sec": 5.992063492063492,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "attempting grasp on pineapple",
+      "description": "attempting grasp on pineapple"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.992063492063492,
+      "t_end_sec": 8.988095238095239,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "lifting empty gripper above drawer",
+      "description": "lifting empty gripper above drawer"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.988095238095239,
+      "t_end_sec": 11.984126984126984,
+      "control_step_start": 453,
+      "control_step_end": 603,
+      "policy_step_start": 75,
+      "policy_step_end": 100,
+      "phrase": "hovering over pineapple",
+      "description": "hovering over pineapple"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 11.984126984126984,
+      "t_end_sec": 14.761904761904763,
+      "control_step_start": 604,
+      "control_step_end": 743,
+      "policy_step_start": 100,
+      "policy_step_end": 123,
+      "phrase": "remaining stationary above drawer",
+      "description": "remaining stationary above drawer"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 4.285714285714286,
+    "vlm_failure_onset_timestamp": "00:04.3",
+    "vlm_failure_onset_frame": 216,
+    "vlm_failure_onset_step": 216,
+    "vlm_coarse_failure_onset_seconds": 5.0,
+    "vlm_failure_window_start_seconds": 4.5,
+    "vlm_failure_window_end_seconds": 6.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the pineapple inside the open drawer, but the gripper closed without securing a firm hold, causing the end-effector to lift up empty.",
+    "vlm_failure_reason_summary": "failed grasp on pineapple; lifting empty gripper",
+    "vlm_recovery_action": "Re-align the end-effector directly over the pineapple inside the drawer, descend further to enclose the object, close the gripper firmly, and lift it out.",
+    "vlm_recovery_action_summary": "realign gripper; grasp pineapple firmly; lift out",
+    "vlm_prevention_action": "Position the gripper lower and more centered over the pineapple's main body before closing the fingers to ensure a reliable grasp.",
+    "vlm_prevention_action_summary": "position gripper lower; secure grip on pineapple",
+    "vlm_justification": "At timestamp 00:05, the robot closes its gripper on the pineapple, but upon lifting at 00:06, the pineapple remains in the drawer untouched while the empty gripper retreats upwards.",
+    "vlm_temporal_justification": "By 01:55 in the clip, the robot gripper finishes closing its fingers completely without establishing a grip on the pineapple. The object remains sitting undisturbed in the drawer as the gripper closes on empty space.",
+    "vlm_video_fps": 50.4,
+    "vlm_video_num_frames": 744,
+    "vlm_video_duration_seconds": 14.742063492063492,
+    "vlm_usage": {
+      "input_tokens": 21900,
+      "output_tokens": 375,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 5.0,\n \"failure_window_start_seconds\": 4.5,\n \"failure_window_end_seconds\": 6.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the pineapple inside the open drawer, but the gripper closed without securing a firm hold, causing the end-effector to lift up empty.\",\n \"failure_reason_summary\": \"failed grasp on pineapple; lifting empty gripper\",\n \"recovery_action\": \"Re-align the end-effector directly over the pineapple inside the drawer, descend further to enclose the object, close the gripper firmly, and lift it out.\",\n \"recovery_action_summary\": \"realign gripper; grasp pineapple firmly; lift out\",\n \"prevention_action\": \"Position the gripper lower and more centered over the pineapple's main body before closing the fingers to ensure a reliable grasp.\",\n \"prevention_action_summary\": \"position gripper lower; secure grip on pineapple\",\n \"justification\": \"At timestamp 00:05, the robot closes its gripper on the pineapple, but upon lifting at 00:06, the pineapple remains in the drawer untouched while the empty gripper retreats upwards.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 216,
+    "failure_sim_step": 216,
+    "failure_policy_step": 36,
+    "failure_chunk_index": 0,
+    "first_post_failure_policy_step": 37,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-align the end-effector directly over the pineapple inside the drawer, descend further to enclose the object, close the gripper firmly, and lift it out."
+  }
+}
\ No newline at end of file
diff --git a/pineapple_out_of_drawer/ep004/labels.npz b/pineapple_out_of_drawer/ep004/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..9e3633e5201faf89e33e6e0983bd284546f8ae2e
--- /dev/null
+++ b/pineapple_out_of_drawer/ep004/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:39349476bf257491fd5b992e66bb379d8a65175dce3f9a8756b34ebcf4d47531
+size 3931
diff --git a/pineapple_out_of_drawer/ep005/example.md b/pineapple_out_of_drawer/ep005/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..ac7d804286d5486372edbc6b4b32811941f00797
--- /dev/null
+++ b/pineapple_out_of_drawer/ep005/example.md
@@ -0,0 +1,33 @@
+# t007__ep005__pineapple_out_of_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the pineapple out of the drawer
+- **success:** False
+- **control steps:** 342  (6.8s at 50 Hz)
+- **policy inferences:** 59  (replan_steps=6)
+- **video:** `pineapple_out_of_drawer/ep005/rollout.mp4`
+- **features:** `59` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–149 | 0–24 | reaching toward pineapple in drawer |
+| 3.0–6.0 | 150–299 | 25–49 | closing gripper and lifting empty |
+| 6.0–6.8 | 300–341 | 50–56 | hovering empty gripper above drawer |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 5.34 s (00:05.3)
+- **frame / control_step:** 267 (policy_step=44, chunk_index=3)
+- **coarse onset:** 4.0 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot lowered its gripper into the open drawer to grasp the pineapple, but closed its fingers without establishing a firm hold and lifted away empty.
+  Summary: failed grasp attempt; pulling away empty
+- **recovery:** Lower the end effector back into the open drawer, align the gripper fingers firmly around the pineapple, close the grasp securely, and remove it from the drawer.
+  Summary: re-align gripper; grasp pineapple securely; remove from drawer
+- **prevention:** Ensure precise positioning and depth alignment around the pineapple before closing the gripper fingers so the object is firmly secured before lifting.
+  Summary: align gripper around pineapple; close tightly before lifting
+- **justification:** At approximately 4.0 seconds, the robot closes its gripper over the pineapple inside the drawer. It fails to secure a grip and lifts its arm up empty by 5.0 seconds, leaving the pineapple inside and failing to complete the task.
diff --git a/pineapple_out_of_drawer/ep005/labels.json b/pineapple_out_of_drawer/ep005/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..6bcb4626c2be3954ab6d778305bebaf37c176b91
--- /dev/null
+++ b/pineapple_out_of_drawer/ep005/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t007__ep005__pineapple_out_of_drawer",
+  "scene_variant": "",
+  "task_id": 7,
+  "task_file": "",
+  "instruction": "Take the pineapple out of the drawer",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:17:34+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0,
+      "control_step_start": 0,
+      "control_step_end": 149,
+      "policy_step_start": 0,
+      "policy_step_end": 24,
+      "phrase": "reaching toward pineapple in drawer",
+      "description": "reaching toward pineapple in drawer"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0,
+      "t_end_sec": 6.0,
+      "control_step_start": 150,
+      "control_step_end": 299,
+      "policy_step_start": 25,
+      "policy_step_end": 49,
+      "phrase": "closing gripper and lifting empty",
+      "description": "closing gripper and lifting empty"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.0,
+      "t_end_sec": 6.84,
+      "control_step_start": 300,
+      "control_step_end": 341,
+      "policy_step_start": 50,
+      "policy_step_end": 56,
+      "phrase": "hovering empty gripper above drawer",
+      "description": "hovering empty gripper above drawer"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 5.34,
+    "vlm_failure_onset_timestamp": "00:05.3",
+    "vlm_failure_onset_frame": 267,
+    "vlm_failure_onset_step": 267,
+    "vlm_coarse_failure_onset_seconds": 4.0,
+    "vlm_failure_window_start_seconds": 3.5,
+    "vlm_failure_window_end_seconds": 4.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot lowered its gripper into the open drawer to grasp the pineapple, but closed its fingers without establishing a firm hold and lifted away empty.",
+    "vlm_failure_reason_summary": "failed grasp attempt; pulling away empty",
+    "vlm_recovery_action": "Lower the end effector back into the open drawer, align the gripper fingers firmly around the pineapple, close the grasp securely, and remove it from the drawer.",
+    "vlm_recovery_action_summary": "re-align gripper; grasp pineapple securely; remove from drawer",
+    "vlm_prevention_action": "Ensure precise positioning and depth alignment around the pineapple before closing the gripper fingers so the object is firmly secured before lifting.",
+    "vlm_prevention_action_summary": "align gripper around pineapple; close tightly before lifting",
+    "vlm_justification": "At approximately 4.0 seconds, the robot closes its gripper over the pineapple inside the drawer. It fails to secure a grip and lifts its arm up empty by 5.0 seconds, leaving the pineapple inside and failing to complete the task.",
+    "vlm_temporal_justification": "The robot lowers its gripper toward the open drawer to fetch the pineapple but aligns incorrectly. At clip timestamp 03:37 (second 217), the gripper fingers begin closing without contacting or securing the pineapple, resulting in the robot lifting away empty.",
+    "vlm_video_fps": 50.0,
+    "vlm_video_num_frames": 342,
+    "vlm_video_duration_seconds": 6.82,
+    "vlm_usage": {
+      "input_tokens": 20694,
+      "output_tokens": 386,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 4.0,\n \"failure_window_start_seconds\": 3.5,\n \"failure_window_end_seconds\": 4.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot lowered its gripper into the open drawer to grasp the pineapple, but closed its fingers without establishing a firm hold and lifted away empty.\",\n \"failure_reason_summary\": \"failed grasp attempt; pulling away empty\",\n \"recovery_action\": \"Lower the end effector back into the open drawer, align the gripper fingers firmly around the pineapple, close the grasp securely, and remove it from the drawer.\",\n \"recovery_action_summary\": \"re-align gripper; grasp pineapple securely; remove from drawer\",\n \"prevention_action\": \"Ensure precise positioning and depth alignment around the pineapple before closing the gripper fingers so the object is firmly secured before lifting.\",\n \"prevention_action_summary\": \"align gripper around pineapple; close tightly before lifting\",\n \"justification\": \"At approximately 4.0 seconds, the robot closes its gripper over the pineapple inside the drawer. It fails to secure a grip and lifts its arm up empty by 5.0 seconds, leaving the pineapple inside and failing to complete the task.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 267,
+    "failure_sim_step": 267,
+    "failure_policy_step": 44,
+    "failure_chunk_index": 3,
+    "first_post_failure_policy_step": 45,
+    "failure_type": "grasp_failure",
+    "correction_action": "Lower the end effector back into the open drawer, align the gripper fingers firmly around the pineapple, close the grasp securely, and remove it from the drawer."
+  }
+}
\ No newline at end of file
diff --git a/pineapple_out_of_drawer/ep005/labels.npz b/pineapple_out_of_drawer/ep005/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..96411aba13d767dff2babe0029f0b6a5bfdc3255
--- /dev/null
+++ b/pineapple_out_of_drawer/ep005/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0bac6485787b4640b9a439c73020cc9261b1f103b7c2c68ee1f7c35ef46ce167
+size 3886
diff --git a/pineapple_out_of_drawer/ep006/example.md b/pineapple_out_of_drawer/ep006/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..609b8276dc427ff58625e6de2a4a918ecd0efd17
--- /dev/null
+++ b/pineapple_out_of_drawer/ep006/example.md
@@ -0,0 +1,23 @@
+# t007__ep006__pineapple_out_of_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the pineapple out of the drawer
+- **success:** True
+- **control steps:** 699  (13.8s at 50 Hz)
+- **policy inferences:** 121  (replan_steps=6)
+- **video:** `pineapple_out_of_drawer/ep006/rollout.mp4`
+- **features:** `121` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | approaching pineapple in drawer |
+| 3.0–6.0 | 152–303 | 25–50 | grasping and lifting pineapple |
+| 6.0–9.0 | 304–455 | 50–75 | moving pineapple over table |
+| 9.0–12.0 | 456–607 | 76–101 | releasing pineapple and retracting |
+| 12.0–13.8 | 608–698 | 101–116 | returning to home position |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/pineapple_out_of_drawer/ep006/labels.json b/pineapple_out_of_drawer/ep006/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..9e0cca753088609dab00fb9f67cd7c79c4889db7
--- /dev/null
+++ b/pineapple_out_of_drawer/ep006/labels.json
@@ -0,0 +1,90 @@
+{
+  "rollout_id": "t007__ep006__pineapple_out_of_drawer",
+  "scene_variant": "",
+  "task_id": 7,
+  "task_file": "",
+  "instruction": "Take the pineapple out of the drawer",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:17:43+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00990099009901,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "approaching pineapple in drawer",
+      "description": "approaching pineapple in drawer"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00990099009901,
+      "t_end_sec": 6.01980198019802,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "grasping and lifting pineapple",
+      "description": "grasping and lifting pineapple"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01980198019802,
+      "t_end_sec": 9.029702970297029,
+      "control_step_start": 304,
+      "control_step_end": 455,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "moving pineapple over table",
+      "description": "moving pineapple over table"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.029702970297029,
+      "t_end_sec": 12.03960396039604,
+      "control_step_start": 456,
+      "control_step_end": 607,
+      "policy_step_start": 76,
+      "policy_step_end": 101,
+      "phrase": "releasing pineapple and retracting",
+      "description": "releasing pineapple and retracting"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 12.03960396039604,
+      "t_end_sec": 13.841584158415841,
+      "control_step_start": 608,
+      "control_step_end": 698,
+      "policy_step_start": 101,
+      "policy_step_end": 116,
+      "phrase": "returning to home position",
+      "description": "returning to home position"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/pineapple_out_of_drawer/ep006/labels.npz b/pineapple_out_of_drawer/ep006/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..5ca32e38589e3e21ba19d99f7183c6b590ecf957
--- /dev/null
+++ b/pineapple_out_of_drawer/ep006/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cda9ab84e1abeb928136036f2fb131b9b5ba4da491c4c05c0a318f06b63edb93
+size 3170
diff --git a/pineapple_out_of_drawer/ep007/example.md b/pineapple_out_of_drawer/ep007/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..55c6d3a6227685f7ede38bd32ef5267d2330634c
--- /dev/null
+++ b/pineapple_out_of_drawer/ep007/example.md
@@ -0,0 +1,22 @@
+# t007__ep007__pineapple_out_of_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the pineapple out of the drawer
+- **success:** True
+- **control steps:** 537  (10.6s at 51 Hz)
+- **policy inferences:** 93  (replan_steps=6)
+- **video:** `pineapple_out_of_drawer/ep007/rollout.mp4`
+- **features:** `93` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching for pineapple |
+| 3.0–6.0 | 152–303 | 25–50 | grasping and lifting pineapple |
+| 6.0–9.0 | 304–455 | 50–75 | placing pineapple on table |
+| 9.0–10.6 | 456–536 | 76–89 | releasing pineapple and retracting |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/pineapple_out_of_drawer/ep007/labels.json b/pineapple_out_of_drawer/ep007/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..97fffa81bcc77dcc373c62f4ee2565ab7bb57fad
--- /dev/null
+++ b/pineapple_out_of_drawer/ep007/labels.json
@@ -0,0 +1,79 @@
+{
+  "rollout_id": "t007__ep007__pineapple_out_of_drawer",
+  "scene_variant": "",
+  "task_id": 7,
+  "task_file": "",
+  "instruction": "Take the pineapple out of the drawer",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:17:51+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0039525691699605,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for pineapple",
+      "description": "reaching for pineapple"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0039525691699605,
+      "t_end_sec": 6.007905138339921,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "grasping and lifting pineapple",
+      "description": "grasping and lifting pineapple"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.007905138339921,
+      "t_end_sec": 9.011857707509881,
+      "control_step_start": 304,
+      "control_step_end": 455,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "placing pineapple on table",
+      "description": "placing pineapple on table"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.011857707509881,
+      "t_end_sec": 10.612648221343873,
+      "control_step_start": 456,
+      "control_step_end": 536,
+      "policy_step_start": 76,
+      "policy_step_end": 89,
+      "phrase": "releasing pineapple and retracting",
+      "description": "releasing pineapple and retracting"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/pineapple_out_of_drawer/ep007/labels.npz b/pineapple_out_of_drawer/ep007/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..6b75afebcb1400640e7a55cc6a56fb9e962b4535
--- /dev/null
+++ b/pineapple_out_of_drawer/ep007/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:af5110f8fd199474443b258b2fbf54293f1327e15793be4c30b9b43eeeccbce9
+size 3113
diff --git a/pineapple_out_of_drawer/ep008/example.md b/pineapple_out_of_drawer/ep008/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..b656f89b06e85a656ffe2f9e19f9c606acabdc7a
--- /dev/null
+++ b/pineapple_out_of_drawer/ep008/example.md
@@ -0,0 +1,22 @@
+# t007__ep008__pineapple_out_of_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the pineapple out of the drawer
+- **success:** True
+- **control steps:** 594  (11.8s at 50 Hz)
+- **policy inferences:** 103  (replan_steps=6)
+- **video:** `pineapple_out_of_drawer/ep008/rollout.mp4`
+- **features:** `103` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | reaching for the pineapple |
+| 3.0–6.0 | 151–301 | 25–50 | grasping the pineapple |
+| 6.0–9.0 | 302–452 | 50–75 | lifting pineapple from drawer |
+| 9.0–11.8 | 453–593 | 75–98 | placing pineapple on table |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/pineapple_out_of_drawer/ep008/labels.json b/pineapple_out_of_drawer/ep008/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..f34bd6a08ed1d0c8de0ea1e42d0ca27ef1daeae0
--- /dev/null
+++ b/pineapple_out_of_drawer/ep008/labels.json
@@ -0,0 +1,79 @@
+{
+  "rollout_id": "t007__ep008__pineapple_out_of_drawer",
+  "scene_variant": "",
+  "task_id": 7,
+  "task_file": "",
+  "instruction": "Take the pineapple out of the drawer",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:17:58+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.996031746031746,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for the pineapple",
+      "description": "reaching for the pineapple"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.996031746031746,
+      "t_end_sec": 5.992063492063492,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "grasping the pineapple",
+      "description": "grasping the pineapple"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.992063492063492,
+      "t_end_sec": 8.988095238095239,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "lifting pineapple from drawer",
+      "description": "lifting pineapple from drawer"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.988095238095239,
+      "t_end_sec": 11.785714285714286,
+      "control_step_start": 453,
+      "control_step_end": 593,
+      "policy_step_start": 75,
+      "policy_step_end": 98,
+      "phrase": "placing pineapple on table",
+      "description": "placing pineapple on table"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/pineapple_out_of_drawer/ep008/labels.npz b/pineapple_out_of_drawer/ep008/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..f32df888a2f010a95ef6512375bdcf3f7488bf6b
--- /dev/null
+++ b/pineapple_out_of_drawer/ep008/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7e1a9bcc41bbc0371a4e8433f9ffedd938b2ef70bc00e400d9563d338d800673
+size 3115
diff --git a/pineapple_out_of_drawer/ep009/example.md b/pineapple_out_of_drawer/ep009/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..452b565d7e18a3e460032d3bd201233fbf63c578
--- /dev/null
+++ b/pineapple_out_of_drawer/ep009/example.md
@@ -0,0 +1,33 @@
+# t007__ep009__pineapple_out_of_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the pineapple out of the drawer
+- **success:** False
+- **control steps:** 426  (8.4s at 51 Hz)
+- **policy inferences:** 74  (replan_steps=6)
+- **video:** `pineapple_out_of_drawer/ep009/rollout.mp4`
+- **features:** `74` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | approaching pineapple in drawer |
+| 3.0–6.0 | 152–303 | 25–50 | attempting grasp and lifting empty |
+| 6.0–8.4 | 304–425 | 50–70 | hovering above open drawer |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 4.901185770750988 s (00:04.9)
+- **frame / control_step:** 248 (policy_step=41, chunk_index=2)
+- **coarse onset:** 4.2 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot lowered its end-effector over the pineapple in the drawer but closed its gripper too high and loosely, failing to secure the pineapple before lifting away empty.
+  Summary: improper gripper positioning; missed grasp on pineapple
+- **recovery:** Lower the end-effector back into the drawer, align the gripper fingers firmly around the pineapple, close to secure it, and lift it out of the drawer.
+  Summary: re-align gripper; grasp pineapple securely; lift out
+- **prevention:** Descend deeper into the drawer and center the gripper fingers around the pineapple before closing them to ensure a firm grasp.
+  Summary: descend deeper into drawer; grasp center of pineapple
+- **justification:** At around 4.2 seconds, the robot attempts to grasp the pineapple but fails to secure it between the gripper fingers. It lifts up empty-handed by 5.0 seconds, leaving the pineapple in the drawer for the remainder of the episode.
diff --git a/pineapple_out_of_drawer/ep009/labels.json b/pineapple_out_of_drawer/ep009/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..503ed4cc6f11e0cb5e4c7c44457a348ea95109cf
--- /dev/null
+++ b/pineapple_out_of_drawer/ep009/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t007__ep009__pineapple_out_of_drawer",
+  "scene_variant": "",
+  "task_id": 7,
+  "task_file": "",
+  "instruction": "Take the pineapple out of the drawer",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:18:30+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0039525691699605,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "approaching pineapple in drawer",
+      "description": "approaching pineapple in drawer"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0039525691699605,
+      "t_end_sec": 6.007905138339921,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "attempting grasp and lifting empty",
+      "description": "attempting grasp and lifting empty"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.007905138339921,
+      "t_end_sec": 8.41897233201581,
+      "control_step_start": 304,
+      "control_step_end": 425,
+      "policy_step_start": 50,
+      "policy_step_end": 70,
+      "phrase": "hovering above open drawer",
+      "description": "hovering above open drawer"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 4.901185770750988,
+    "vlm_failure_onset_timestamp": "00:04.9",
+    "vlm_failure_onset_frame": 248,
+    "vlm_failure_onset_step": 248,
+    "vlm_coarse_failure_onset_seconds": 4.2,
+    "vlm_failure_window_start_seconds": 3.5,
+    "vlm_failure_window_end_seconds": 5.0,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot lowered its end-effector over the pineapple in the drawer but closed its gripper too high and loosely, failing to secure the pineapple before lifting away empty.",
+    "vlm_failure_reason_summary": "improper gripper positioning; missed grasp on pineapple",
+    "vlm_recovery_action": "Lower the end-effector back into the drawer, align the gripper fingers firmly around the pineapple, close to secure it, and lift it out of the drawer.",
+    "vlm_recovery_action_summary": "re-align gripper; grasp pineapple securely; lift out",
+    "vlm_prevention_action": "Descend deeper into the drawer and center the gripper fingers around the pineapple before closing them to ensure a firm grasp.",
+    "vlm_prevention_action_summary": "descend deeper into drawer; grasp center of pineapple",
+    "vlm_justification": "At around 4.2 seconds, the robot attempts to grasp the pineapple but fails to secure it between the gripper fingers. It lifts up empty-handed by 5.0 seconds, leaving the pineapple in the drawer for the remainder of the episode.",
+    "vlm_temporal_justification": "At clip second 187 (03:07), the robot begins closing its gripper fingers while positioned too high above the pineapple, leading directly to a failed grasp as the fingers close without securing the object.",
+    "vlm_video_fps": 50.6,
+    "vlm_video_num_frames": 426,
+    "vlm_video_duration_seconds": 8.399209486166008,
+    "vlm_usage": {
+      "input_tokens": 21580,
+      "output_tokens": 385,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 4.2,\n \"failure_window_start_seconds\": 3.5,\n \"failure_window_end_seconds\": 5.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot lowered its end-effector over the pineapple in the drawer but closed its gripper too high and loosely, failing to secure the pineapple before lifting away empty.\",\n \"failure_reason_summary\": \"improper gripper positioning; missed grasp on pineapple\",\n \"recovery_action\": \"Lower the end-effector back into the drawer, align the gripper fingers firmly around the pineapple, close to secure it, and lift it out of the drawer.\",\n \"recovery_action_summary\": \"re-align gripper; grasp pineapple securely; lift out\",\n \"prevention_action\": \"Descend deeper into the drawer and center the gripper fingers around the pineapple before closing them to ensure a firm grasp.\",\n \"prevention_action_summary\": \"descend deeper into drawer; grasp center of pineapple\",\n \"justification\": \"At around 4.2 seconds, the robot attempts to grasp the pineapple but fails to secure it between the gripper fingers. It lifts up empty-handed by 5.0 seconds, leaving the pineapple in the drawer for the remainder of the episode.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 248,
+    "failure_sim_step": 248,
+    "failure_policy_step": 41,
+    "failure_chunk_index": 2,
+    "first_post_failure_policy_step": 42,
+    "failure_type": "grasp_failure",
+    "correction_action": "Lower the end-effector back into the drawer, align the gripper fingers firmly around the pineapple, close to secure it, and lift it out of the drawer."
+  }
+}
\ No newline at end of file
diff --git a/pineapple_out_of_drawer/ep009/labels.npz b/pineapple_out_of_drawer/ep009/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..52e8b08d6328d89a718404633b7d8405eb674816
--- /dev/null
+++ b/pineapple_out_of_drawer/ep009/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:25dea0bf99cdff8c9247074cc5a7ecee3c0d4c62aaf00b2b59590b3e79d075a7
+size 3827
diff --git a/pineapple_out_of_drawer/ep010/example.md b/pineapple_out_of_drawer/ep010/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..3fa70140618bcb0c68b309559f2f6f8622871f99
--- /dev/null
+++ b/pineapple_out_of_drawer/ep010/example.md
@@ -0,0 +1,35 @@
+# t007__ep010__pineapple_out_of_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the pineapple out of the drawer
+- **success:** False
+- **control steps:** 648  (12.7s at 51 Hz)
+- **policy inferences:** 112  (replan_steps=6)
+- **video:** `pineapple_out_of_drawer/ep010/rollout.mp4`
+- **features:** `112` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–153 | 0–25 | approaching pineapple in drawer |
+| 3.0–6.0 | 154–307 | 25–51 | attempting grasp on pineapple |
+| 6.0–9.0 | 308–461 | 51–76 | re-opening and re-attempting grasp |
+| 9.0–12.0 | 462–615 | 77–102 | attempting grasp again |
+| 12.0–12.7 | 616–647 | 102–107 | hovering above pineapple |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 5.17578125 s (00:05.2)
+- **frame / control_step:** 265 (policy_step=44, chunk_index=1)
+- **coarse onset:** 5.0 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempts to grasp the pineapple inside the open drawer, but the gripper fails to establish a secure hold and slips off the object.
+  Summary: failed grasp on pineapple; gripper slipping
+- **recovery:** Re-align the gripper directly over the center of the pineapple, open the fingers, and close them tightly around the object before lifting.
+  Summary: re-align gripper; grasp pineapple firmly
+- **prevention:** Lower and center the end effector properly over the pineapple before closing the gripper jaws to ensure a solid grasp.
+  Summary: center gripper over pineapple; secure firm grip
+- **justification:** At around 5.0 seconds, the robot closes its gripper on the pineapple inside the drawer, but fails to achieve a stable grip, leading to repeated failed grasp attempts.
diff --git a/pineapple_out_of_drawer/ep010/labels.json b/pineapple_out_of_drawer/ep010/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..575a9f56dcb0881562fa33c2230a38f63041e5af
--- /dev/null
+++ b/pineapple_out_of_drawer/ep010/labels.json
@@ -0,0 +1,120 @@
+{
+  "rollout_id": "t007__ep010__pineapple_out_of_drawer",
+  "scene_variant": "",
+  "task_id": 7,
+  "task_file": "",
+  "instruction": "Take the pineapple out of the drawer",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:19:01+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0078125,
+      "control_step_start": 0,
+      "control_step_end": 153,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "approaching pineapple in drawer",
+      "description": "approaching pineapple in drawer"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0078125,
+      "t_end_sec": 6.015625,
+      "control_step_start": 154,
+      "control_step_end": 307,
+      "policy_step_start": 25,
+      "policy_step_end": 51,
+      "phrase": "attempting grasp on pineapple",
+      "description": "attempting grasp on pineapple"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.015625,
+      "t_end_sec": 9.0234375,
+      "control_step_start": 308,
+      "control_step_end": 461,
+      "policy_step_start": 51,
+      "policy_step_end": 76,
+      "phrase": "re-opening and re-attempting grasp",
+      "description": "re-opening and re-attempting grasp"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.0234375,
+      "t_end_sec": 12.03125,
+      "control_step_start": 462,
+      "control_step_end": 615,
+      "policy_step_start": 77,
+      "policy_step_end": 102,
+      "phrase": "attempting grasp again",
+      "description": "attempting grasp again"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 12.03125,
+      "t_end_sec": 12.65625,
+      "control_step_start": 616,
+      "control_step_end": 647,
+      "policy_step_start": 102,
+      "policy_step_end": 107,
+      "phrase": "hovering above pineapple",
+      "description": "hovering above pineapple"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 5.17578125,
+    "vlm_failure_onset_timestamp": "00:05.2",
+    "vlm_failure_onset_frame": 265,
+    "vlm_failure_onset_step": 265,
+    "vlm_coarse_failure_onset_seconds": 5.0,
+    "vlm_failure_window_start_seconds": 4.5,
+    "vlm_failure_window_end_seconds": 5.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempts to grasp the pineapple inside the open drawer, but the gripper fails to establish a secure hold and slips off the object.",
+    "vlm_failure_reason_summary": "failed grasp on pineapple; gripper slipping",
+    "vlm_recovery_action": "Re-align the gripper directly over the center of the pineapple, open the fingers, and close them tightly around the object before lifting.",
+    "vlm_recovery_action_summary": "re-align gripper; grasp pineapple firmly",
+    "vlm_prevention_action": "Lower and center the end effector properly over the pineapple before closing the gripper jaws to ensure a solid grasp.",
+    "vlm_prevention_action_summary": "center gripper over pineapple; secure firm grip",
+    "vlm_justification": "At around 5.0 seconds, the robot closes its gripper on the pineapple inside the drawer, but fails to achieve a stable grip, leading to repeated failed grasp attempts.",
+    "vlm_temporal_justification": "At 02:43 (163 seconds into the clip), the robot gripper finishes closing its fingers completely without securing the pineapple inside the drawer, resulting in a grasp failure as it begins to lift empty.",
+    "vlm_video_fps": 51.2,
+    "vlm_video_num_frames": 648,
+    "vlm_video_duration_seconds": 12.63671875,
+    "vlm_usage": {
+      "input_tokens": 22147,
+      "output_tokens": 350,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 5.0,\n \"failure_window_start_seconds\": 4.5,\n \"failure_window_end_seconds\": 5.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempts to grasp the pineapple inside the open drawer, but the gripper fails to establish a secure hold and slips off the object.\",\n \"failure_reason_summary\": \"failed grasp on pineapple; gripper slipping\",\n \"recovery_action\": \"Re-align the gripper directly over the center of the pineapple, open the fingers, and close them tightly around the object before lifting.\",\n \"recovery_action_summary\": \"re-align gripper; grasp pineapple firmly\",\n \"prevention_action\": \"Lower and center the end effector properly over the pineapple before closing the gripper jaws to ensure a solid grasp.\",\n \"prevention_action_summary\": \"center gripper over pineapple; secure firm grip\",\n \"justification\": \"At around 5.0 seconds, the robot closes its gripper on the pineapple inside the drawer, but fails to achieve a stable grip, leading to repeated failed grasp attempts.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 265,
+    "failure_sim_step": 265,
+    "failure_policy_step": 44,
+    "failure_chunk_index": 1,
+    "first_post_failure_policy_step": 45,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-align the gripper directly over the center of the pineapple, open the fingers, and close them tightly around the object before lifting."
+  }
+}
\ No newline at end of file
diff --git a/pineapple_out_of_drawer/ep010/labels.npz b/pineapple_out_of_drawer/ep010/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..5b623b1d45b1083116858b91739852cd13979111
--- /dev/null
+++ b/pineapple_out_of_drawer/ep010/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:109960523ac4ff91b7f9d237276f8072414a9d0080931cc5520c183b235eb054
+size 3826
diff --git a/pineapple_out_of_drawer/ep011/example.md b/pineapple_out_of_drawer/ep011/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..49c69053fc450f5bccabbe564646c54d7fe9cb91
--- /dev/null
+++ b/pineapple_out_of_drawer/ep011/example.md
@@ -0,0 +1,34 @@
+# t007__ep011__pineapple_out_of_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the pineapple out of the drawer
+- **success:** False
+- **control steps:** 564  (11.2s at 50 Hz)
+- **policy inferences:** 97  (replan_steps=6)
+- **video:** `pineapple_out_of_drawer/ep011/rollout.mp4`
+- **features:** `97` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | positioning end-effector over open drawer |
+| 3.0–6.0 | 151–301 | 25–50 | lowering gripper and missing pineapple grasp |
+| 6.0–9.0 | 302–452 | 50–75 | raising empty gripper and descending again |
+| 9.0–11.2 | 453–563 | 75–93 | hovering empty gripper above open drawer |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 4.940476190476191 s (00:04.9)
+- **frame / control_step:** 249 (policy_step=41, chunk_index=3)
+- **coarse onset:** 5.5 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot lowered its end-effector into the open drawer and closed its gripper, but missed the pineapple and retracted empty.
+  Summary: failed grasp on pineapple; lifting empty gripper
+- **recovery:** Re-align the gripper over the pineapple inside the open drawer, lower far enough to encapsulate the object, close the gripper firmly, and lift the pineapple out.
+  Summary: re-align gripper; grasp pineapple; lift out of drawer
+- **prevention:** Descend deeper and position the gripper jaws precisely around the middle of the pineapple before closing them to guarantee a firm hold.
+  Summary: position gripper lower; grasp pineapple firmly
+- **justification:** Around 5.5 seconds into the rollout, the robot attempts to grasp the pineapple in the open drawer, but closes its gripper prematurely without securing the object and pulls away empty.
diff --git a/pineapple_out_of_drawer/ep011/labels.json b/pineapple_out_of_drawer/ep011/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..9bc17b9908091ff2e1267d9b91f663f2c8c6804e
--- /dev/null
+++ b/pineapple_out_of_drawer/ep011/labels.json
@@ -0,0 +1,109 @@
+{
+  "rollout_id": "t007__ep011__pineapple_out_of_drawer",
+  "scene_variant": "",
+  "task_id": 7,
+  "task_file": "",
+  "instruction": "Take the pineapple out of the drawer",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:19:33+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.996031746031746,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "positioning end-effector over open drawer",
+      "description": "positioning end-effector over open drawer"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.996031746031746,
+      "t_end_sec": 5.992063492063492,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "lowering gripper and missing pineapple grasp",
+      "description": "lowering gripper and missing pineapple grasp"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.992063492063492,
+      "t_end_sec": 8.988095238095239,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "raising empty gripper and descending again",
+      "description": "raising empty gripper and descending again"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.988095238095239,
+      "t_end_sec": 11.190476190476192,
+      "control_step_start": 453,
+      "control_step_end": 563,
+      "policy_step_start": 75,
+      "policy_step_end": 93,
+      "phrase": "hovering empty gripper above open drawer",
+      "description": "hovering empty gripper above open drawer"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 4.940476190476191,
+    "vlm_failure_onset_timestamp": "00:04.9",
+    "vlm_failure_onset_frame": 249,
+    "vlm_failure_onset_step": 249,
+    "vlm_coarse_failure_onset_seconds": 5.5,
+    "vlm_failure_window_start_seconds": 5.0,
+    "vlm_failure_window_end_seconds": 6.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot lowered its end-effector into the open drawer and closed its gripper, but missed the pineapple and retracted empty.",
+    "vlm_failure_reason_summary": "failed grasp on pineapple; lifting empty gripper",
+    "vlm_recovery_action": "Re-align the gripper over the pineapple inside the open drawer, lower far enough to encapsulate the object, close the gripper firmly, and lift the pineapple out.",
+    "vlm_recovery_action_summary": "re-align gripper; grasp pineapple; lift out of drawer",
+    "vlm_prevention_action": "Descend deeper and position the gripper jaws precisely around the middle of the pineapple before closing them to guarantee a firm hold.",
+    "vlm_prevention_action_summary": "position gripper lower; grasp pineapple firmly",
+    "vlm_justification": "Around 5.5 seconds into the rollout, the robot attempts to grasp the pineapple in the open drawer, but closes its gripper prematurely without securing the object and pulls away empty.",
+    "vlm_temporal_justification": "At 02:03 (second 123 of the clip), the robot closes its gripper fingers completely without securing the pineapple, clearly failing the grasp attempt.",
+    "vlm_video_fps": 50.4,
+    "vlm_video_num_frames": 564,
+    "vlm_video_duration_seconds": 11.170634920634921,
+    "vlm_usage": {
+      "input_tokens": 21640,
+      "output_tokens": 351,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 5.5,\n \"failure_window_start_seconds\": 5.0,\n \"failure_window_end_seconds\": 6.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot lowered its end-effector into the open drawer and closed its gripper, but missed the pineapple and retracted empty.\",\n \"failure_reason_summary\": \"failed grasp on pineapple; lifting empty gripper\",\n \"recovery_action\": \"Re-align the gripper over the pineapple inside the open drawer, lower far enough to encapsulate the object, close the gripper firmly, and lift the pineapple out.\",\n \"recovery_action_summary\": \"re-align gripper; grasp pineapple; lift out of drawer\",\n \"prevention_action\": \"Descend deeper and position the gripper jaws precisely around the middle of the pineapple before closing them to guarantee a firm hold.\",\n \"prevention_action_summary\": \"position gripper lower; grasp pineapple firmly\",\n \"justification\": \"Around 5.5 seconds into the rollout, the robot attempts to grasp the pineapple in the open drawer, but closes its gripper prematurely without securing the object and pulls away empty.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 249,
+    "failure_sim_step": 249,
+    "failure_policy_step": 41,
+    "failure_chunk_index": 3,
+    "first_post_failure_policy_step": 42,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-align the gripper over the pineapple inside the open drawer, lower far enough to encapsulate the object, close the gripper firmly, and lift the pineapple out."
+  }
+}
\ No newline at end of file
diff --git a/pineapple_out_of_drawer/ep011/labels.npz b/pineapple_out_of_drawer/ep011/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..726ca20d5e411f1a64f3a6a19c2535686fc0c40d
--- /dev/null
+++ b/pineapple_out_of_drawer/ep011/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f7770bd16d6f03abe229a8dfea04d3919f920c5e470668044cd3d80437f7139f
+size 3880
diff --git a/pineapple_out_of_drawer/ep012/example.md b/pineapple_out_of_drawer/ep012/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..d3af12636ea29b15b101a2f8aafc5c61d7646bf9
--- /dev/null
+++ b/pineapple_out_of_drawer/ep012/example.md
@@ -0,0 +1,35 @@
+# t007__ep012__pineapple_out_of_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the pineapple out of the drawer
+- **success:** False
+- **control steps:** 744  (14.8s at 50 Hz)
+- **policy inferences:** 128  (replan_steps=6)
+- **video:** `pineapple_out_of_drawer/ep012/rollout.mp4`
+- **features:** `128` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | approaching open drawer |
+| 3.0–6.0 | 151–301 | 25–50 | lowering gripper to grasp pineapple |
+| 6.0–9.0 | 302–452 | 50–75 | retracting empty gripper and repositioning |
+| 9.0–12.0 | 453–603 | 75–100 | attempting second grasp on pineapple |
+| 12.0–14.8 | 604–743 | 100–123 | retracting empty gripper |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 5.337301587301588 s (00:05.3)
+- **frame / control_step:** 269 (policy_step=44, chunk_index=5)
+- **coarse onset:** 5.0 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot lowered its gripper into the drawer to grasp the pineapple, but the fingers failed to secure a firm grip and closed prematurely, causing the gripper to lift up empty.
+  Summary: unsuccessful grasp of pineapple; gripper lifts empty
+- **recovery:** Re-align the gripper over the pineapple inside the drawer, lower the end-effector deeper to enclose the pineapple, close the gripper firmly, and lift it out.
+  Summary: re-align gripper; lower further into drawer; grasp pineapple securely
+- **prevention:** Center the gripper directly over the pineapple's center of mass and descend sufficiently deep into the drawer before closing the fingers.
+  Summary: center gripper over pineapple; lower deeper before closing
+- **justification:** At around 5.0 seconds, the robot's gripper attempts to grasp the pineapple in the open drawer but closes without firmly grasping the object, pulling back up empty and repeating this failed grasp cycle.
diff --git a/pineapple_out_of_drawer/ep012/labels.json b/pineapple_out_of_drawer/ep012/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..f51c112f542e5a7d862b77e24c0272c5a2122a5f
--- /dev/null
+++ b/pineapple_out_of_drawer/ep012/labels.json
@@ -0,0 +1,120 @@
+{
+  "rollout_id": "t007__ep012__pineapple_out_of_drawer",
+  "scene_variant": "",
+  "task_id": 7,
+  "task_file": "",
+  "instruction": "Take the pineapple out of the drawer",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:20:07+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.996031746031746,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "approaching open drawer",
+      "description": "approaching open drawer"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.996031746031746,
+      "t_end_sec": 5.992063492063492,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "lowering gripper to grasp pineapple",
+      "description": "lowering gripper to grasp pineapple"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.992063492063492,
+      "t_end_sec": 8.988095238095239,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "retracting empty gripper and repositioning",
+      "description": "retracting empty gripper and repositioning"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.988095238095239,
+      "t_end_sec": 11.984126984126984,
+      "control_step_start": 453,
+      "control_step_end": 603,
+      "policy_step_start": 75,
+      "policy_step_end": 100,
+      "phrase": "attempting second grasp on pineapple",
+      "description": "attempting second grasp on pineapple"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 11.984126984126984,
+      "t_end_sec": 14.761904761904763,
+      "control_step_start": 604,
+      "control_step_end": 743,
+      "policy_step_start": 100,
+      "policy_step_end": 123,
+      "phrase": "retracting empty gripper",
+      "description": "retracting empty gripper"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 5.337301587301588,
+    "vlm_failure_onset_timestamp": "00:05.3",
+    "vlm_failure_onset_frame": 269,
+    "vlm_failure_onset_step": 269,
+    "vlm_coarse_failure_onset_seconds": 5.0,
+    "vlm_failure_window_start_seconds": 4.5,
+    "vlm_failure_window_end_seconds": 6.0,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot lowered its gripper into the drawer to grasp the pineapple, but the fingers failed to secure a firm grip and closed prematurely, causing the gripper to lift up empty.",
+    "vlm_failure_reason_summary": "unsuccessful grasp of pineapple; gripper lifts empty",
+    "vlm_recovery_action": "Re-align the gripper over the pineapple inside the drawer, lower the end-effector deeper to enclose the pineapple, close the gripper firmly, and lift it out.",
+    "vlm_recovery_action_summary": "re-align gripper; lower further into drawer; grasp pineapple securely",
+    "vlm_prevention_action": "Center the gripper directly over the pineapple's center of mass and descend sufficiently deep into the drawer before closing the fingers.",
+    "vlm_prevention_action_summary": "center gripper over pineapple; lower deeper before closing",
+    "vlm_justification": "At around 5.0 seconds, the robot's gripper attempts to grasp the pineapple in the open drawer but closes without firmly grasping the object, pulling back up empty and repeating this failed grasp cycle.",
+    "vlm_temporal_justification": "At 02:48 (second 168 of the clip), the gripper fingers close completely without properly encompassing or securing the pineapple, making the grasp failure visibly obvious before the robot lifts up empty.",
+    "vlm_video_fps": 50.4,
+    "vlm_video_num_frames": 744,
+    "vlm_video_duration_seconds": 14.742063492063492,
+    "vlm_usage": {
+      "input_tokens": 21902,
+      "output_tokens": 379,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 5.0,\n \"failure_window_start_seconds\": 4.5,\n \"failure_window_end_seconds\": 6.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot lowered its gripper into the drawer to grasp the pineapple, but the fingers failed to secure a firm grip and closed prematurely, causing the gripper to lift up empty.\",\n \"failure_reason_summary\": \"unsuccessful grasp of pineapple; gripper lifts empty\",\n \"recovery_action\": \"Re-align the gripper over the pineapple inside the drawer, lower the end-effector deeper to enclose the pineapple, close the gripper firmly, and lift it out.\",\n \"recovery_action_summary\": \"re-align gripper; lower further into drawer; grasp pineapple securely\",\n \"prevention_action\": \"Center the gripper directly over the pineapple's center of mass and descend sufficiently deep into the drawer before closing the fingers.\",\n \"prevention_action_summary\": \"center gripper over pineapple; lower deeper before closing\",\n \"justification\": \"At around 5.0 seconds, the robot's gripper attempts to grasp the pineapple in the open drawer but closes without firmly grasping the object, pulling back up empty and repeating this failed grasp cycle.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 269,
+    "failure_sim_step": 269,
+    "failure_policy_step": 44,
+    "failure_chunk_index": 5,
+    "first_post_failure_policy_step": 45,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-align the gripper over the pineapple inside the drawer, lower the end-effector deeper to enclose the pineapple, close the gripper firmly, and lift it out."
+  }
+}
\ No newline at end of file
diff --git a/pineapple_out_of_drawer/ep012/labels.npz b/pineapple_out_of_drawer/ep012/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..1c283d4cc665311d02276df259f0fc4b3c412415
--- /dev/null
+++ b/pineapple_out_of_drawer/ep012/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8857f54be1802dacc43bb5e97cb37f7a2eaf7b4cab968fa4a39038cd406a31c2
+size 3951
diff --git a/pineapple_out_of_drawer/ep013/example.md b/pineapple_out_of_drawer/ep013/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..a97856878e1935b4c9b39f0a67fac57071e45628
--- /dev/null
+++ b/pineapple_out_of_drawer/ep013/example.md
@@ -0,0 +1,27 @@
+# t007__ep013__pineapple_out_of_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the pineapple out of the drawer
+- **success:** True
+- **control steps:** 1228  (24.2s at 51 Hz)
+- **policy inferences:** 209  (replan_steps=6)
+- **video:** `pineapple_out_of_drawer/ep013/rollout.mp4`
+- **features:** `209` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | idle above table |
+| 3.0–6.0 | 152–303 | 25–50 | reaching for pineapple |
+| 6.0–9.0 | 304–455 | 50–75 | grasping pineapple |
+| 9.0–12.0 | 456–607 | 76–101 | lifting pineapple out |
+| 12.0–15.0 | 608–759 | 101–126 | holding pineapple near drawer |
+| 15.0–18.0 | 760–911 | 126–151 | moving pineapple to side |
+| 18.0–21.0 | 912–1063 | 152–177 | lowering pineapple to table |
+| 21.0–24.0 | 1064–1215 | 177–202 | releasing pineapple on table |
+| 24.0–24.2 | 1216–1227 | 202–204 | idle near table |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/pineapple_out_of_drawer/ep013/labels.json b/pineapple_out_of_drawer/ep013/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..30be98ac7a25100a1d70bc0c69746865b435b006
--- /dev/null
+++ b/pineapple_out_of_drawer/ep013/labels.json
@@ -0,0 +1,134 @@
+{
+  "rollout_id": "t007__ep013__pineapple_out_of_drawer",
+  "scene_variant": "",
+  "task_id": 7,
+  "task_file": "",
+  "instruction": "Take the pineapple out of the drawer",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:20:18+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9980276134122286,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "idle above table",
+      "description": "idle above table"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9980276134122286,
+      "t_end_sec": 5.996055226824457,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "reaching for pineapple",
+      "description": "reaching for pineapple"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.996055226824457,
+      "t_end_sec": 8.994082840236686,
+      "control_step_start": 304,
+      "control_step_end": 455,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "grasping pineapple",
+      "description": "grasping pineapple"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.994082840236686,
+      "t_end_sec": 11.992110453648914,
+      "control_step_start": 456,
+      "control_step_end": 607,
+      "policy_step_start": 76,
+      "policy_step_end": 101,
+      "phrase": "lifting pineapple out",
+      "description": "lifting pineapple out"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 11.992110453648914,
+      "t_end_sec": 14.990138067061142,
+      "control_step_start": 608,
+      "control_step_end": 759,
+      "policy_step_start": 101,
+      "policy_step_end": 126,
+      "phrase": "holding pineapple near drawer",
+      "description": "holding pineapple near drawer"
+    },
+    {
+      "segment_index": 5,
+      "t_start_sec": 14.990138067061142,
+      "t_end_sec": 17.988165680473372,
+      "control_step_start": 760,
+      "control_step_end": 911,
+      "policy_step_start": 126,
+      "policy_step_end": 151,
+      "phrase": "moving pineapple to side",
+      "description": "moving pineapple to side"
+    },
+    {
+      "segment_index": 6,
+      "t_start_sec": 17.988165680473372,
+      "t_end_sec": 20.986193293885602,
+      "control_step_start": 912,
+      "control_step_end": 1063,
+      "policy_step_start": 152,
+      "policy_step_end": 177,
+      "phrase": "lowering pineapple to table",
+      "description": "lowering pineapple to table"
+    },
+    {
+      "segment_index": 7,
+      "t_start_sec": 20.986193293885602,
+      "t_end_sec": 23.98422090729783,
+      "control_step_start": 1064,
+      "control_step_end": 1215,
+      "policy_step_start": 177,
+      "policy_step_end": 202,
+      "phrase": "releasing pineapple on table",
+      "description": "releasing pineapple on table"
+    },
+    {
+      "segment_index": 8,
+      "t_start_sec": 23.98422090729783,
+      "t_end_sec": 24.220907297830372,
+      "control_step_start": 1216,
+      "control_step_end": 1227,
+      "policy_step_start": 202,
+      "policy_step_end": 204,
+      "phrase": "idle near table",
+      "description": "idle near table"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/pineapple_out_of_drawer/ep013/labels.npz b/pineapple_out_of_drawer/ep013/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..68f45cebf085e0e50d64150ba7329c5e28c73b69
--- /dev/null
+++ b/pineapple_out_of_drawer/ep013/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bae9fecf4cf28a7b1fb55f0726d7a51f6abb243d569d6bcb29d8705da6a65bdf
+size 3269
diff --git a/pineapple_out_of_drawer/ep014/example.md b/pineapple_out_of_drawer/ep014/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..745681849418e22fe0fc29a33e33ed9480f44efd
--- /dev/null
+++ b/pineapple_out_of_drawer/ep014/example.md
@@ -0,0 +1,38 @@
+# t007__ep014__pineapple_out_of_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the pineapple out of the drawer
+- **success:** False
+- **control steps:** 1083  (21.4s at 50 Hz)
+- **policy inferences:** 187  (replan_steps=6)
+- **video:** `pineapple_out_of_drawer/ep014/rollout.mp4`
+- **features:** `187` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching toward open drawer |
+| 3.0–6.0 | 152–303 | 25–50 | lowering end-effector into drawer |
+| 6.0–9.0 | 304–455 | 50–75 | attempting to grasp pineapple |
+| 9.0–12.0 | 456–607 | 76–101 | slipping off pineapple |
+| 12.0–15.0 | 608–759 | 101–126 | opening and closing empty gripper |
+| 15.0–18.1 | 760–911 | 126–151 | hovering above pineapple |
+| 18.1–21.1 | 912–1063 | 152–177 | hovering idle above drawer |
+| 21.1–21.4 | 1064–1082 | 177–180 | remaining stationary |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 6.4950495049504955 s (00:06.5)
+- **frame / control_step:** 328 (policy_step=54, chunk_index=4)
+- **coarse onset:** 6.5 s (refined=False)
+- **mode:** grasp_failure
+- **reason:** The robot positioned its end-effector over the pineapple toy inside the open drawer but failed to establish a secure grip when closing its fingers, leading to repeated failed grasp attempts.
+  Summary: failed grasp on pineapple; missing object engagement
+- **recovery:** Re-center the gripper over the pineapple, descend deeper into the open drawer, and close the fingers tightly around the object before lifting.
+  Summary: re-align gripper; lower into drawer; grasp pineapple securely
+- **prevention:** Position the end-effector lower relative to the pineapple toy before closing the fingers to ensure a deep and stable grasp.
+  Summary: position gripper lower; grasp pineapple firmly
+- **justification:** At around 6.5 seconds, the robot's gripper attempts to grasp the pineapple in the open drawer but fails to close deeply around it, causing the fingers to slip off without grabbing the object. Subsequent attempts also fail to secure a grip.
diff --git a/pineapple_out_of_drawer/ep014/labels.json b/pineapple_out_of_drawer/ep014/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..b56c41f8182c981ee0aaef9d4bc694bf7efc367e
--- /dev/null
+++ b/pineapple_out_of_drawer/ep014/labels.json
@@ -0,0 +1,153 @@
+{
+  "rollout_id": "t007__ep014__pineapple_out_of_drawer",
+  "scene_variant": "",
+  "task_id": 7,
+  "task_file": "",
+  "instruction": "Take the pineapple out of the drawer",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:20:57+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00990099009901,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching toward open drawer",
+      "description": "reaching toward open drawer"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00990099009901,
+      "t_end_sec": 6.01980198019802,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "lowering end-effector into drawer",
+      "description": "lowering end-effector into drawer"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01980198019802,
+      "t_end_sec": 9.029702970297029,
+      "control_step_start": 304,
+      "control_step_end": 455,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "attempting to grasp pineapple",
+      "description": "attempting to grasp pineapple"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.029702970297029,
+      "t_end_sec": 12.03960396039604,
+      "control_step_start": 456,
+      "control_step_end": 607,
+      "policy_step_start": 76,
+      "policy_step_end": 101,
+      "phrase": "slipping off pineapple",
+      "description": "slipping off pineapple"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 12.03960396039604,
+      "t_end_sec": 15.049504950495049,
+      "control_step_start": 608,
+      "control_step_end": 759,
+      "policy_step_start": 101,
+      "policy_step_end": 126,
+      "phrase": "opening and closing empty gripper",
+      "description": "opening and closing empty gripper"
+    },
+    {
+      "segment_index": 5,
+      "t_start_sec": 15.049504950495049,
+      "t_end_sec": 18.059405940594058,
+      "control_step_start": 760,
+      "control_step_end": 911,
+      "policy_step_start": 126,
+      "policy_step_end": 151,
+      "phrase": "hovering above pineapple",
+      "description": "hovering above pineapple"
+    },
+    {
+      "segment_index": 6,
+      "t_start_sec": 18.059405940594058,
+      "t_end_sec": 21.06930693069307,
+      "control_step_start": 912,
+      "control_step_end": 1063,
+      "policy_step_start": 152,
+      "policy_step_end": 177,
+      "phrase": "hovering idle above drawer",
+      "description": "hovering idle above drawer"
+    },
+    {
+      "segment_index": 7,
+      "t_start_sec": 21.06930693069307,
+      "t_end_sec": 21.445544554455445,
+      "control_step_start": 1064,
+      "control_step_end": 1082,
+      "policy_step_start": 177,
+      "policy_step_end": 180,
+      "phrase": "remaining stationary",
+      "description": "remaining stationary"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 6.4950495049504955,
+    "vlm_failure_onset_timestamp": "00:06.5",
+    "vlm_failure_onset_frame": 328,
+    "vlm_failure_onset_step": 328,
+    "vlm_coarse_failure_onset_seconds": 6.5,
+    "vlm_failure_window_start_seconds": 5.5,
+    "vlm_failure_window_end_seconds": 7.5,
+    "vlm_temporal_refined": false,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot positioned its end-effector over the pineapple toy inside the open drawer but failed to establish a secure grip when closing its fingers, leading to repeated failed grasp attempts.",
+    "vlm_failure_reason_summary": "failed grasp on pineapple; missing object engagement",
+    "vlm_recovery_action": "Re-center the gripper over the pineapple, descend deeper into the open drawer, and close the fingers tightly around the object before lifting.",
+    "vlm_recovery_action_summary": "re-align gripper; lower into drawer; grasp pineapple securely",
+    "vlm_prevention_action": "Position the end-effector lower relative to the pineapple toy before closing the fingers to ensure a deep and stable grasp.",
+    "vlm_prevention_action_summary": "position gripper lower; grasp pineapple firmly",
+    "vlm_justification": "At around 6.5 seconds, the robot's gripper attempts to grasp the pineapple in the open drawer but fails to close deeply around it, causing the fingers to slip off without grabbing the object. Subsequent attempts also fail to secure a grip.",
+    "vlm_temporal_justification": "At 01:40 in the clip, the robot closes its gripper around the pineapple toy, but fails to establish a secure grip, marking the exact onset of the grasp failure.",
+    "vlm_video_fps": 50.5,
+    "vlm_video_num_frames": 1083,
+    "vlm_video_duration_seconds": 21.425742574257427,
+    "vlm_usage": {
+      "input_tokens": 22407,
+      "output_tokens": 374,
+      "cached_tokens": 1429
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 6.5,\n \"failure_window_start_seconds\": 5.5,\n \"failure_window_end_seconds\": 7.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot positioned its end-effector over the pineapple toy inside the open drawer but failed to establish a secure grip when closing its fingers, leading to repeated failed grasp attempts.\",\n \"failure_reason_summary\": \"failed grasp on pineapple; missing object engagement\",\n \"recovery_action\": \"Re-center the gripper over the pineapple, descend deeper into the open drawer, and close the fingers tightly around the object before lifting.\",\n \"recovery_action_summary\": \"re-align gripper; lower into drawer; grasp pineapple securely\",\n \"prevention_action\": \"Position the end-effector lower relative to the pineapple toy before closing the fingers to ensure a deep and stable grasp.\",\n \"prevention_action_summary\": \"position gripper lower; grasp pineapple firmly\",\n \"justification\": \"At around 6.5 seconds, the robot's gripper attempts to grasp the pineapple in the open drawer but fails to close deeply around it, causing the fingers to slip off without grabbing the object. Subsequent attempts also fail to secure a grip.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 328,
+    "failure_sim_step": 328,
+    "failure_policy_step": 54,
+    "failure_chunk_index": 4,
+    "first_post_failure_policy_step": 55,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-center the gripper over the pineapple, descend deeper into the open drawer, and close the fingers tightly around the object before lifting."
+  }
+}
\ No newline at end of file
diff --git a/pineapple_out_of_drawer/ep014/labels.npz b/pineapple_out_of_drawer/ep014/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..e7f9654d099c1b880e33cd595ccd48ce42526ae6
--- /dev/null
+++ b/pineapple_out_of_drawer/ep014/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8393adb9d59c17aeb5f72ed4dbfd03da170db527f3dc8bab385f9189d9c1aa01
+size 4068
diff --git a/pineapple_out_of_drawer/ep015/example.md b/pineapple_out_of_drawer/ep015/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..297a08fe8939b99a0a1951007c1eb6b29e893df5
--- /dev/null
+++ b/pineapple_out_of_drawer/ep015/example.md
@@ -0,0 +1,35 @@
+# t007__ep015__pineapple_out_of_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the pineapple out of the drawer
+- **success:** False
+- **control steps:** 691  (13.7s at 50 Hz)
+- **policy inferences:** 121  (replan_steps=6)
+- **video:** `pineapple_out_of_drawer/ep015/rollout.mp4`
+- **features:** `121` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | approaching pineapple in open drawer |
+| 3.0–6.0 | 151–301 | 25–50 | lowering gripper onto pineapple |
+| 6.0–9.0 | 302–452 | 50–75 | lifting gripper without pineapple |
+| 9.0–12.0 | 453–603 | 75–100 | retracting empty gripper |
+| 12.0–13.7 | 604–690 | 100–115 | hovering above tabletop |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 6.003976143141154 s (00:06.0)
+- **frame / control_step:** 302 (policy_step=50, chunk_index=2)
+- **coarse onset:** 6.0 s (refined=False)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the pineapple inside the open drawer, but the gripper slipped off without securing a hold, leaving the pineapple in the drawer as the arm lifted away.
+  Summary: failed grasp on pineapple; lifting empty gripper
+- **recovery:** Lower the end-effector back into the open drawer, position the gripper fingers squarely around the body of the pineapple, close the fingers firmly, and lift it clear of the drawer.
+  Summary: re-align over pineapple; grasp firmly; lift out of drawer
+- **prevention:** Ensure the gripper descends sufficiently deep and centered over the pineapple before closing the fingers to achieve a firm grip prior to lifting.
+  Summary: center gripper on pineapple; secure grip before lifting
+- **justification:** At around 6.0 seconds, the robot closes its gripper around the pineapple in the drawer, but slips off as it lifts up at 6.7 seconds, leaving the object unmoved in the drawer.
diff --git a/pineapple_out_of_drawer/ep015/labels.json b/pineapple_out_of_drawer/ep015/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..9772f16bfbd733eac06fcaa6faf30db42b920555
--- /dev/null
+++ b/pineapple_out_of_drawer/ep015/labels.json
@@ -0,0 +1,120 @@
+{
+  "rollout_id": "t007__ep015__pineapple_out_of_drawer",
+  "scene_variant": "",
+  "task_id": 7,
+  "task_file": "",
+  "instruction": "Take the pineapple out of the drawer",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:21:30+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.001988071570577,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "approaching pineapple in open drawer",
+      "description": "approaching pineapple in open drawer"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.001988071570577,
+      "t_end_sec": 6.003976143141154,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "lowering gripper onto pineapple",
+      "description": "lowering gripper onto pineapple"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.003976143141154,
+      "t_end_sec": 9.00596421471173,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "lifting gripper without pineapple",
+      "description": "lifting gripper without pineapple"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.00596421471173,
+      "t_end_sec": 12.007952286282308,
+      "control_step_start": 453,
+      "control_step_end": 603,
+      "policy_step_start": 75,
+      "policy_step_end": 100,
+      "phrase": "retracting empty gripper",
+      "description": "retracting empty gripper"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 12.007952286282308,
+      "t_end_sec": 13.737574552683897,
+      "control_step_start": 604,
+      "control_step_end": 690,
+      "policy_step_start": 100,
+      "policy_step_end": 115,
+      "phrase": "hovering above tabletop",
+      "description": "hovering above tabletop"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 6.003976143141154,
+    "vlm_failure_onset_timestamp": "00:06.0",
+    "vlm_failure_onset_frame": 302,
+    "vlm_failure_onset_step": 302,
+    "vlm_coarse_failure_onset_seconds": 6.0,
+    "vlm_failure_window_start_seconds": 5.5,
+    "vlm_failure_window_end_seconds": 7.0,
+    "vlm_temporal_refined": false,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the pineapple inside the open drawer, but the gripper slipped off without securing a hold, leaving the pineapple in the drawer as the arm lifted away.",
+    "vlm_failure_reason_summary": "failed grasp on pineapple; lifting empty gripper",
+    "vlm_recovery_action": "Lower the end-effector back into the open drawer, position the gripper fingers squarely around the body of the pineapple, close the fingers firmly, and lift it clear of the drawer.",
+    "vlm_recovery_action_summary": "re-align over pineapple; grasp firmly; lift out of drawer",
+    "vlm_prevention_action": "Ensure the gripper descends sufficiently deep and centered over the pineapple before closing the fingers to achieve a firm grip prior to lifting.",
+    "vlm_prevention_action_summary": "center gripper on pineapple; secure grip before lifting",
+    "vlm_justification": "At around 6.0 seconds, the robot closes its gripper around the pineapple in the drawer, but slips off as it lifts up at 6.7 seconds, leaving the object unmoved in the drawer.",
+    "vlm_temporal_justification": "The robot arm begins lifting upward after attempting to grasp the pineapple, but the gripper fails to establish a hold and slides off the object, leaving the pineapple in the drawer.",
+    "vlm_video_fps": 50.3,
+    "vlm_video_num_frames": 691,
+    "vlm_video_duration_seconds": 13.717693836978132,
+    "vlm_usage": {
+      "input_tokens": 21839,
+      "output_tokens": 376,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 6.0,\n \"failure_window_start_seconds\": 5.5,\n \"failure_window_end_seconds\": 7.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the pineapple inside the open drawer, but the gripper slipped off without securing a hold, leaving the pineapple in the drawer as the arm lifted away.\",\n \"failure_reason_summary\": \"failed grasp on pineapple; lifting empty gripper\",\n \"recovery_action\": \"Lower the end-effector back into the open drawer, position the gripper fingers squarely around the body of the pineapple, close the fingers firmly, and lift it clear of the drawer.\",\n \"recovery_action_summary\": \"re-align over pineapple; grasp firmly; lift out of drawer\",\n \"prevention_action\": \"Ensure the gripper descends sufficiently deep and centered over the pineapple before closing the fingers to achieve a firm grip prior to lifting.\",\n \"prevention_action_summary\": \"center gripper on pineapple; secure grip before lifting\",\n \"justification\": \"At around 6.0 seconds, the robot closes its gripper around the pineapple in the drawer, but slips off as it lifts up at 6.7 seconds, leaving the object unmoved in the drawer.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 302,
+    "failure_sim_step": 302,
+    "failure_policy_step": 50,
+    "failure_chunk_index": 2,
+    "first_post_failure_policy_step": 51,
+    "failure_type": "grasp_failure",
+    "correction_action": "Lower the end-effector back into the open drawer, position the gripper fingers squarely around the body of the pineapple, close the fingers firmly, and lift it clear of the drawer."
+  }
+}
\ No newline at end of file
diff --git a/pineapple_out_of_drawer/ep015/labels.npz b/pineapple_out_of_drawer/ep015/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..cc6988519f00d358d63b7c0e9481e350c789c107
--- /dev/null
+++ b/pineapple_out_of_drawer/ep015/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d67f5a59cba0fe1e4fa9e8f670070b54a6d3cb04d2400018ecec009c41ae43a8
+size 3981
diff --git a/pineapple_out_of_drawer/ep016/example.md b/pineapple_out_of_drawer/ep016/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..ffcfe7331a069d3a6bc11fae27dbe32f7f48d7b6
--- /dev/null
+++ b/pineapple_out_of_drawer/ep016/example.md
@@ -0,0 +1,37 @@
+# t007__ep016__pineapple_out_of_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the pineapple out of the drawer
+- **success:** False
+- **control steps:** 1025  (20.4s at 50 Hz)
+- **policy inferences:** 177  (replan_steps=6)
+- **video:** `pineapple_out_of_drawer/ep016/rollout.mp4`
+- **features:** `177` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | hovering above tabletop |
+| 3.0–6.0 | 151–301 | 25–50 | approaching open drawer |
+| 6.0–9.0 | 302–452 | 50–75 | aligning over pineapple |
+| 9.0–12.0 | 453–603 | 75–100 | lowering into drawer |
+| 12.0–15.0 | 604–754 | 100–125 | closing gripper on object |
+| 15.0–18.0 | 755–905 | 125–150 | lifting empty gripper |
+| 18.0–20.4 | 906–1024 | 151–170 | retracting arm |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 12.107355864811133 s (00:12.1)
+- **frame / control_step:** 609 (policy_step=101, chunk_index=3)
+- **coarse onset:** 12.0 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the pineapple inside the open drawer, but the gripper failed to position itself securely around the object, causing the fingers to slip and lift up empty.
+  Summary: improper gripper alignment; failed grasp on pineapple
+- **recovery:** Lower the arm back into the open drawer, realign the gripper with the center of the pineapple, close the gripper securely, and lift the pineapple out of the drawer.
+  Summary: realign gripper; grasp pineapple; lift out of drawer
+- **prevention:** Center the end-effector over the main body of the pineapple and ensure a firm grip is established before initiating the upward lift.
+  Summary: center gripper on pineapple; secure grasp before lifting
+- **justification:** The robot lowers its end-effector into the drawer around 10-12 seconds to pick up the pineapple, but fails to securely grasp it. It lifts its empty gripper back up around 15 seconds, leaving the pineapple unmoved inside the drawer.
diff --git a/pineapple_out_of_drawer/ep016/labels.json b/pineapple_out_of_drawer/ep016/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..64614966f8106fdf68c30ab7332e6624cf2a15bc
--- /dev/null
+++ b/pineapple_out_of_drawer/ep016/labels.json
@@ -0,0 +1,142 @@
+{
+  "rollout_id": "t007__ep016__pineapple_out_of_drawer",
+  "scene_variant": "",
+  "task_id": 7,
+  "task_file": "",
+  "instruction": "Take the pineapple out of the drawer",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:22:01+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.001988071570577,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "hovering above tabletop",
+      "description": "hovering above tabletop"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.001988071570577,
+      "t_end_sec": 6.003976143141154,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "approaching open drawer",
+      "description": "approaching open drawer"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.003976143141154,
+      "t_end_sec": 9.00596421471173,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "aligning over pineapple",
+      "description": "aligning over pineapple"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.00596421471173,
+      "t_end_sec": 12.007952286282308,
+      "control_step_start": 453,
+      "control_step_end": 603,
+      "policy_step_start": 75,
+      "policy_step_end": 100,
+      "phrase": "lowering into drawer",
+      "description": "lowering into drawer"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 12.007952286282308,
+      "t_end_sec": 15.009940357852884,
+      "control_step_start": 604,
+      "control_step_end": 754,
+      "policy_step_start": 100,
+      "policy_step_end": 125,
+      "phrase": "closing gripper on object",
+      "description": "closing gripper on object"
+    },
+    {
+      "segment_index": 5,
+      "t_start_sec": 15.009940357852884,
+      "t_end_sec": 18.01192842942346,
+      "control_step_start": 755,
+      "control_step_end": 905,
+      "policy_step_start": 125,
+      "policy_step_end": 150,
+      "phrase": "lifting empty gripper",
+      "description": "lifting empty gripper"
+    },
+    {
+      "segment_index": 6,
+      "t_start_sec": 18.01192842942346,
+      "t_end_sec": 20.377733598409545,
+      "control_step_start": 906,
+      "control_step_end": 1024,
+      "policy_step_start": 151,
+      "policy_step_end": 170,
+      "phrase": "retracting arm",
+      "description": "retracting arm"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 12.107355864811133,
+    "vlm_failure_onset_timestamp": "00:12.1",
+    "vlm_failure_onset_frame": 609,
+    "vlm_failure_onset_step": 609,
+    "vlm_coarse_failure_onset_seconds": 12.0,
+    "vlm_failure_window_start_seconds": 10.5,
+    "vlm_failure_window_end_seconds": 14.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the pineapple inside the open drawer, but the gripper failed to position itself securely around the object, causing the fingers to slip and lift up empty.",
+    "vlm_failure_reason_summary": "improper gripper alignment; failed grasp on pineapple",
+    "vlm_recovery_action": "Lower the arm back into the open drawer, realign the gripper with the center of the pineapple, close the gripper securely, and lift the pineapple out of the drawer.",
+    "vlm_recovery_action_summary": "realign gripper; grasp pineapple; lift out of drawer",
+    "vlm_prevention_action": "Center the end-effector over the main body of the pineapple and ensure a firm grip is established before initiating the upward lift.",
+    "vlm_prevention_action_summary": "center gripper on pineapple; secure grasp before lifting",
+    "vlm_justification": "The robot lowers its end-effector into the drawer around 10-12 seconds to pick up the pineapple, but fails to securely grasp it. It lifts its empty gripper back up around 15 seconds, leaving the pineapple unmoved inside the drawer.",
+    "vlm_temporal_justification": "At 02:36 (156 seconds into the clip), the gripper fully closes without securing the pineapple and begins lifting upward empty, leaving the object behind in the drawer.",
+    "vlm_video_fps": 50.3,
+    "vlm_video_num_frames": 1025,
+    "vlm_video_duration_seconds": 20.357852882703778,
+    "vlm_usage": {
+      "input_tokens": 22219,
+      "output_tokens": 389,
+      "cached_tokens": 1361
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 12.0,\n \"failure_window_start_seconds\": 10.5,\n \"failure_window_end_seconds\": 14.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the pineapple inside the open drawer, but the gripper failed to position itself securely around the object, causing the fingers to slip and lift up empty.\",\n \"failure_reason_summary\": \"improper gripper alignment; failed grasp on pineapple\",\n \"recovery_action\": \"Lower the arm back into the open drawer, realign the gripper with the center of the pineapple, close the gripper securely, and lift the pineapple out of the drawer.\",\n \"recovery_action_summary\": \"realign gripper; grasp pineapple; lift out of drawer\",\n \"prevention_action\": \"Center the end-effector over the main body of the pineapple and ensure a firm grip is established before initiating the upward lift.\",\n \"prevention_action_summary\": \"center gripper on pineapple; secure grasp before lifting\",\n \"justification\": \"The robot lowers its end-effector into the drawer around 10-12 seconds to pick up the pineapple, but fails to securely grasp it. It lifts its empty gripper back up around 15 seconds, leaving the pineapple unmoved inside the drawer.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 609,
+    "failure_sim_step": 609,
+    "failure_policy_step": 101,
+    "failure_chunk_index": 3,
+    "first_post_failure_policy_step": 102,
+    "failure_type": "grasp_failure",
+    "correction_action": "Lower the arm back into the open drawer, realign the gripper with the center of the pineapple, close the gripper securely, and lift the pineapple out of the drawer."
+  }
+}
\ No newline at end of file
diff --git a/pineapple_out_of_drawer/ep016/labels.npz b/pineapple_out_of_drawer/ep016/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..8c74cfc3e4ac12eb3d7008cec3de66a670f4dc96
--- /dev/null
+++ b/pineapple_out_of_drawer/ep016/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ebaeb650671e219de0738ef8856b05b85aad924e3e1977b59073e4976251d3ef
+size 4007
diff --git a/pineapple_out_of_drawer/ep017/example.md b/pineapple_out_of_drawer/ep017/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..c05afbf172d6ef0cf757b35ea6f62dcb6a32fddf
--- /dev/null
+++ b/pineapple_out_of_drawer/ep017/example.md
@@ -0,0 +1,36 @@
+# t007__ep017__pineapple_out_of_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the pineapple out of the drawer
+- **success:** False
+- **control steps:** 902  (17.8s at 51 Hz)
+- **policy inferences:** 155  (replan_steps=6)
+- **video:** `pineapple_out_of_drawer/ep017/rollout.mp4`
+- **features:** `155` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | holding arm above table |
+| 3.0–6.0 | 152–303 | 25–50 | lowering toward open drawer |
+| 6.0–9.0 | 304–455 | 50–75 | positioning gripper near pineapple |
+| 9.0–12.0 | 456–607 | 76–101 | closing gripper on drawer edge |
+| 12.0–15.0 | 608–759 | 101–126 | holding gripper against drawer rim |
+| 15.0–17.8 | 760–901 | 126–150 | idle / no progress |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 9.03353057199211 s (00:09.0)
+- **frame / control_step:** 458 (policy_step=76, chunk_index=2)
+- **coarse onset:** 11.5 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot misaligned its end-effector over the right edge of the drawer wall instead of centering over the pineapple, causing the gripper to close without grasping the object.
+  Summary: misaligned gripper; failed grasp of pineapple
+- **recovery:** Open the gripper, lift the arm to clear the drawer, center the end-effector over the pineapple inside the drawer, and attempt the grasp again.
+  Summary: open gripper; re-align above pineapple; grasp pineapple
+- **prevention:** Align the gripper fingers precisely around the center of the pineapple before initiating the closing motion.
+  Summary: center gripper over pineapple; execute grasp
+- **justification:** Around 11.5 seconds into the video, the robot lowers its gripper to the right edge of the open drawer and closes its fingers, missing the pineapple entirely and becoming stuck against the edge.
diff --git a/pineapple_out_of_drawer/ep017/labels.json b/pineapple_out_of_drawer/ep017/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..9121f1cb1e9f61d5394c43846853c2b0e3732829
--- /dev/null
+++ b/pineapple_out_of_drawer/ep017/labels.json
@@ -0,0 +1,131 @@
+{
+  "rollout_id": "t007__ep017__pineapple_out_of_drawer",
+  "scene_variant": "",
+  "task_id": 7,
+  "task_file": "",
+  "instruction": "Take the pineapple out of the drawer",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:22:34+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9980276134122286,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "holding arm above table",
+      "description": "holding arm above table"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9980276134122286,
+      "t_end_sec": 5.996055226824457,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "lowering toward open drawer",
+      "description": "lowering toward open drawer"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.996055226824457,
+      "t_end_sec": 8.994082840236686,
+      "control_step_start": 304,
+      "control_step_end": 455,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "positioning gripper near pineapple",
+      "description": "positioning gripper near pineapple"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.994082840236686,
+      "t_end_sec": 11.992110453648914,
+      "control_step_start": 456,
+      "control_step_end": 607,
+      "policy_step_start": 76,
+      "policy_step_end": 101,
+      "phrase": "closing gripper on drawer edge",
+      "description": "closing gripper on drawer edge"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 11.992110453648914,
+      "t_end_sec": 14.990138067061142,
+      "control_step_start": 608,
+      "control_step_end": 759,
+      "policy_step_start": 101,
+      "policy_step_end": 126,
+      "phrase": "holding gripper against drawer rim",
+      "description": "holding gripper against drawer rim"
+    },
+    {
+      "segment_index": 5,
+      "t_start_sec": 14.990138067061142,
+      "t_end_sec": 17.790927021696252,
+      "control_step_start": 760,
+      "control_step_end": 901,
+      "policy_step_start": 126,
+      "policy_step_end": 150,
+      "phrase": "idle / no progress",
+      "description": "idle / no progress"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 9.03353057199211,
+    "vlm_failure_onset_timestamp": "00:09.0",
+    "vlm_failure_onset_frame": 458,
+    "vlm_failure_onset_step": 458,
+    "vlm_coarse_failure_onset_seconds": 11.5,
+    "vlm_failure_window_start_seconds": 11.0,
+    "vlm_failure_window_end_seconds": 12.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot misaligned its end-effector over the right edge of the drawer wall instead of centering over the pineapple, causing the gripper to close without grasping the object.",
+    "vlm_failure_reason_summary": "misaligned gripper; failed grasp of pineapple",
+    "vlm_recovery_action": "Open the gripper, lift the arm to clear the drawer, center the end-effector over the pineapple inside the drawer, and attempt the grasp again.",
+    "vlm_recovery_action_summary": "open gripper; re-align above pineapple; grasp pineapple",
+    "vlm_prevention_action": "Align the gripper fingers precisely around the center of the pineapple before initiating the closing motion.",
+    "vlm_prevention_action_summary": "center gripper over pineapple; execute grasp",
+    "vlm_justification": "Around 11.5 seconds into the video, the robot lowers its gripper to the right edge of the open drawer and closes its fingers, missing the pineapple entirely and becoming stuck against the edge.",
+    "vlm_temporal_justification": "At second 27 (frame 458 of original rollout), the robot gripper begins closing its fingers while positioned over the right wall of the drawer rather than centered around the pineapple, leading directly to the failed grasp.",
+    "vlm_video_fps": 50.7,
+    "vlm_video_num_frames": 902,
+    "vlm_video_duration_seconds": 17.77120315581854,
+    "vlm_usage": {
+      "input_tokens": 22218,
+      "output_tokens": 366,
+      "cached_tokens": 1225
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 11.5,\n \"failure_window_start_seconds\": 11.0,\n \"failure_window_end_seconds\": 12.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot misaligned its end-effector over the right edge of the drawer wall instead of centering over the pineapple, causing the gripper to close without grasping the object.\",\n \"failure_reason_summary\": \"misaligned gripper; failed grasp of pineapple\",\n \"recovery_action\": \"Open the gripper, lift the arm to clear the drawer, center the end-effector over the pineapple inside the drawer, and attempt the grasp again.\",\n \"recovery_action_summary\": \"open gripper; re-align above pineapple; grasp pineapple\",\n \"prevention_action\": \"Align the gripper fingers precisely around the center of the pineapple before initiating the closing motion.\",\n \"prevention_action_summary\": \"center gripper over pineapple; execute grasp\",\n \"justification\": \"Around 11.5 seconds into the video, the robot lowers its gripper to the right edge of the open drawer and closes its fingers, missing the pineapple entirely and becoming stuck against the edge.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 458,
+    "failure_sim_step": 458,
+    "failure_policy_step": 76,
+    "failure_chunk_index": 2,
+    "first_post_failure_policy_step": 77,
+    "failure_type": "grasp_failure",
+    "correction_action": "Open the gripper, lift the arm to clear the drawer, center the end-effector over the pineapple inside the drawer, and attempt the grasp again."
+  }
+}
\ No newline at end of file
diff --git a/pineapple_out_of_drawer/ep017/labels.npz b/pineapple_out_of_drawer/ep017/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..289de403f90d48be6d02e577baa8d1d21bf69531
--- /dev/null
+++ b/pineapple_out_of_drawer/ep017/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:002716e2e8247ada8c1297c1b103df59d7d0eed7d9f441fbab7284ac9e281361
+size 3916
diff --git a/pineapple_out_of_drawer/ep018/example.md b/pineapple_out_of_drawer/ep018/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..0a185fcdbb31e23f68512d8019729aafd8e7ec28
--- /dev/null
+++ b/pineapple_out_of_drawer/ep018/example.md
@@ -0,0 +1,26 @@
+# t007__ep018__pineapple_out_of_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the pineapple out of the drawer
+- **success:** True
+- **control steps:** 1072  (21.3s at 50 Hz)
+- **policy inferences:** 186  (replan_steps=6)
+- **video:** `pineapple_out_of_drawer/ep018/rollout.mp4`
+- **features:** `186` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | idle at home position |
+| 3.0–6.0 | 151–301 | 25–50 | reaching toward pineapple in drawer |
+| 6.0–9.0 | 302–452 | 50–75 | grasping pineapple inside drawer |
+| 9.0–12.0 | 453–603 | 75–100 | holding pineapple in drawer |
+| 12.0–15.0 | 604–754 | 100–125 | lifting pineapple from drawer |
+| 15.0–18.0 | 755–905 | 125–150 | moving pineapple out of drawer |
+| 18.0–21.0 | 906–1056 | 151–176 | lowering pineapple toward table |
+| 21.0–21.3 | 1057–1071 | 176–178 | placing pineapple on table |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/pineapple_out_of_drawer/ep018/labels.json b/pineapple_out_of_drawer/ep018/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..c43ebe8283d488ce1d73412013030b594d86fc44
--- /dev/null
+++ b/pineapple_out_of_drawer/ep018/labels.json
@@ -0,0 +1,123 @@
+{
+  "rollout_id": "t007__ep018__pineapple_out_of_drawer",
+  "scene_variant": "",
+  "task_id": 7,
+  "task_file": "",
+  "instruction": "Take the pineapple out of the drawer",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:22:43+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.996031746031746,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "idle at home position",
+      "description": "idle at home position"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.996031746031746,
+      "t_end_sec": 5.992063492063492,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "reaching toward pineapple in drawer",
+      "description": "reaching toward pineapple in drawer"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.992063492063492,
+      "t_end_sec": 8.988095238095239,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "grasping pineapple inside drawer",
+      "description": "grasping pineapple inside drawer"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.988095238095239,
+      "t_end_sec": 11.984126984126984,
+      "control_step_start": 453,
+      "control_step_end": 603,
+      "policy_step_start": 75,
+      "policy_step_end": 100,
+      "phrase": "holding pineapple in drawer",
+      "description": "holding pineapple in drawer"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 11.984126984126984,
+      "t_end_sec": 14.980158730158731,
+      "control_step_start": 604,
+      "control_step_end": 754,
+      "policy_step_start": 100,
+      "policy_step_end": 125,
+      "phrase": "lifting pineapple from drawer",
+      "description": "lifting pineapple from drawer"
+    },
+    {
+      "segment_index": 5,
+      "t_start_sec": 14.980158730158731,
+      "t_end_sec": 17.976190476190478,
+      "control_step_start": 755,
+      "control_step_end": 905,
+      "policy_step_start": 125,
+      "policy_step_end": 150,
+      "phrase": "moving pineapple out of drawer",
+      "description": "moving pineapple out of drawer"
+    },
+    {
+      "segment_index": 6,
+      "t_start_sec": 17.976190476190478,
+      "t_end_sec": 20.97222222222222,
+      "control_step_start": 906,
+      "control_step_end": 1056,
+      "policy_step_start": 151,
+      "policy_step_end": 176,
+      "phrase": "lowering pineapple toward table",
+      "description": "lowering pineapple toward table"
+    },
+    {
+      "segment_index": 7,
+      "t_start_sec": 20.97222222222222,
+      "t_end_sec": 21.26984126984127,
+      "control_step_start": 1057,
+      "control_step_end": 1071,
+      "policy_step_start": 176,
+      "policy_step_end": 178,
+      "phrase": "placing pineapple on table",
+      "description": "placing pineapple on table"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/pineapple_out_of_drawer/ep018/labels.npz b/pineapple_out_of_drawer/ep018/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..bc0c62dc17f7fec390af0ce02c7ddc6c7a8cfa59
--- /dev/null
+++ b/pineapple_out_of_drawer/ep018/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:800723479d05c3fb36e5e43882fb4ca04eb9404b9105b3a9a0734d9633dee740
+size 3272
diff --git a/pineapple_out_of_drawer/ep019/example.md b/pineapple_out_of_drawer/ep019/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..589ba6c7d441be902a9725c87bc23c63dae4dd63
--- /dev/null
+++ b/pineapple_out_of_drawer/ep019/example.md
@@ -0,0 +1,28 @@
+# t007__ep019__pineapple_out_of_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the pineapple out of the drawer
+- **success:** True
+- **control steps:** 1427  (28.3s at 50 Hz)
+- **policy inferences:** 246  (replan_steps=6)
+- **video:** `pineapple_out_of_drawer/ep019/rollout.mp4`
+- **features:** `246` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | idle / no motion |
+| 3.0–6.0 | 152–303 | 25–50 | reaching toward drawer |
+| 6.0–9.0 | 304–455 | 50–75 | positioning gripper above drawer |
+| 9.0–12.0 | 456–607 | 76–101 | adjusting position over drawer |
+| 12.0–15.0 | 608–759 | 101–126 | lowering gripper toward pineapple |
+| 15.0–18.1 | 760–911 | 126–151 | reaching into drawer for pineapple |
+| 18.1–21.1 | 912–1063 | 152–177 | grasping and lifting pineapple |
+| 21.1–24.1 | 1064–1215 | 177–202 | moving pineapple to table |
+| 24.1–27.1 | 1216–1367 | 202–227 | placing pineapple on table |
+| 27.1–28.3 | 1368–1426 | 228–237 | idle after releasing pineapple |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/pineapple_out_of_drawer/ep019/labels.json b/pineapple_out_of_drawer/ep019/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..2213a002c84161796efc676fbd12e5d0d9b4a271
--- /dev/null
+++ b/pineapple_out_of_drawer/ep019/labels.json
@@ -0,0 +1,145 @@
+{
+  "rollout_id": "t007__ep019__pineapple_out_of_drawer",
+  "scene_variant": "",
+  "task_id": 7,
+  "task_file": "",
+  "instruction": "Take the pineapple out of the drawer",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:22:54+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00990099009901,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "idle / no motion",
+      "description": "idle / no motion"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00990099009901,
+      "t_end_sec": 6.01980198019802,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "reaching toward drawer",
+      "description": "reaching toward drawer"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01980198019802,
+      "t_end_sec": 9.029702970297029,
+      "control_step_start": 304,
+      "control_step_end": 455,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "positioning gripper above drawer",
+      "description": "positioning gripper above drawer"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.029702970297029,
+      "t_end_sec": 12.03960396039604,
+      "control_step_start": 456,
+      "control_step_end": 607,
+      "policy_step_start": 76,
+      "policy_step_end": 101,
+      "phrase": "adjusting position over drawer",
+      "description": "adjusting position over drawer"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 12.03960396039604,
+      "t_end_sec": 15.049504950495049,
+      "control_step_start": 608,
+      "control_step_end": 759,
+      "policy_step_start": 101,
+      "policy_step_end": 126,
+      "phrase": "lowering gripper toward pineapple",
+      "description": "lowering gripper toward pineapple"
+    },
+    {
+      "segment_index": 5,
+      "t_start_sec": 15.049504950495049,
+      "t_end_sec": 18.059405940594058,
+      "control_step_start": 760,
+      "control_step_end": 911,
+      "policy_step_start": 126,
+      "policy_step_end": 151,
+      "phrase": "reaching into drawer for pineapple",
+      "description": "reaching into drawer for pineapple"
+    },
+    {
+      "segment_index": 6,
+      "t_start_sec": 18.059405940594058,
+      "t_end_sec": 21.06930693069307,
+      "control_step_start": 912,
+      "control_step_end": 1063,
+      "policy_step_start": 152,
+      "policy_step_end": 177,
+      "phrase": "grasping and lifting pineapple",
+      "description": "grasping and lifting pineapple"
+    },
+    {
+      "segment_index": 7,
+      "t_start_sec": 21.06930693069307,
+      "t_end_sec": 24.07920792079208,
+      "control_step_start": 1064,
+      "control_step_end": 1215,
+      "policy_step_start": 177,
+      "policy_step_end": 202,
+      "phrase": "moving pineapple to table",
+      "description": "moving pineapple to table"
+    },
+    {
+      "segment_index": 8,
+      "t_start_sec": 24.07920792079208,
+      "t_end_sec": 27.08910891089109,
+      "control_step_start": 1216,
+      "control_step_end": 1367,
+      "policy_step_start": 202,
+      "policy_step_end": 227,
+      "phrase": "placing pineapple on table",
+      "description": "placing pineapple on table"
+    },
+    {
+      "segment_index": 9,
+      "t_start_sec": 27.08910891089109,
+      "t_end_sec": 28.257425742574256,
+      "control_step_start": 1368,
+      "control_step_end": 1426,
+      "policy_step_start": 228,
+      "policy_step_end": 237,
+      "phrase": "idle after releasing pineapple",
+      "description": "idle after releasing pineapple"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/pineapple_out_of_drawer/ep019/labels.npz b/pineapple_out_of_drawer/ep019/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..af763aa1c9ac591653cbb147674d59623a13351f
--- /dev/null
+++ b/pineapple_out_of_drawer/ep019/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:deca0b026a14567464416834eda13f4a993f78a11fdcce431e9a3ebd4543201a
+size 3361
diff --git a/pineapple_out_of_drawer/ep020/example.md b/pineapple_out_of_drawer/ep020/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..fd509c8691f38fba36418467fbd0fcad2c52a28c
--- /dev/null
+++ b/pineapple_out_of_drawer/ep020/example.md
@@ -0,0 +1,37 @@
+# t007__ep020__pineapple_out_of_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the pineapple out of the drawer
+- **success:** False
+- **control steps:** 946  (18.8s at 50 Hz)
+- **policy inferences:** 162  (replan_steps=6)
+- **video:** `pineapple_out_of_drawer/ep020/rollout.mp4`
+- **features:** `162` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | holding position above table |
+| 3.0–6.0 | 151–301 | 25–50 | moving gripper toward drawer |
+| 6.0–9.0 | 302–452 | 50–75 | lowering end-effector to drawer |
+| 9.0–12.0 | 453–603 | 75–100 | attempting grasp on pineapple |
+| 12.0–15.0 | 604–754 | 100–125 | lifting empty gripper upward |
+| 15.0–18.0 | 755–905 | 125–150 | hovering above open drawer |
+| 18.0–18.8 | 906–945 | 151–157 | idle above drawer |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 10.218253968253968 s (00:10.2)
+- **frame / control_step:** 515 (policy_step=85, chunk_index=5)
+- **coarse onset:** 11.5 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the pineapple inside the open drawer, but the gripper closed off-center and failed to clamp securely, lifting away empty.
+  Summary: failed grasp on pineapple; gripper lifts away empty
+- **recovery:** Lower the end-effector back into the drawer, position the fingers around the center of the pineapple, close the gripper firmly, and lift the pineapple out.
+  Summary: re-align gripper; grasp pineapple firmly; lift out of drawer
+- **prevention:** Center the gripper over the middle of the pineapple and descend deeper into the drawer before closing the fingers to ensure a secure grip.
+  Summary: center gripper over pineapple; descend fully before closing
+- **justification:** At approximately 11.5 seconds, the robot's gripper attempts to grasp the pineapple but fails to secure it. The gripper lifts up and away empty by 13.5 seconds, leaving the pineapple in the drawer for the rest of the rollout.
diff --git a/pineapple_out_of_drawer/ep020/labels.json b/pineapple_out_of_drawer/ep020/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..c7adf8d7b082e5c8da8d635992f5249644aa69f3
--- /dev/null
+++ b/pineapple_out_of_drawer/ep020/labels.json
@@ -0,0 +1,142 @@
+{
+  "rollout_id": "t007__ep020__pineapple_out_of_drawer",
+  "scene_variant": "",
+  "task_id": 7,
+  "task_file": "",
+  "instruction": "Take the pineapple out of the drawer",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:23:25+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.996031746031746,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "holding position above table",
+      "description": "holding position above table"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.996031746031746,
+      "t_end_sec": 5.992063492063492,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "moving gripper toward drawer",
+      "description": "moving gripper toward drawer"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.992063492063492,
+      "t_end_sec": 8.988095238095239,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "lowering end-effector to drawer",
+      "description": "lowering end-effector to drawer"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.988095238095239,
+      "t_end_sec": 11.984126984126984,
+      "control_step_start": 453,
+      "control_step_end": 603,
+      "policy_step_start": 75,
+      "policy_step_end": 100,
+      "phrase": "attempting grasp on pineapple",
+      "description": "attempting grasp on pineapple"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 11.984126984126984,
+      "t_end_sec": 14.980158730158731,
+      "control_step_start": 604,
+      "control_step_end": 754,
+      "policy_step_start": 100,
+      "policy_step_end": 125,
+      "phrase": "lifting empty gripper upward",
+      "description": "lifting empty gripper upward"
+    },
+    {
+      "segment_index": 5,
+      "t_start_sec": 14.980158730158731,
+      "t_end_sec": 17.976190476190478,
+      "control_step_start": 755,
+      "control_step_end": 905,
+      "policy_step_start": 125,
+      "policy_step_end": 150,
+      "phrase": "hovering above open drawer",
+      "description": "hovering above open drawer"
+    },
+    {
+      "segment_index": 6,
+      "t_start_sec": 17.976190476190478,
+      "t_end_sec": 18.76984126984127,
+      "control_step_start": 906,
+      "control_step_end": 945,
+      "policy_step_start": 151,
+      "policy_step_end": 157,
+      "phrase": "idle above drawer",
+      "description": "idle above drawer"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 10.218253968253968,
+    "vlm_failure_onset_timestamp": "00:10.2",
+    "vlm_failure_onset_frame": 515,
+    "vlm_failure_onset_step": 515,
+    "vlm_coarse_failure_onset_seconds": 11.5,
+    "vlm_failure_window_start_seconds": 10.5,
+    "vlm_failure_window_end_seconds": 13.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the pineapple inside the open drawer, but the gripper closed off-center and failed to clamp securely, lifting away empty.",
+    "vlm_failure_reason_summary": "failed grasp on pineapple; gripper lifts away empty",
+    "vlm_recovery_action": "Lower the end-effector back into the drawer, position the fingers around the center of the pineapple, close the gripper firmly, and lift the pineapple out.",
+    "vlm_recovery_action_summary": "re-align gripper; grasp pineapple firmly; lift out of drawer",
+    "vlm_prevention_action": "Center the gripper over the middle of the pineapple and descend deeper into the drawer before closing the fingers to ensure a secure grip.",
+    "vlm_prevention_action_summary": "center gripper over pineapple; descend fully before closing",
+    "vlm_justification": "At approximately 11.5 seconds, the robot's gripper attempts to grasp the pineapple but fails to secure it. The gripper lifts up and away empty by 13.5 seconds, leaving the pineapple in the drawer for the rest of the rollout.",
+    "vlm_temporal_justification": "At second 01:26 (frame 86 of the clip), the robot gripper begins closing off-center over the pineapple in the drawer, failing to enclose the object properly before lifting away empty.",
+    "vlm_video_fps": 50.4,
+    "vlm_video_num_frames": 946,
+    "vlm_video_duration_seconds": 18.75,
+    "vlm_usage": {
+      "input_tokens": 22151,
+      "output_tokens": 387,
+      "cached_tokens": 1293
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 11.5,\n \"failure_window_start_seconds\": 10.5,\n \"failure_window_end_seconds\": 13.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the pineapple inside the open drawer, but the gripper closed off-center and failed to clamp securely, lifting away empty.\",\n \"failure_reason_summary\": \"failed grasp on pineapple; gripper lifts away empty\",\n \"recovery_action\": \"Lower the end-effector back into the drawer, position the fingers around the center of the pineapple, close the gripper firmly, and lift the pineapple out.\",\n \"recovery_action_summary\": \"re-align gripper; grasp pineapple firmly; lift out of drawer\",\n \"prevention_action\": \"Center the gripper over the middle of the pineapple and descend deeper into the drawer before closing the fingers to ensure a secure grip.\",\n \"prevention_action_summary\": \"center gripper over pineapple; descend fully before closing\",\n \"justification\": \"At approximately 11.5 seconds, the robot's gripper attempts to grasp the pineapple but fails to secure it. The gripper lifts up and away empty by 13.5 seconds, leaving the pineapple in the drawer for the rest of the rollout.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 515,
+    "failure_sim_step": 515,
+    "failure_policy_step": 85,
+    "failure_chunk_index": 5,
+    "first_post_failure_policy_step": 86,
+    "failure_type": "grasp_failure",
+    "correction_action": "Lower the end-effector back into the drawer, position the fingers around the center of the pineapple, close the gripper firmly, and lift the pineapple out."
+  }
+}
\ No newline at end of file
diff --git a/pineapple_out_of_drawer/ep020/labels.npz b/pineapple_out_of_drawer/ep020/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..798194f952ad1bc0b0e95d175073e855dd44f30d
--- /dev/null
+++ b/pineapple_out_of_drawer/ep020/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:35163ecf8043b59dea34e1dfebad75b7ec1b16d60758585c6140a612249f4bda
+size 4009
diff --git a/pineapple_out_of_drawer/ep021/example.md b/pineapple_out_of_drawer/ep021/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..def6c916173c6998c274e097a5d9ebeb1f829032
--- /dev/null
+++ b/pineapple_out_of_drawer/ep021/example.md
@@ -0,0 +1,38 @@
+# t007__ep021__pineapple_out_of_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the pineapple out of the drawer
+- **success:** False
+- **control steps:** 1118  (22.2s at 50 Hz)
+- **policy inferences:** 192  (replan_steps=6)
+- **video:** `pineapple_out_of_drawer/ep021/rollout.mp4`
+- **features:** `192` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | idle above drawer |
+| 3.0–6.0 | 151–301 | 25–50 | reaching toward pineapple |
+| 6.0–9.0 | 302–452 | 50–75 | attempting grasp on pineapple |
+| 9.0–12.0 | 453–603 | 75–100 | lifting empty; repositioning gripper |
+| 12.0–15.0 | 604–754 | 100–125 | attempting second grasp |
+| 15.0–18.0 | 755–905 | 125–150 | lifting up empty |
+| 18.0–21.0 | 906–1056 | 151–176 | hovering above drawer |
+| 21.0–22.2 | 1057–1117 | 176–186 | idle above drawer |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 7.5396825396825395 s (00:07.5)
+- **frame / control_step:** 380 (policy_step=63, chunk_index=2)
+- **coarse onset:** 8.0 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the pineapple inside the open drawer, but failed to establish a secure hold, causing the gripper to slip off and lift up empty.
+  Summary: unsuccessful grasp attempt; gripper slipping off object
+- **recovery:** Re-align the gripper directly over the pineapple, lower it deeper into the drawer, and close the fingers firmly around the object before lifting it out.
+  Summary: realign gripper; grasp pineapple firmly; lift out of drawer
+- **prevention:** Lower the gripper further into the drawer and center it accurately over the pineapple before closing the fingers to ensure a firm grasp.
+  Summary: position gripper lower; center grip on pineapple
+- **justification:** At 00:08, the gripper closes around the pineapple and lifts upwards, but the object slips out of the fingers and remains in the drawer. Subsequent attempts also fail to secure the object.
diff --git a/pineapple_out_of_drawer/ep021/labels.json b/pineapple_out_of_drawer/ep021/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..7837f4cbe2c286181dec73c1baded0473f55fe8c
--- /dev/null
+++ b/pineapple_out_of_drawer/ep021/labels.json
@@ -0,0 +1,153 @@
+{
+  "rollout_id": "t007__ep021__pineapple_out_of_drawer",
+  "scene_variant": "",
+  "task_id": 7,
+  "task_file": "",
+  "instruction": "Take the pineapple out of the drawer",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:23:58+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.996031746031746,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "idle above drawer",
+      "description": "idle above drawer"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.996031746031746,
+      "t_end_sec": 5.992063492063492,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "reaching toward pineapple",
+      "description": "reaching toward pineapple"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.992063492063492,
+      "t_end_sec": 8.988095238095239,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "attempting grasp on pineapple",
+      "description": "attempting grasp on pineapple"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.988095238095239,
+      "t_end_sec": 11.984126984126984,
+      "control_step_start": 453,
+      "control_step_end": 603,
+      "policy_step_start": 75,
+      "policy_step_end": 100,
+      "phrase": "lifting empty; repositioning gripper",
+      "description": "lifting empty; repositioning gripper"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 11.984126984126984,
+      "t_end_sec": 14.980158730158731,
+      "control_step_start": 604,
+      "control_step_end": 754,
+      "policy_step_start": 100,
+      "policy_step_end": 125,
+      "phrase": "attempting second grasp",
+      "description": "attempting second grasp"
+    },
+    {
+      "segment_index": 5,
+      "t_start_sec": 14.980158730158731,
+      "t_end_sec": 17.976190476190478,
+      "control_step_start": 755,
+      "control_step_end": 905,
+      "policy_step_start": 125,
+      "policy_step_end": 150,
+      "phrase": "lifting up empty",
+      "description": "lifting up empty"
+    },
+    {
+      "segment_index": 6,
+      "t_start_sec": 17.976190476190478,
+      "t_end_sec": 20.97222222222222,
+      "control_step_start": 906,
+      "control_step_end": 1056,
+      "policy_step_start": 151,
+      "policy_step_end": 176,
+      "phrase": "hovering above drawer",
+      "description": "hovering above drawer"
+    },
+    {
+      "segment_index": 7,
+      "t_start_sec": 20.97222222222222,
+      "t_end_sec": 22.182539682539684,
+      "control_step_start": 1057,
+      "control_step_end": 1117,
+      "policy_step_start": 176,
+      "policy_step_end": 186,
+      "phrase": "idle above drawer",
+      "description": "idle above drawer"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 7.5396825396825395,
+    "vlm_failure_onset_timestamp": "00:07.5",
+    "vlm_failure_onset_frame": 380,
+    "vlm_failure_onset_step": 380,
+    "vlm_coarse_failure_onset_seconds": 8.0,
+    "vlm_failure_window_start_seconds": 7.0,
+    "vlm_failure_window_end_seconds": 9.0,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the pineapple inside the open drawer, but failed to establish a secure hold, causing the gripper to slip off and lift up empty.",
+    "vlm_failure_reason_summary": "unsuccessful grasp attempt; gripper slipping off object",
+    "vlm_recovery_action": "Re-align the gripper directly over the pineapple, lower it deeper into the drawer, and close the fingers firmly around the object before lifting it out.",
+    "vlm_recovery_action_summary": "realign gripper; grasp pineapple firmly; lift out of drawer",
+    "vlm_prevention_action": "Lower the gripper further into the drawer and center it accurately over the pineapple before closing the fingers to ensure a firm grasp.",
+    "vlm_prevention_action_summary": "position gripper lower; center grip on pineapple",
+    "vlm_justification": "At 00:08, the gripper closes around the pineapple and lifts upwards, but the object slips out of the fingers and remains in the drawer. Subsequent attempts also fail to secure the object.",
+    "vlm_temporal_justification": "At second 128 (02:08 in the clip), the robot arm begins moving upward after closing its gripper around the pineapple, but the pineapple remains in the drawer, making the grasp failure visibly apparent.",
+    "vlm_video_fps": 50.4,
+    "vlm_video_num_frames": 1118,
+    "vlm_video_duration_seconds": 22.162698412698415,
+    "vlm_usage": {
+      "input_tokens": 22341,
+      "output_tokens": 373,
+      "cached_tokens": 1497
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 8.0,\n \"failure_window_start_seconds\": 7.0,\n \"failure_window_end_seconds\": 9.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the pineapple inside the open drawer, but failed to establish a secure hold, causing the gripper to slip off and lift up empty.\",\n \"failure_reason_summary\": \"unsuccessful grasp attempt; gripper slipping off object\",\n \"recovery_action\": \"Re-align the gripper directly over the pineapple, lower it deeper into the drawer, and close the fingers firmly around the object before lifting it out.\",\n \"recovery_action_summary\": \"realign gripper; grasp pineapple firmly; lift out of drawer\",\n \"prevention_action\": \"Lower the gripper further into the drawer and center it accurately over the pineapple before closing the fingers to ensure a firm grasp.\",\n \"prevention_action_summary\": \"position gripper lower; center grip on pineapple\",\n \"justification\": \"At 00:08, the gripper closes around the pineapple and lifts upwards, but the object slips out of the fingers and remains in the drawer. Subsequent attempts also fail to secure the object.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 380,
+    "failure_sim_step": 380,
+    "failure_policy_step": 63,
+    "failure_chunk_index": 2,
+    "first_post_failure_policy_step": 64,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-align the gripper directly over the pineapple, lower it deeper into the drawer, and close the fingers firmly around the object before lifting it out."
+  }
+}
\ No newline at end of file
diff --git a/pineapple_out_of_drawer/ep021/labels.npz b/pineapple_out_of_drawer/ep021/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..54a08deac718fe0ce47eee9dd34cafc91a2c0d69
--- /dev/null
+++ b/pineapple_out_of_drawer/ep021/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9467e1318159bdb999b9f24461a5e6632e650e9d3f6073143b4f5f0473529f79
+size 4018
diff --git a/pineapple_out_of_drawer/ep022/example.md b/pineapple_out_of_drawer/ep022/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..78a5c47be6c1db19d2d7583a69f63e11b2780d09
--- /dev/null
+++ b/pineapple_out_of_drawer/ep022/example.md
@@ -0,0 +1,37 @@
+# t007__ep022__pineapple_out_of_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the pineapple out of the drawer
+- **success:** False
+- **control steps:** 967  (19.1s at 51 Hz)
+- **policy inferences:** 166  (replan_steps=6)
+- **video:** `pineapple_out_of_drawer/ep022/rollout.mp4`
+- **features:** `166` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | hovering above drawer |
+| 3.0–6.0 | 152–303 | 25–50 | lowering arm toward drawer |
+| 6.0–9.0 | 304–455 | 50–75 | positioning gripper over pineapple |
+| 9.0–12.0 | 456–607 | 76–101 | closing gripper on pineapple |
+| 12.0–15.0 | 608–759 | 101–126 | lifting arm; losing grip on pineapple |
+| 15.0–18.0 | 760–911 | 126–151 | retracting empty gripper |
+| 18.0–19.1 | 912–966 | 152–161 | idle above drawer |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 11.321499013806706 s (00:11.3)
+- **frame / control_step:** 574 (policy_step=95, chunk_index=4)
+- **coarse onset:** 12.0 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the pineapple near its top leaves, but the fingers slipped off as the arm lifted, leaving the pineapple in the drawer.
+  Summary: failed grasp on pineapple; slipping during lift
+- **recovery:** Lower the end-effector back into the drawer, position the gripper securely around the pineapple's main body, close the fingers, and lift it out.
+  Summary: re-align gripper; grasp pineapple body; lift out of drawer
+- **prevention:** Position the gripper lower around the solid body of the pineapple rather than the top leaves to establish a secure hold before lifting.
+  Summary: grasp pineapple center; secure firm grip before lifting
+- **justification:** At around 12.0 seconds, the robot attempts to lift the pineapple, but its grip on the upper leaves is insufficient, causing the pineapple to slip out of the gripper and remain in the drawer.
diff --git a/pineapple_out_of_drawer/ep022/labels.json b/pineapple_out_of_drawer/ep022/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..23bd3ead6b1c17d61b228a9a12694a93d9db1a27
--- /dev/null
+++ b/pineapple_out_of_drawer/ep022/labels.json
@@ -0,0 +1,142 @@
+{
+  "rollout_id": "t007__ep022__pineapple_out_of_drawer",
+  "scene_variant": "",
+  "task_id": 7,
+  "task_file": "",
+  "instruction": "Take the pineapple out of the drawer",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:24:32+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9980276134122286,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "hovering above drawer",
+      "description": "hovering above drawer"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9980276134122286,
+      "t_end_sec": 5.996055226824457,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "lowering arm toward drawer",
+      "description": "lowering arm toward drawer"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.996055226824457,
+      "t_end_sec": 8.994082840236686,
+      "control_step_start": 304,
+      "control_step_end": 455,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "positioning gripper over pineapple",
+      "description": "positioning gripper over pineapple"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.994082840236686,
+      "t_end_sec": 11.992110453648914,
+      "control_step_start": 456,
+      "control_step_end": 607,
+      "policy_step_start": 76,
+      "policy_step_end": 101,
+      "phrase": "closing gripper on pineapple",
+      "description": "closing gripper on pineapple"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 11.992110453648914,
+      "t_end_sec": 14.990138067061142,
+      "control_step_start": 608,
+      "control_step_end": 759,
+      "policy_step_start": 101,
+      "policy_step_end": 126,
+      "phrase": "lifting arm; losing grip on pineapple",
+      "description": "lifting arm; losing grip on pineapple"
+    },
+    {
+      "segment_index": 5,
+      "t_start_sec": 14.990138067061142,
+      "t_end_sec": 17.988165680473372,
+      "control_step_start": 760,
+      "control_step_end": 911,
+      "policy_step_start": 126,
+      "policy_step_end": 151,
+      "phrase": "retracting empty gripper",
+      "description": "retracting empty gripper"
+    },
+    {
+      "segment_index": 6,
+      "t_start_sec": 17.988165680473372,
+      "t_end_sec": 19.072978303747533,
+      "control_step_start": 912,
+      "control_step_end": 966,
+      "policy_step_start": 152,
+      "policy_step_end": 161,
+      "phrase": "idle above drawer",
+      "description": "idle above drawer"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 11.321499013806706,
+    "vlm_failure_onset_timestamp": "00:11.3",
+    "vlm_failure_onset_frame": 574,
+    "vlm_failure_onset_step": 574,
+    "vlm_coarse_failure_onset_seconds": 12.0,
+    "vlm_failure_window_start_seconds": 11.0,
+    "vlm_failure_window_end_seconds": 14.0,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the pineapple near its top leaves, but the fingers slipped off as the arm lifted, leaving the pineapple in the drawer.",
+    "vlm_failure_reason_summary": "failed grasp on pineapple; slipping during lift",
+    "vlm_recovery_action": "Lower the end-effector back into the drawer, position the gripper securely around the pineapple's main body, close the fingers, and lift it out.",
+    "vlm_recovery_action_summary": "re-align gripper; grasp pineapple body; lift out of drawer",
+    "vlm_prevention_action": "Position the gripper lower around the solid body of the pineapple rather than the top leaves to establish a secure hold before lifting.",
+    "vlm_prevention_action_summary": "grasp pineapple center; secure firm grip before lifting",
+    "vlm_justification": "At around 12.0 seconds, the robot attempts to lift the pineapple, but its grip on the upper leaves is insufficient, causing the pineapple to slip out of the gripper and remain in the drawer.",
+    "vlm_temporal_justification": "At second 01:58 (frame 118 of the clip), the robot's gripper fully closes on the top green leaves of the pineapple instead of around its body, making the grasp failure inevitable once lifting begins.",
+    "vlm_video_fps": 50.7,
+    "vlm_video_num_frames": 967,
+    "vlm_video_duration_seconds": 19.05325443786982,
+    "vlm_usage": {
+      "input_tokens": 22277,
+      "output_tokens": 380,
+      "cached_tokens": 1293
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 12.0,\n \"failure_window_start_seconds\": 11.0,\n \"failure_window_end_seconds\": 14.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the pineapple near its top leaves, but the fingers slipped off as the arm lifted, leaving the pineapple in the drawer.\",\n \"failure_reason_summary\": \"failed grasp on pineapple; slipping during lift\",\n \"recovery_action\": \"Lower the end-effector back into the drawer, position the gripper securely around the pineapple's main body, close the fingers, and lift it out.\",\n \"recovery_action_summary\": \"re-align gripper; grasp pineapple body; lift out of drawer\",\n \"prevention_action\": \"Position the gripper lower around the solid body of the pineapple rather than the top leaves to establish a secure hold before lifting.\",\n \"prevention_action_summary\": \"grasp pineapple center; secure firm grip before lifting\",\n \"justification\": \"At around 12.0 seconds, the robot attempts to lift the pineapple, but its grip on the upper leaves is insufficient, causing the pineapple to slip out of the gripper and remain in the drawer.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 574,
+    "failure_sim_step": 574,
+    "failure_policy_step": 95,
+    "failure_chunk_index": 4,
+    "first_post_failure_policy_step": 96,
+    "failure_type": "grasp_failure",
+    "correction_action": "Lower the end-effector back into the drawer, position the gripper securely around the pineapple's main body, close the fingers, and lift it out."
+  }
+}
\ No newline at end of file
diff --git a/pineapple_out_of_drawer/ep022/labels.npz b/pineapple_out_of_drawer/ep022/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..5a57343c3b5619b781974f22dcfb3edc60f65dda
--- /dev/null
+++ b/pineapple_out_of_drawer/ep022/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ba47aa0a8863ed9e1a646453b7254b3d6633a8a4a46dcac6cf9c18203e2daec4
+size 3967
diff --git a/pineapple_out_of_drawer/ep023/example.md b/pineapple_out_of_drawer/ep023/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..066594fc2ca6fcd59b69215a2d4d8073533eef41
--- /dev/null
+++ b/pineapple_out_of_drawer/ep023/example.md
@@ -0,0 +1,36 @@
+# t007__ep023__pineapple_out_of_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the pineapple out of the drawer
+- **success:** False
+- **control steps:** 832  (16.5s at 50 Hz)
+- **policy inferences:** 143  (replan_steps=6)
+- **video:** `pineapple_out_of_drawer/ep023/rollout.mp4`
+- **features:** `143` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | holding position above table |
+| 3.0–6.0 | 152–303 | 25–50 | descending toward open drawer |
+| 6.0–9.0 | 304–455 | 50–75 | reaching into drawer for pineapple |
+| 9.0–12.0 | 456–607 | 76–101 | lifting empty gripper from drawer |
+| 12.0–15.0 | 608–759 | 101–126 | hovering empty-handed above table |
+| 15.0–16.5 | 760–831 | 126–138 | remaining stationary above table |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 7.683168316831683 s (00:07.7)
+- **frame / control_step:** 388 (policy_step=64, chunk_index=4)
+- **coarse onset:** 8.0 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the pineapple inside the open drawer, but the gripper failed to establish a secure hold and lifted up empty.
+  Summary: failed grasp on pineapple; lifting empty gripper
+- **recovery:** Re-align the end-effector over the pineapple in the drawer, lower the gripper further to grasp it securely, and lift it out of the drawer.
+  Summary: realign gripper over pineapple; grasp and lift out
+- **prevention:** Position the gripper directly over the center of the pineapple and descend low enough into the drawer box to close the fingers firmly around the object.
+  Summary: center gripper around pineapple; ensure deep grasp
+- **justification:** The robot reaches into the open drawer at around 8.0 seconds to grab the pineapple, but closes its fingers loosely or misaligned. By 11.0 seconds, the arm retracts upwards while leaving the pineapple behind in the drawer.
diff --git a/pineapple_out_of_drawer/ep023/labels.json b/pineapple_out_of_drawer/ep023/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..ea65c2d6fc57f3e5ccdc466047634c455bc37933
--- /dev/null
+++ b/pineapple_out_of_drawer/ep023/labels.json
@@ -0,0 +1,131 @@
+{
+  "rollout_id": "t007__ep023__pineapple_out_of_drawer",
+  "scene_variant": "",
+  "task_id": 7,
+  "task_file": "",
+  "instruction": "Take the pineapple out of the drawer",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:25:06+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00990099009901,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "holding position above table",
+      "description": "holding position above table"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00990099009901,
+      "t_end_sec": 6.01980198019802,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "descending toward open drawer",
+      "description": "descending toward open drawer"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01980198019802,
+      "t_end_sec": 9.029702970297029,
+      "control_step_start": 304,
+      "control_step_end": 455,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "reaching into drawer for pineapple",
+      "description": "reaching into drawer for pineapple"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.029702970297029,
+      "t_end_sec": 12.03960396039604,
+      "control_step_start": 456,
+      "control_step_end": 607,
+      "policy_step_start": 76,
+      "policy_step_end": 101,
+      "phrase": "lifting empty gripper from drawer",
+      "description": "lifting empty gripper from drawer"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 12.03960396039604,
+      "t_end_sec": 15.049504950495049,
+      "control_step_start": 608,
+      "control_step_end": 759,
+      "policy_step_start": 101,
+      "policy_step_end": 126,
+      "phrase": "hovering empty-handed above table",
+      "description": "hovering empty-handed above table"
+    },
+    {
+      "segment_index": 5,
+      "t_start_sec": 15.049504950495049,
+      "t_end_sec": 16.475247524752476,
+      "control_step_start": 760,
+      "control_step_end": 831,
+      "policy_step_start": 126,
+      "policy_step_end": 138,
+      "phrase": "remaining stationary above table",
+      "description": "remaining stationary above table"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 7.683168316831683,
+    "vlm_failure_onset_timestamp": "00:07.7",
+    "vlm_failure_onset_frame": 388,
+    "vlm_failure_onset_step": 388,
+    "vlm_coarse_failure_onset_seconds": 8.0,
+    "vlm_failure_window_start_seconds": 7.0,
+    "vlm_failure_window_end_seconds": 11.0,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the pineapple inside the open drawer, but the gripper failed to establish a secure hold and lifted up empty.",
+    "vlm_failure_reason_summary": "failed grasp on pineapple; lifting empty gripper",
+    "vlm_recovery_action": "Re-align the end-effector over the pineapple in the drawer, lower the gripper further to grasp it securely, and lift it out of the drawer.",
+    "vlm_recovery_action_summary": "realign gripper over pineapple; grasp and lift out",
+    "vlm_prevention_action": "Position the gripper directly over the center of the pineapple and descend low enough into the drawer box to close the fingers firmly around the object.",
+    "vlm_prevention_action_summary": "center gripper around pineapple; ensure deep grasp",
+    "vlm_justification": "The robot reaches into the open drawer at around 8.0 seconds to grab the pineapple, but closes its fingers loosely or misaligned. By 11.0 seconds, the arm retracts upwards while leaving the pineapple behind in the drawer.",
+    "vlm_temporal_justification": "At 02:16 in the clip, the robot's gripper fingers close completely together without securing the pineapple, marking the precise onset of the grasp failure.",
+    "vlm_video_fps": 50.5,
+    "vlm_video_num_frames": 832,
+    "vlm_video_duration_seconds": 16.455445544554454,
+    "vlm_usage": {
+      "input_tokens": 22084,
+      "output_tokens": 368,
+      "cached_tokens": 1089
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 8.0,\n \"failure_window_start_seconds\": 7.0,\n \"failure_window_end_seconds\": 11.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the pineapple inside the open drawer, but the gripper failed to establish a secure hold and lifted up empty.\",\n \"failure_reason_summary\": \"failed grasp on pineapple; lifting empty gripper\",\n \"recovery_action\": \"Re-align the end-effector over the pineapple in the drawer, lower the gripper further to grasp it securely, and lift it out of the drawer.\",\n \"recovery_action_summary\": \"realign gripper over pineapple; grasp and lift out\",\n \"prevention_action\": \"Position the gripper directly over the center of the pineapple and descend low enough into the drawer box to close the fingers firmly around the object.\",\n \"prevention_action_summary\": \"center gripper around pineapple; ensure deep grasp\",\n \"justification\": \"The robot reaches into the open drawer at around 8.0 seconds to grab the pineapple, but closes its fingers loosely or misaligned. By 11.0 seconds, the arm retracts upwards while leaving the pineapple behind in the drawer.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 388,
+    "failure_sim_step": 388,
+    "failure_policy_step": 64,
+    "failure_chunk_index": 4,
+    "first_post_failure_policy_step": 65,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-align the end-effector over the pineapple in the drawer, lower the gripper further to grasp it securely, and lift it out of the drawer."
+  }
+}
\ No newline at end of file
diff --git a/pineapple_out_of_drawer/ep023/labels.npz b/pineapple_out_of_drawer/ep023/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..82abe6b6551ae8b0171c433b1c638e261285aafa
--- /dev/null
+++ b/pineapple_out_of_drawer/ep023/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4a3a410c47b0eab6b198bc4f0dae44a66f598e3430bfee6069e33dff91583837
+size 3978
diff --git a/pineapple_out_of_drawer/ep024/example.md b/pineapple_out_of_drawer/ep024/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..21afe9b8670f8e96912b76037e1ab5d3a759e5f1
--- /dev/null
+++ b/pineapple_out_of_drawer/ep024/example.md
@@ -0,0 +1,33 @@
+# t007__ep024__pineapple_out_of_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the pineapple out of the drawer
+- **success:** False
+- **control steps:** 455  (9.0s at 50 Hz)
+- **policy inferences:** 78  (replan_steps=6)
+- **video:** `pineapple_out_of_drawer/ep024/rollout.mp4`
+- **features:** `78` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | moving arm behind drawer |
+| 3.0–6.0 | 152–303 | 25–50 | hovering behind drawer box |
+| 6.0–9.0 | 304–454 | 50–75 | lowering gripper behind drawer |
+
+## Failure
+
+- **yes**
+- **onset type:** operator_intervention
+- **onset:** 2.910891089108911 s (00:02.9)
+- **frame / control_step:** 147 (policy_step=24, chunk_index=3)
+- **coarse onset:** 4.0 s (refined=True)
+- **mode:** stuck_or_no_progress
+- **reason:** The robot hovered and moved aimlessly behind the drawer structure instead of reaching into the drawer to pick up the pineapple.
+  Summary: hovering behind drawer; failing to reach pineapple
+- **recovery:** Reposition the gripper directly over the open drawer compartment, lower to grasp the pineapple securely, and lift it out of the drawer.
+  Summary: reposition over drawer; grasp and remove pineapple
+- **prevention:** Direct the arm trajectory toward the open drawer section containing the pineapple rather than lowering behind the drawer box.
+  Summary: target pineapple in drawer; avoid back of drawer box
+- **justification:** By around 4.0 seconds, the robot arm reaches a hovering position behind the drawer box and fails to make any useful progress toward reaching or grabbing the pineapple inside the drawer.
diff --git a/pineapple_out_of_drawer/ep024/labels.json b/pineapple_out_of_drawer/ep024/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..7fe244c738b048f623094b1178036546b4a6aef0
--- /dev/null
+++ b/pineapple_out_of_drawer/ep024/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t007__ep024__pineapple_out_of_drawer",
+  "scene_variant": "",
+  "task_id": 7,
+  "task_file": "",
+  "instruction": "Take the pineapple out of the drawer",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:25:45+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00990099009901,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "moving arm behind drawer",
+      "description": "moving arm behind drawer"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00990099009901,
+      "t_end_sec": 6.01980198019802,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "hovering behind drawer box",
+      "description": "hovering behind drawer box"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01980198019802,
+      "t_end_sec": 9.009900990099009,
+      "control_step_start": 304,
+      "control_step_end": 454,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "lowering gripper behind drawer",
+      "description": "lowering gripper behind drawer"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "stuck_or_no_progress",
+    "vlm_failure_onset_type": "operator_intervention",
+    "vlm_failure_onset_seconds": 2.910891089108911,
+    "vlm_failure_onset_timestamp": "00:02.9",
+    "vlm_failure_onset_frame": 147,
+    "vlm_failure_onset_step": 147,
+    "vlm_coarse_failure_onset_seconds": 4.0,
+    "vlm_failure_window_start_seconds": 3.5,
+    "vlm_failure_window_end_seconds": 4.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot hovered and moved aimlessly behind the drawer structure instead of reaching into the drawer to pick up the pineapple.",
+    "vlm_failure_reason_summary": "hovering behind drawer; failing to reach pineapple",
+    "vlm_recovery_action": "Reposition the gripper directly over the open drawer compartment, lower to grasp the pineapple securely, and lift it out of the drawer.",
+    "vlm_recovery_action_summary": "reposition over drawer; grasp and remove pineapple",
+    "vlm_prevention_action": "Direct the arm trajectory toward the open drawer section containing the pineapple rather than lowering behind the drawer box.",
+    "vlm_prevention_action_summary": "target pineapple in drawer; avoid back of drawer box",
+    "vlm_justification": "By around 4.0 seconds, the robot arm reaches a hovering position behind the drawer box and fails to make any useful progress toward reaching or grabbing the pineapple inside the drawer.",
+    "vlm_temporal_justification": "At 01:37 (second 97), the robotic arm begins moving from its home position. Instead of trajectory planning towards the pineapple inside the open drawer, it moves downwards behind the drawer structure, initiating the aimless hovering behavior.",
+    "vlm_video_fps": 50.5,
+    "vlm_video_num_frames": 455,
+    "vlm_video_duration_seconds": 8.990099009900991,
+    "vlm_usage": {
+      "input_tokens": 21637,
+      "output_tokens": 358,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"stuck_or_no_progress\",\n \"onset_type\": \"operator_intervention\",\n \"failure_onset_seconds\": 4.0,\n \"failure_window_start_seconds\": 3.5,\n \"failure_window_end_seconds\": 4.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot hovered and moved aimlessly behind the drawer structure instead of reaching into the drawer to pick up the pineapple.\",\n \"failure_reason_summary\": \"hovering behind drawer; failing to reach pineapple\",\n \"recovery_action\": \"Reposition the gripper directly over the open drawer compartment, lower to grasp the pineapple securely, and lift it out of the drawer.\",\n \"recovery_action_summary\": \"reposition over drawer; grasp and remove pineapple\",\n \"prevention_action\": \"Direct the arm trajectory toward the open drawer section containing the pineapple rather than lowering behind the drawer box.\",\n \"prevention_action_summary\": \"target pineapple in drawer; avoid back of drawer box\",\n \"justification\": \"By around 4.0 seconds, the robot arm reaches a hovering position behind the drawer box and fails to make any useful progress toward reaching or grabbing the pineapple inside the drawer.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 147,
+    "failure_sim_step": 147,
+    "failure_policy_step": 24,
+    "failure_chunk_index": 3,
+    "first_post_failure_policy_step": 25,
+    "failure_type": "stuck_or_no_progress",
+    "correction_action": "Reposition the gripper directly over the open drawer compartment, lower to grasp the pineapple securely, and lift it out of the drawer."
+  }
+}
\ No newline at end of file
diff --git a/pineapple_out_of_drawer/ep024/labels.npz b/pineapple_out_of_drawer/ep024/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..e7b31516102ee9c9e0b05dcaef86f9b8fc521c89
--- /dev/null
+++ b/pineapple_out_of_drawer/ep024/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:26e6c7f826bb369b38c82003708d8aa0f1ff6eca5cdb3916edb44beddb339f93
+size 3750
diff --git a/pineapple_out_of_drawer/ep025/example.md b/pineapple_out_of_drawer/ep025/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..47b06d5b0788873ce74220a175f6afd324bcf800
--- /dev/null
+++ b/pineapple_out_of_drawer/ep025/example.md
@@ -0,0 +1,34 @@
+# t007__ep025__pineapple_out_of_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the pineapple out of the drawer
+- **success:** False
+- **control steps:** 540  (10.5s at 51 Hz)
+- **policy inferences:** 93  (replan_steps=6)
+- **video:** `pineapple_out_of_drawer/ep025/rollout.mp4`
+- **features:** `93` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–153 | 0–25 | hovering above container |
+| 3.0–6.0 | 154–307 | 25–51 | lowering gripper towards pineapple |
+| 6.0–9.0 | 308–461 | 51–76 | closing gripper around pineapple |
+| 9.0–10.5 | 462–539 | 77–89 | remaining stationary inside container |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 8.557504873294347 s (00:08.6)
+- **frame / control_step:** 439 (policy_step=73, chunk_index=1)
+- **coarse onset:** 8.5 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot end-effector lowered into the container and closed its fingers around the pineapple, but failed to establish a secure grasp to lift it out.
+  Summary: failed to grasp pineapple; object remains in container
+- **recovery:** Open the gripper, reposition the end-effector directly over the main body of the pineapple, and execute a firm grasp before lifting it out.
+  Summary: re-align gripper; firmly grasp pineapple; lift out of box
+- **prevention:** Ensure the gripper descends deep enough around the pineapple's center of mass before closing the fingers to guarantee a successful grasp.
+  Summary: descend deeper over object; secure firm grip
+- **justification:** At approximately 8.5 seconds, the robot attempts to grasp the pineapple in the container but fails to grab it securely, leaving the pineapple untouched as the rollout ends shortly after.
diff --git a/pineapple_out_of_drawer/ep025/labels.json b/pineapple_out_of_drawer/ep025/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..5f8c0047e36511eb1dc7974a1b834bada43d1817
--- /dev/null
+++ b/pineapple_out_of_drawer/ep025/labels.json
@@ -0,0 +1,109 @@
+{
+  "rollout_id": "t007__ep025__pineapple_out_of_drawer",
+  "scene_variant": "",
+  "task_id": 7,
+  "task_file": "",
+  "instruction": "Take the pineapple out of the drawer",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:26:18+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0019493177387915,
+      "control_step_start": 0,
+      "control_step_end": 153,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "hovering above container",
+      "description": "hovering above container"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0019493177387915,
+      "t_end_sec": 6.003898635477583,
+      "control_step_start": 154,
+      "control_step_end": 307,
+      "policy_step_start": 25,
+      "policy_step_end": 51,
+      "phrase": "lowering gripper towards pineapple",
+      "description": "lowering gripper towards pineapple"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.003898635477583,
+      "t_end_sec": 9.005847953216374,
+      "control_step_start": 308,
+      "control_step_end": 461,
+      "policy_step_start": 51,
+      "policy_step_end": 76,
+      "phrase": "closing gripper around pineapple",
+      "description": "closing gripper around pineapple"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.005847953216374,
+      "t_end_sec": 10.526315789473685,
+      "control_step_start": 462,
+      "control_step_end": 539,
+      "policy_step_start": 77,
+      "policy_step_end": 89,
+      "phrase": "remaining stationary inside container",
+      "description": "remaining stationary inside container"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 8.557504873294347,
+    "vlm_failure_onset_timestamp": "00:08.6",
+    "vlm_failure_onset_frame": 439,
+    "vlm_failure_onset_step": 439,
+    "vlm_coarse_failure_onset_seconds": 8.5,
+    "vlm_failure_window_start_seconds": 8.0,
+    "vlm_failure_window_end_seconds": 9.0,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot end-effector lowered into the container and closed its fingers around the pineapple, but failed to establish a secure grasp to lift it out.",
+    "vlm_failure_reason_summary": "failed to grasp pineapple; object remains in container",
+    "vlm_recovery_action": "Open the gripper, reposition the end-effector directly over the main body of the pineapple, and execute a firm grasp before lifting it out.",
+    "vlm_recovery_action_summary": "re-align gripper; firmly grasp pineapple; lift out of box",
+    "vlm_prevention_action": "Ensure the gripper descends deep enough around the pineapple's center of mass before closing the fingers to guarantee a successful grasp.",
+    "vlm_prevention_action_summary": "descend deeper over object; secure firm grip",
+    "vlm_justification": "At approximately 8.5 seconds, the robot attempts to grasp the pineapple in the container but fails to grab it securely, leaving the pineapple untouched as the rollout ends shortly after.",
+    "vlm_temporal_justification": "At timestamp 02:37, the robot gripper finishes closing its fingers completely together without securing the pineapple, making the grasp failure visibly apparent as it begins to lift away empty.",
+    "vlm_video_fps": 51.3,
+    "vlm_video_num_frames": 540,
+    "vlm_video_duration_seconds": 10.50682261208577,
+    "vlm_usage": {
+      "input_tokens": 18810,
+      "output_tokens": 358,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 8.5,\n \"failure_window_start_seconds\": 8.0,\n \"failure_window_end_seconds\": 9.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot end-effector lowered into the container and closed its fingers around the pineapple, but failed to establish a secure grasp to lift it out.\",\n \"failure_reason_summary\": \"failed to grasp pineapple; object remains in container\",\n \"recovery_action\": \"Open the gripper, reposition the end-effector directly over the main body of the pineapple, and execute a firm grasp before lifting it out.\",\n \"recovery_action_summary\": \"re-align gripper; firmly grasp pineapple; lift out of box\",\n \"prevention_action\": \"Ensure the gripper descends deep enough around the pineapple's center of mass before closing the fingers to guarantee a successful grasp.\",\n \"prevention_action_summary\": \"descend deeper over object; secure firm grip\",\n \"justification\": \"At approximately 8.5 seconds, the robot attempts to grasp the pineapple in the container but fails to grab it securely, leaving the pineapple untouched as the rollout ends shortly after.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 439,
+    "failure_sim_step": 439,
+    "failure_policy_step": 73,
+    "failure_chunk_index": 1,
+    "first_post_failure_policy_step": 74,
+    "failure_type": "grasp_failure",
+    "correction_action": "Open the gripper, reposition the end-effector directly over the main body of the pineapple, and execute a firm grasp before lifting it out."
+  }
+}
\ No newline at end of file
diff --git a/pineapple_out_of_drawer/ep025/labels.npz b/pineapple_out_of_drawer/ep025/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..b12e44f011812a8d3390c34e9cee99fd1fdc03c1
--- /dev/null
+++ b/pineapple_out_of_drawer/ep025/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:18e77ed9e64826639b2f04472ef071782730e908ecc8e56f62dca457ecde9d76
+size 3891
diff --git a/pineapple_out_of_drawer/ep026/example.md b/pineapple_out_of_drawer/ep026/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..fa76f08ac78edbe83b013e467a116b5e5577f64e
--- /dev/null
+++ b/pineapple_out_of_drawer/ep026/example.md
@@ -0,0 +1,23 @@
+# t007__ep026__pineapple_out_of_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the pineapple out of the drawer
+- **success:** True
+- **control steps:** 715  (14.1s at 51 Hz)
+- **policy inferences:** 123  (replan_steps=6)
+- **video:** `pineapple_out_of_drawer/ep026/rollout.mp4`
+- **features:** `123` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | approaching pineapple in drawer |
+| 3.0–6.0 | 152–303 | 25–50 | grasping pineapple in drawer |
+| 6.0–9.0 | 304–455 | 50–75 | lifting pineapple from drawer |
+| 9.0–12.0 | 456–607 | 76–101 | moving pineapple over table |
+| 12.0–14.1 | 608–714 | 101–119 | releasing pineapple onto table |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/pineapple_out_of_drawer/ep026/labels.json b/pineapple_out_of_drawer/ep026/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..9d2c056c63941ce1ac9e0e54a400bcc7e39f7b21
--- /dev/null
+++ b/pineapple_out_of_drawer/ep026/labels.json
@@ -0,0 +1,90 @@
+{
+  "rollout_id": "t007__ep026__pineapple_out_of_drawer",
+  "scene_variant": "",
+  "task_id": 7,
+  "task_file": "",
+  "instruction": "Take the pineapple out of the drawer",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:26:28+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9921259842519685,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "approaching pineapple in drawer",
+      "description": "approaching pineapple in drawer"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9921259842519685,
+      "t_end_sec": 5.984251968503937,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "grasping pineapple in drawer",
+      "description": "grasping pineapple in drawer"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.984251968503937,
+      "t_end_sec": 8.976377952755906,
+      "control_step_start": 304,
+      "control_step_end": 455,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "lifting pineapple from drawer",
+      "description": "lifting pineapple from drawer"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.976377952755906,
+      "t_end_sec": 11.968503937007874,
+      "control_step_start": 456,
+      "control_step_end": 607,
+      "policy_step_start": 76,
+      "policy_step_end": 101,
+      "phrase": "moving pineapple over table",
+      "description": "moving pineapple over table"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 11.968503937007874,
+      "t_end_sec": 14.0748031496063,
+      "control_step_start": 608,
+      "control_step_end": 714,
+      "policy_step_start": 101,
+      "policy_step_end": 119,
+      "phrase": "releasing pineapple onto table",
+      "description": "releasing pineapple onto table"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/pineapple_out_of_drawer/ep026/labels.npz b/pineapple_out_of_drawer/ep026/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..f75fb1b25b314101b4d1024db232e2d4e07389c7
--- /dev/null
+++ b/pineapple_out_of_drawer/ep026/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c79fe2b38722bf54b18024b602e7434f744d8ec97755832f937e8d8ef5cbb2ce
+size 3134
diff --git a/pineapple_out_of_drawer/ep027/example.md b/pineapple_out_of_drawer/ep027/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..f9b6316f0a46917458383ea55c02526bb4e68f1d
--- /dev/null
+++ b/pineapple_out_of_drawer/ep027/example.md
@@ -0,0 +1,35 @@
+# t007__ep027__pineapple_out_of_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the pineapple out of the drawer
+- **success:** False
+- **control steps:** 685  (13.6s at 50 Hz)
+- **policy inferences:** 118  (replan_steps=6)
+- **video:** `pineapple_out_of_drawer/ep027/rollout.mp4`
+- **features:** `118` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | descending toward open drawer |
+| 3.0–6.0 | 151–301 | 25–50 | lowering gripper; attempting grasp |
+| 6.0–9.0 | 302–452 | 50–75 | hovering motionless inside drawer |
+| 9.0–12.0 | 453–603 | 75–100 | remaining stationary over drawer |
+| 12.0–13.6 | 604–684 | 100–114 | idle near pineapple |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 3.5912698412698414 s (00:03.6)
+- **frame / control_step:** 181 (policy_step=30, chunk_index=1)
+- **coarse onset:** 6.0 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot lowered its end-effector into the open drawer to grasp the pineapple but failed to align properly, closing its fingers around empty space without securing the object.
+  Summary: failed grasp on pineapple; misaligned gripper inside drawer
+- **recovery:** Open the gripper, adjust the position to center over the pineapple inside the open drawer, and close the gripper to securely grasp and lift it out.
+  Summary: open gripper; re-align over pineapple; grasp and lift
+- **prevention:** Center the end-effector carefully over the pineapple's center of mass before closing the gripper to establish a stable hold.
+  Summary: center gripper over object; grasp firmly
+- **justification:** At around 6.0 seconds, the robot attempts to close its gripper around the pineapple in the open drawer, but misses the grasp and remains stuck hovering over the object for the rest of the rollout.
diff --git a/pineapple_out_of_drawer/ep027/labels.json b/pineapple_out_of_drawer/ep027/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..9b12220533ee9dc2791bf636e176849b96c0d104
--- /dev/null
+++ b/pineapple_out_of_drawer/ep027/labels.json
@@ -0,0 +1,120 @@
+{
+  "rollout_id": "t007__ep027__pineapple_out_of_drawer",
+  "scene_variant": "",
+  "task_id": 7,
+  "task_file": "",
+  "instruction": "Take the pineapple out of the drawer",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:27:03+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.996031746031746,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "descending toward open drawer",
+      "description": "descending toward open drawer"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.996031746031746,
+      "t_end_sec": 5.992063492063492,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "lowering gripper; attempting grasp",
+      "description": "lowering gripper; attempting grasp"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.992063492063492,
+      "t_end_sec": 8.988095238095239,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "hovering motionless inside drawer",
+      "description": "hovering motionless inside drawer"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.988095238095239,
+      "t_end_sec": 11.984126984126984,
+      "control_step_start": 453,
+      "control_step_end": 603,
+      "policy_step_start": 75,
+      "policy_step_end": 100,
+      "phrase": "remaining stationary over drawer",
+      "description": "remaining stationary over drawer"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 11.984126984126984,
+      "t_end_sec": 13.591269841269842,
+      "control_step_start": 604,
+      "control_step_end": 684,
+      "policy_step_start": 100,
+      "policy_step_end": 114,
+      "phrase": "idle near pineapple",
+      "description": "idle near pineapple"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 3.5912698412698414,
+    "vlm_failure_onset_timestamp": "00:03.6",
+    "vlm_failure_onset_frame": 181,
+    "vlm_failure_onset_step": 181,
+    "vlm_coarse_failure_onset_seconds": 6.0,
+    "vlm_failure_window_start_seconds": 5.5,
+    "vlm_failure_window_end_seconds": 7.0,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot lowered its end-effector into the open drawer to grasp the pineapple but failed to align properly, closing its fingers around empty space without securing the object.",
+    "vlm_failure_reason_summary": "failed grasp on pineapple; misaligned gripper inside drawer",
+    "vlm_recovery_action": "Open the gripper, adjust the position to center over the pineapple inside the open drawer, and close the gripper to securely grasp and lift it out.",
+    "vlm_recovery_action_summary": "open gripper; re-align over pineapple; grasp and lift",
+    "vlm_prevention_action": "Center the end-effector carefully over the pineapple's center of mass before closing the gripper to establish a stable hold.",
+    "vlm_prevention_action_summary": "center gripper over object; grasp firmly",
+    "vlm_justification": "At around 6.0 seconds, the robot attempts to close its gripper around the pineapple in the open drawer, but misses the grasp and remains stuck hovering over the object for the rest of the rollout.",
+    "vlm_temporal_justification": "At second 30 (00:30 in the clip), the robot finishes lowering into the drawer and begins closing its gripper fingers while misaligned with the pineapple, closing on empty space.",
+    "vlm_video_fps": 50.4,
+    "vlm_video_num_frames": 685,
+    "vlm_video_duration_seconds": 13.571428571428571,
+    "vlm_usage": {
+      "input_tokens": 21837,
+      "output_tokens": 367,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 6.0,\n \"failure_window_start_seconds\": 5.5,\n \"failure_window_end_seconds\": 7.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot lowered its end-effector into the open drawer to grasp the pineapple but failed to align properly, closing its fingers around empty space without securing the object.\",\n \"failure_reason_summary\": \"failed grasp on pineapple; misaligned gripper inside drawer\",\n \"recovery_action\": \"Open the gripper, adjust the position to center over the pineapple inside the open drawer, and close the gripper to securely grasp and lift it out.\",\n \"recovery_action_summary\": \"open gripper; re-align over pineapple; grasp and lift\",\n \"prevention_action\": \"Center the end-effector carefully over the pineapple's center of mass before closing the gripper to establish a stable hold.\",\n \"prevention_action_summary\": \"center gripper over object; grasp firmly\",\n \"justification\": \"At around 6.0 seconds, the robot attempts to close its gripper around the pineapple in the open drawer, but misses the grasp and remains stuck hovering over the object for the rest of the rollout.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 181,
+    "failure_sim_step": 181,
+    "failure_policy_step": 30,
+    "failure_chunk_index": 1,
+    "first_post_failure_policy_step": 31,
+    "failure_type": "grasp_failure",
+    "correction_action": "Open the gripper, adjust the position to center over the pineapple inside the open drawer, and close the gripper to securely grasp and lift it out."
+  }
+}
\ No newline at end of file
diff --git a/pineapple_out_of_drawer/ep027/labels.npz b/pineapple_out_of_drawer/ep027/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..ed2b5648b4d2bec3eedcf23319bbb504024a5b89
--- /dev/null
+++ b/pineapple_out_of_drawer/ep027/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c4fc072e8bb319ac056ccec3b90d51e8fbff1c5403766400edc5ecadcb315e4a
+size 3936
diff --git a/pineapple_out_of_drawer/ep028/example.md b/pineapple_out_of_drawer/ep028/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..8da350dd3b4a1b7656843f672be786d000c4d7b8
--- /dev/null
+++ b/pineapple_out_of_drawer/ep028/example.md
@@ -0,0 +1,35 @@
+# t007__ep028__pineapple_out_of_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the pineapple out of the drawer
+- **success:** False
+- **control steps:** 620  (12.3s at 50 Hz)
+- **policy inferences:** 107  (replan_steps=6)
+- **video:** `pineapple_out_of_drawer/ep028/rollout.mp4`
+- **features:** `107` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | lowering arm toward open drawer |
+| 3.0–6.0 | 151–301 | 25–50 | approaching pineapple inside drawer |
+| 6.0–9.0 | 302–452 | 50–75 | attempting grasp on pineapple |
+| 9.0–12.0 | 453–603 | 75–100 | lifting empty gripper |
+| 12.0–12.3 | 604–619 | 100–103 | hovering above drawer |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 7.003968253968254 s (00:07.0)
+- **frame / control_step:** 353 (policy_step=58, chunk_index=5)
+- **coarse onset:** 7.0 s (refined=False)
+- **mode:** grasp_failure
+- **reason:** The robot descended into the drawer and closed its end-effector over the pineapple, but failed to secure a firm grip, causing the pineapple to slip out as the arm lifted away empty.
+  Summary: failed grasp on pineapple; slipping from gripper
+- **recovery:** Lower the end-effector back into the drawer, align the gripper fingers around the body of the pineapple, close firmly, and lift it out of the drawer.
+  Summary: re-align gripper; grasp pineapple securely; lift out
+- **prevention:** Position the gripper deeper and more centrally around the pineapple before closing the fingers to ensure a secure hold.
+  Summary: center gripper on pineapple; grasp deeper
+- **justification:** Around 7.0 seconds into the rollout, the robot closes its gripper on the pineapple but fails to establish a secure grip. As the arm retracts upwards, the pineapple remains sitting in the open drawer.
diff --git a/pineapple_out_of_drawer/ep028/labels.json b/pineapple_out_of_drawer/ep028/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..d5ea780b35f1dc61a4d569dcaec77d8d2959b266
--- /dev/null
+++ b/pineapple_out_of_drawer/ep028/labels.json
@@ -0,0 +1,120 @@
+{
+  "rollout_id": "t007__ep028__pineapple_out_of_drawer",
+  "scene_variant": "",
+  "task_id": 7,
+  "task_file": "",
+  "instruction": "Take the pineapple out of the drawer",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:27:36+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.996031746031746,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "lowering arm toward open drawer",
+      "description": "lowering arm toward open drawer"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.996031746031746,
+      "t_end_sec": 5.992063492063492,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "approaching pineapple inside drawer",
+      "description": "approaching pineapple inside drawer"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.992063492063492,
+      "t_end_sec": 8.988095238095239,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "attempting grasp on pineapple",
+      "description": "attempting grasp on pineapple"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.988095238095239,
+      "t_end_sec": 11.984126984126984,
+      "control_step_start": 453,
+      "control_step_end": 603,
+      "policy_step_start": 75,
+      "policy_step_end": 100,
+      "phrase": "lifting empty gripper",
+      "description": "lifting empty gripper"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 11.984126984126984,
+      "t_end_sec": 12.301587301587302,
+      "control_step_start": 604,
+      "control_step_end": 619,
+      "policy_step_start": 100,
+      "policy_step_end": 103,
+      "phrase": "hovering above drawer",
+      "description": "hovering above drawer"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 7.003968253968254,
+    "vlm_failure_onset_timestamp": "00:07.0",
+    "vlm_failure_onset_frame": 353,
+    "vlm_failure_onset_step": 353,
+    "vlm_coarse_failure_onset_seconds": 7.0,
+    "vlm_failure_window_start_seconds": 6.0,
+    "vlm_failure_window_end_seconds": 8.0,
+    "vlm_temporal_refined": false,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot descended into the drawer and closed its end-effector over the pineapple, but failed to secure a firm grip, causing the pineapple to slip out as the arm lifted away empty.",
+    "vlm_failure_reason_summary": "failed grasp on pineapple; slipping from gripper",
+    "vlm_recovery_action": "Lower the end-effector back into the drawer, align the gripper fingers around the body of the pineapple, close firmly, and lift it out of the drawer.",
+    "vlm_recovery_action_summary": "re-align gripper; grasp pineapple securely; lift out",
+    "vlm_prevention_action": "Position the gripper deeper and more centrally around the pineapple before closing the fingers to ensure a secure hold.",
+    "vlm_prevention_action_summary": "center gripper on pineapple; grasp deeper",
+    "vlm_justification": "Around 7.0 seconds into the rollout, the robot closes its gripper on the pineapple but fails to establish a secure grip. As the arm retracts upwards, the pineapple remains sitting in the open drawer.",
+    "vlm_temporal_justification": "At 00:34, the robot arm descends into the drawer and closes its end-effector around the pineapple, but fails to establish a secure grasp.",
+    "vlm_video_fps": 50.4,
+    "vlm_video_num_frames": 620,
+    "vlm_video_duration_seconds": 12.281746031746032,
+    "vlm_usage": {
+      "input_tokens": 21716,
+      "output_tokens": 365,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 7.0,\n \"failure_window_start_seconds\": 6.0,\n \"failure_window_end_seconds\": 8.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot descended into the drawer and closed its end-effector over the pineapple, but failed to secure a firm grip, causing the pineapple to slip out as the arm lifted away empty.\",\n \"failure_reason_summary\": \"failed grasp on pineapple; slipping from gripper\",\n \"recovery_action\": \"Lower the end-effector back into the drawer, align the gripper fingers around the body of the pineapple, close firmly, and lift it out of the drawer.\",\n \"recovery_action_summary\": \"re-align gripper; grasp pineapple securely; lift out\",\n \"prevention_action\": \"Position the gripper deeper and more centrally around the pineapple before closing the fingers to ensure a secure hold.\",\n \"prevention_action_summary\": \"center gripper on pineapple; grasp deeper\",\n \"justification\": \"Around 7.0 seconds into the rollout, the robot closes its gripper on the pineapple but fails to establish a secure grip. As the arm retracts upwards, the pineapple remains sitting in the open drawer.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 353,
+    "failure_sim_step": 353,
+    "failure_policy_step": 58,
+    "failure_chunk_index": 5,
+    "first_post_failure_policy_step": 59,
+    "failure_type": "grasp_failure",
+    "correction_action": "Lower the end-effector back into the drawer, align the gripper fingers around the body of the pineapple, close firmly, and lift it out of the drawer."
+  }
+}
\ No newline at end of file
diff --git a/pineapple_out_of_drawer/ep028/labels.npz b/pineapple_out_of_drawer/ep028/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..ebabf921b11a52b279f309b62fa82207801724b2
--- /dev/null
+++ b/pineapple_out_of_drawer/ep028/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:61f120cda83001c554c2de69834317faa05940e0aa59c2f5d6e526677f25c289
+size 3889
diff --git a/pineapple_out_of_drawer/ep029/example.md b/pineapple_out_of_drawer/ep029/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..89fcdb5f3bcdfd4ceead642773a2c708aae3d237
--- /dev/null
+++ b/pineapple_out_of_drawer/ep029/example.md
@@ -0,0 +1,36 @@
+# t007__ep029__pineapple_out_of_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the pineapple out of the drawer
+- **success:** False
+- **control steps:** 857  (16.9s at 51 Hz)
+- **policy inferences:** 147  (replan_steps=6)
+- **video:** `pineapple_out_of_drawer/ep029/rollout.mp4`
+- **features:** `147` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching toward pineapple in drawer |
+| 3.0–6.0 | 152–303 | 25–50 | attempting grasp on pineapple |
+| 6.0–9.0 | 304–455 | 50–75 | slipping off pineapple; hovering above |
+| 9.0–12.0 | 456–607 | 76–101 | hovering over pineapple in drawer |
+| 12.0–15.0 | 608–759 | 101–126 | remaining stationary above drawer |
+| 15.0–16.9 | 760–856 | 126–142 | idle above open drawer |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 4.8023715415019765 s (00:04.8)
+- **frame / control_step:** 243 (policy_step=40, chunk_index=3)
+- **coarse onset:** 5.0 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot end-effector lowered into the drawer and closed its fingers around the upper tip of the pineapple, failing to establish a secure grasp and slipping off when trying to lift it.
+  Summary: failed grasp on pineapple; gripper slipped off object
+- **recovery:** Re-align the gripper directly over the center of the pineapple, lower further into the drawer, close the fingers firmly around the object, and lift it clear of the drawer.
+  Summary: realign gripper over pineapple; secure deep grasp; lift out of drawer
+- **prevention:** Lower the end-effector further into the open drawer to envelop the widest portion of the pineapple before closing the gripper fingers.
+  Summary: align gripper lower; grasp pineapple firmly; lift out
+- **justification:** At 5.0 seconds, the robot gripper closes around the pineapple without making adequate contact to secure it. As a result, the gripper slips off the pineapple, leaving it inside the open drawer while the arm fails to make progress for the rest of the rollout.
diff --git a/pineapple_out_of_drawer/ep029/labels.json b/pineapple_out_of_drawer/ep029/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..2a6ba6d28ba51d2d607418e32ebd0de45765987a
--- /dev/null
+++ b/pineapple_out_of_drawer/ep029/labels.json
@@ -0,0 +1,131 @@
+{
+  "rollout_id": "t007__ep029__pineapple_out_of_drawer",
+  "scene_variant": "",
+  "task_id": 7,
+  "task_file": "",
+  "instruction": "Take the pineapple out of the drawer",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:28:11+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0039525691699605,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching toward pineapple in drawer",
+      "description": "reaching toward pineapple in drawer"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0039525691699605,
+      "t_end_sec": 6.007905138339921,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "attempting grasp on pineapple",
+      "description": "attempting grasp on pineapple"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.007905138339921,
+      "t_end_sec": 9.011857707509881,
+      "control_step_start": 304,
+      "control_step_end": 455,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "slipping off pineapple; hovering above",
+      "description": "slipping off pineapple; hovering above"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.011857707509881,
+      "t_end_sec": 12.015810276679842,
+      "control_step_start": 456,
+      "control_step_end": 607,
+      "policy_step_start": 76,
+      "policy_step_end": 101,
+      "phrase": "hovering over pineapple in drawer",
+      "description": "hovering over pineapple in drawer"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 12.015810276679842,
+      "t_end_sec": 15.019762845849803,
+      "control_step_start": 608,
+      "control_step_end": 759,
+      "policy_step_start": 101,
+      "policy_step_end": 126,
+      "phrase": "remaining stationary above drawer",
+      "description": "remaining stationary above drawer"
+    },
+    {
+      "segment_index": 5,
+      "t_start_sec": 15.019762845849803,
+      "t_end_sec": 16.936758893280633,
+      "control_step_start": 760,
+      "control_step_end": 856,
+      "policy_step_start": 126,
+      "policy_step_end": 142,
+      "phrase": "idle above open drawer",
+      "description": "idle above open drawer"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 4.8023715415019765,
+    "vlm_failure_onset_timestamp": "00:04.8",
+    "vlm_failure_onset_frame": 243,
+    "vlm_failure_onset_step": 243,
+    "vlm_coarse_failure_onset_seconds": 5.0,
+    "vlm_failure_window_start_seconds": 4.5,
+    "vlm_failure_window_end_seconds": 5.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot end-effector lowered into the drawer and closed its fingers around the upper tip of the pineapple, failing to establish a secure grasp and slipping off when trying to lift it.",
+    "vlm_failure_reason_summary": "failed grasp on pineapple; gripper slipped off object",
+    "vlm_recovery_action": "Re-align the gripper directly over the center of the pineapple, lower further into the drawer, close the fingers firmly around the object, and lift it clear of the drawer.",
+    "vlm_recovery_action_summary": "realign gripper over pineapple; secure deep grasp; lift out of drawer",
+    "vlm_prevention_action": "Lower the end-effector further into the open drawer to envelop the widest portion of the pineapple before closing the gripper fingers.",
+    "vlm_prevention_action_summary": "align gripper lower; grasp pineapple firmly; lift out",
+    "vlm_justification": "At 5.0 seconds, the robot gripper closes around the pineapple without making adequate contact to secure it. As a result, the gripper slips off the pineapple, leaving it inside the open drawer while the arm fails to make progress for the rest of the rollout.",
+    "vlm_temporal_justification": "At timestamp 02:22 (142 seconds), the robot end-effector stops descending and begins closing its fingers around the top tip of the pineapple rather than reaching deep enough around its body, initiating the insecure grasp.",
+    "vlm_video_fps": 50.6,
+    "vlm_video_num_frames": 857,
+    "vlm_video_duration_seconds": 16.91699604743083,
+    "vlm_usage": {
+      "input_tokens": 22157,
+      "output_tokens": 401,
+      "cached_tokens": 1157
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 5.0,\n \"failure_window_start_seconds\": 4.5,\n \"failure_window_end_seconds\": 5.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot end-effector lowered into the drawer and closed its fingers around the upper tip of the pineapple, failing to establish a secure grasp and slipping off when trying to lift it.\",\n \"failure_reason_summary\": \"failed grasp on pineapple; gripper slipped off object\",\n \"recovery_action\": \"Re-align the gripper directly over the center of the pineapple, lower further into the drawer, close the fingers firmly around the object, and lift it clear of the drawer.\",\n \"recovery_action_summary\": \"realign gripper over pineapple; secure deep grasp; lift out of drawer\",\n \"prevention_action\": \"Lower the end-effector further into the open drawer to envelop the widest portion of the pineapple before closing the gripper fingers.\",\n \"prevention_action_summary\": \"align gripper lower; grasp pineapple firmly; lift out\",\n \"justification\": \"At 5.0 seconds, the robot gripper closes around the pineapple without making adequate contact to secure it. As a result, the gripper slips off the pineapple, leaving it inside the open drawer while the arm fails to make progress for the rest of the rollout.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 243,
+    "failure_sim_step": 243,
+    "failure_policy_step": 40,
+    "failure_chunk_index": 3,
+    "first_post_failure_policy_step": 41,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-align the gripper directly over the center of the pineapple, lower further into the drawer, close the fingers firmly around the object, and lift it clear of the drawer."
+  }
+}
\ No newline at end of file
diff --git a/pineapple_out_of_drawer/ep029/labels.npz b/pineapple_out_of_drawer/ep029/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..e09b0bb11539df1a0d2836c3e9c1c4c0329eee37
--- /dev/null
+++ b/pineapple_out_of_drawer/ep029/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:255a89fee8fdfc2f7491a255f8c0615c19926fd85089865b524edfddea9f9374
+size 4013
diff --git a/pineapple_out_of_drawer/ep030/example.md b/pineapple_out_of_drawer/ep030/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..209c96d1bae2544fa63ace3c49c9b723910d0092
--- /dev/null
+++ b/pineapple_out_of_drawer/ep030/example.md
@@ -0,0 +1,25 @@
+# t007__ep030__pineapple_out_of_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the pineapple out of the drawer
+- **success:** True
+- **control steps:** 941  (18.7s at 50 Hz)
+- **policy inferences:** 161  (replan_steps=6)
+- **video:** `pineapple_out_of_drawer/ep030/rollout.mp4`
+- **features:** `161` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | positioning gripper above drawer |
+| 3.0–6.0 | 151–301 | 25–50 | lowering gripper toward pineapple |
+| 6.0–9.0 | 302–452 | 50–75 | grasping pineapple toy |
+| 9.0–12.0 | 453–603 | 75–100 | lifting pineapple out of drawer |
+| 12.0–15.0 | 604–754 | 100–125 | moving pineapple over table |
+| 15.0–18.0 | 755–905 | 125–150 | placing pineapple on table |
+| 18.0–18.7 | 906–940 | 151–156 | releasing pineapple |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/pineapple_out_of_drawer/ep030/labels.json b/pineapple_out_of_drawer/ep030/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..18c57d42e15959cdbc9c5dfdbdb5340a98e329e3
--- /dev/null
+++ b/pineapple_out_of_drawer/ep030/labels.json
@@ -0,0 +1,112 @@
+{
+  "rollout_id": "t007__ep030__pineapple_out_of_drawer",
+  "scene_variant": "",
+  "task_id": 7,
+  "task_file": "",
+  "instruction": "Take the pineapple out of the drawer",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:28:24+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.001988071570577,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "positioning gripper above drawer",
+      "description": "positioning gripper above drawer"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.001988071570577,
+      "t_end_sec": 6.003976143141154,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "lowering gripper toward pineapple",
+      "description": "lowering gripper toward pineapple"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.003976143141154,
+      "t_end_sec": 9.00596421471173,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "grasping pineapple toy",
+      "description": "grasping pineapple toy"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.00596421471173,
+      "t_end_sec": 12.007952286282308,
+      "control_step_start": 453,
+      "control_step_end": 603,
+      "policy_step_start": 75,
+      "policy_step_end": 100,
+      "phrase": "lifting pineapple out of drawer",
+      "description": "lifting pineapple out of drawer"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 12.007952286282308,
+      "t_end_sec": 15.009940357852884,
+      "control_step_start": 604,
+      "control_step_end": 754,
+      "policy_step_start": 100,
+      "policy_step_end": 125,
+      "phrase": "moving pineapple over table",
+      "description": "moving pineapple over table"
+    },
+    {
+      "segment_index": 5,
+      "t_start_sec": 15.009940357852884,
+      "t_end_sec": 18.01192842942346,
+      "control_step_start": 755,
+      "control_step_end": 905,
+      "policy_step_start": 125,
+      "policy_step_end": 150,
+      "phrase": "placing pineapple on table",
+      "description": "placing pineapple on table"
+    },
+    {
+      "segment_index": 6,
+      "t_start_sec": 18.01192842942346,
+      "t_end_sec": 18.707753479125248,
+      "control_step_start": 906,
+      "control_step_end": 940,
+      "policy_step_start": 151,
+      "policy_step_end": 156,
+      "phrase": "releasing pineapple",
+      "description": "releasing pineapple"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/pineapple_out_of_drawer/ep030/labels.npz b/pineapple_out_of_drawer/ep030/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..105cf2eba5b981b9e4c00bf6f59421a1aec597cb
--- /dev/null
+++ b/pineapple_out_of_drawer/ep030/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f627f7b000a2365c35a8751e3b4639b3a503cd40c5169189a3a96ff264e5da8b
+size 3222
diff --git a/pineapple_out_of_drawer/ep031/example.md b/pineapple_out_of_drawer/ep031/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..9e964b0e989824a45753d5c724ee9a0e43f1e92b
--- /dev/null
+++ b/pineapple_out_of_drawer/ep031/example.md
@@ -0,0 +1,27 @@
+# t007__ep031__pineapple_out_of_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the pineapple out of the drawer
+- **success:** True
+- **control steps:** 1226  (24.3s at 50 Hz)
+- **policy inferences:** 211  (replan_steps=6)
+- **video:** `pineapple_out_of_drawer/ep031/rollout.mp4`
+- **features:** `211` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | idle |
+| 3.0–6.0 | 151–301 | 25–50 | moving toward pineapple |
+| 6.0–9.0 | 302–452 | 50–75 | approaching pineapple in drawer |
+| 9.0–12.0 | 453–603 | 75–100 | positioning gripper over pineapple |
+| 12.0–15.0 | 604–754 | 100–125 | grasping pineapple |
+| 15.0–18.0 | 755–905 | 125–150 | lifting pineapple out of drawer |
+| 18.0–21.0 | 906–1056 | 151–176 | placing pineapple on table |
+| 21.0–24.0 | 1057–1207 | 176–201 | releasing pineapple and retracting |
+| 24.0–24.3 | 1208–1225 | 201–204 | idle |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/pineapple_out_of_drawer/ep031/labels.json b/pineapple_out_of_drawer/ep031/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..11f49e5a1b22472bd1b7d7d1d5136a4b3126bc16
--- /dev/null
+++ b/pineapple_out_of_drawer/ep031/labels.json
@@ -0,0 +1,134 @@
+{
+  "rollout_id": "t007__ep031__pineapple_out_of_drawer",
+  "scene_variant": "",
+  "task_id": 7,
+  "task_file": "",
+  "instruction": "Take the pineapple out of the drawer",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:28:34+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.996031746031746,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "idle",
+      "description": "idle"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.996031746031746,
+      "t_end_sec": 5.992063492063492,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "moving toward pineapple",
+      "description": "moving toward pineapple"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.992063492063492,
+      "t_end_sec": 8.988095238095239,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "approaching pineapple in drawer",
+      "description": "approaching pineapple in drawer"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.988095238095239,
+      "t_end_sec": 11.984126984126984,
+      "control_step_start": 453,
+      "control_step_end": 603,
+      "policy_step_start": 75,
+      "policy_step_end": 100,
+      "phrase": "positioning gripper over pineapple",
+      "description": "positioning gripper over pineapple"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 11.984126984126984,
+      "t_end_sec": 14.980158730158731,
+      "control_step_start": 604,
+      "control_step_end": 754,
+      "policy_step_start": 100,
+      "policy_step_end": 125,
+      "phrase": "grasping pineapple",
+      "description": "grasping pineapple"
+    },
+    {
+      "segment_index": 5,
+      "t_start_sec": 14.980158730158731,
+      "t_end_sec": 17.976190476190478,
+      "control_step_start": 755,
+      "control_step_end": 905,
+      "policy_step_start": 125,
+      "policy_step_end": 150,
+      "phrase": "lifting pineapple out of drawer",
+      "description": "lifting pineapple out of drawer"
+    },
+    {
+      "segment_index": 6,
+      "t_start_sec": 17.976190476190478,
+      "t_end_sec": 20.97222222222222,
+      "control_step_start": 906,
+      "control_step_end": 1056,
+      "policy_step_start": 151,
+      "policy_step_end": 176,
+      "phrase": "placing pineapple on table",
+      "description": "placing pineapple on table"
+    },
+    {
+      "segment_index": 7,
+      "t_start_sec": 20.97222222222222,
+      "t_end_sec": 23.96825396825397,
+      "control_step_start": 1057,
+      "control_step_end": 1207,
+      "policy_step_start": 176,
+      "policy_step_end": 201,
+      "phrase": "releasing pineapple and retracting",
+      "description": "releasing pineapple and retracting"
+    },
+    {
+      "segment_index": 8,
+      "t_start_sec": 23.96825396825397,
+      "t_end_sec": 24.325396825396826,
+      "control_step_start": 1208,
+      "control_step_end": 1225,
+      "policy_step_start": 201,
+      "policy_step_end": 204,
+      "phrase": "idle",
+      "description": "idle"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/pineapple_out_of_drawer/ep031/labels.npz b/pineapple_out_of_drawer/ep031/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..1d3af7859076fe6461d6ea491a34952ccf0f47e5
--- /dev/null
+++ b/pineapple_out_of_drawer/ep031/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ad2b6145b732107b1b831c3bd0bc29e47c0879b802a563816b3df54e1bd21cb0
+size 3286
diff --git a/pineapple_out_of_drawer/ep032/example.md b/pineapple_out_of_drawer/ep032/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..9770348c5e3234e17a39fa8e8dd4fe2ae3dfe7e9
--- /dev/null
+++ b/pineapple_out_of_drawer/ep032/example.md
@@ -0,0 +1,24 @@
+# t007__ep032__pineapple_out_of_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the pineapple out of the drawer
+- **success:** True
+- **control steps:** 898  (17.8s at 50 Hz)
+- **policy inferences:** 155  (replan_steps=6)
+- **video:** `pineapple_out_of_drawer/ep032/rollout.mp4`
+- **features:** `155` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | idle / no progress |
+| 3.0–6.0 | 152–303 | 25–50 | reaching toward pineapple |
+| 6.0–9.0 | 304–455 | 50–75 | grasping pineapple in drawer |
+| 9.0–12.0 | 456–607 | 76–101 | lifting pineapple from drawer |
+| 12.0–15.0 | 608–759 | 101–126 | moving and releasing pineapple |
+| 15.0–17.8 | 760–897 | 126–149 | retracting robot arm |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/pineapple_out_of_drawer/ep032/labels.json b/pineapple_out_of_drawer/ep032/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..723fde4d32713832b86cc32cce44b46c8d1dff79
--- /dev/null
+++ b/pineapple_out_of_drawer/ep032/labels.json
@@ -0,0 +1,101 @@
+{
+  "rollout_id": "t007__ep032__pineapple_out_of_drawer",
+  "scene_variant": "",
+  "task_id": 7,
+  "task_file": "",
+  "instruction": "Take the pineapple out of the drawer",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:28:44+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00990099009901,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "idle / no progress",
+      "description": "idle / no progress"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00990099009901,
+      "t_end_sec": 6.01980198019802,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "reaching toward pineapple",
+      "description": "reaching toward pineapple"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01980198019802,
+      "t_end_sec": 9.029702970297029,
+      "control_step_start": 304,
+      "control_step_end": 455,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "grasping pineapple in drawer",
+      "description": "grasping pineapple in drawer"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.029702970297029,
+      "t_end_sec": 12.03960396039604,
+      "control_step_start": 456,
+      "control_step_end": 607,
+      "policy_step_start": 76,
+      "policy_step_end": 101,
+      "phrase": "lifting pineapple from drawer",
+      "description": "lifting pineapple from drawer"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 12.03960396039604,
+      "t_end_sec": 15.049504950495049,
+      "control_step_start": 608,
+      "control_step_end": 759,
+      "policy_step_start": 101,
+      "policy_step_end": 126,
+      "phrase": "moving and releasing pineapple",
+      "description": "moving and releasing pineapple"
+    },
+    {
+      "segment_index": 5,
+      "t_start_sec": 15.049504950495049,
+      "t_end_sec": 17.782178217821784,
+      "control_step_start": 760,
+      "control_step_end": 897,
+      "policy_step_start": 126,
+      "policy_step_end": 149,
+      "phrase": "retracting robot arm",
+      "description": "retracting robot arm"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/pineapple_out_of_drawer/ep032/labels.npz b/pineapple_out_of_drawer/ep032/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..36bd43583fb7e7779e336761ba96c5a9d4f14446
--- /dev/null
+++ b/pineapple_out_of_drawer/ep032/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:033d05404769735cd38d3c662f62ad55d814e01a0b5d467cb85db169ca594745
+size 3210
diff --git a/pineapple_out_of_drawer/ep033/example.md b/pineapple_out_of_drawer/ep033/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..5695d7e0fccdef2a654e65838927d76dec785580
--- /dev/null
+++ b/pineapple_out_of_drawer/ep033/example.md
@@ -0,0 +1,36 @@
+# t007__ep033__pineapple_out_of_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the pineapple out of the drawer
+- **success:** False
+- **control steps:** 903  (17.9s at 50 Hz)
+- **policy inferences:** 156  (replan_steps=6)
+- **video:** `pineapple_out_of_drawer/ep033/rollout.mp4`
+- **features:** `156` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | lowering arm toward drawer |
+| 3.0–6.0 | 151–301 | 25–50 | approaching pineapple in drawer |
+| 6.0–9.0 | 302–452 | 50–75 | closing gripper on pineapple |
+| 9.0–12.0 | 453–603 | 75–100 | idle inside open drawer |
+| 12.0–15.0 | 604–754 | 100–125 | idle inside open drawer |
+| 15.0–17.9 | 755–902 | 125–150 | idle inside open drawer |
+
+## Failure
+
+- **yes**
+- **onset type:** operator_intervention
+- **onset:** 7.182539682539683 s (00:07.2)
+- **frame / control_step:** 362 (policy_step=60, chunk_index=2)
+- **coarse onset:** 8.5 s (refined=True)
+- **mode:** stuck_or_no_progress
+- **reason:** The robot arm descended into the drawer and closed its end effector on the pineapple, but then froze completely and ceased moving for the remainder of the trajectory without lifting the object out.
+  Summary: robot arm froze after gripping pineapple; no attempt to lift
+- **recovery:** Resume arm control, grasp the pineapple securely, and pull it upward and out of the open drawer.
+  Summary: lift gripper upward; remove pineapple from drawer
+- **prevention:** Ensure the policy outputs a continuous sequence of upward motion control commands immediately following the grasp phase.
+  Summary: command post-grasp lifting motion; avoid policy freeze
+- **justification:** None
diff --git a/pineapple_out_of_drawer/ep033/labels.json b/pineapple_out_of_drawer/ep033/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..b46badb184b2a278a57ae92ad0513c0b76824c97
--- /dev/null
+++ b/pineapple_out_of_drawer/ep033/labels.json
@@ -0,0 +1,131 @@
+{
+  "rollout_id": "t007__ep033__pineapple_out_of_drawer",
+  "scene_variant": "",
+  "task_id": 7,
+  "task_file": "",
+  "instruction": "Take the pineapple out of the drawer",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:29:17+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.996031746031746,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "lowering arm toward drawer",
+      "description": "lowering arm toward drawer"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.996031746031746,
+      "t_end_sec": 5.992063492063492,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "approaching pineapple in drawer",
+      "description": "approaching pineapple in drawer"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.992063492063492,
+      "t_end_sec": 8.988095238095239,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "closing gripper on pineapple",
+      "description": "closing gripper on pineapple"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.988095238095239,
+      "t_end_sec": 11.984126984126984,
+      "control_step_start": 453,
+      "control_step_end": 603,
+      "policy_step_start": 75,
+      "policy_step_end": 100,
+      "phrase": "idle inside open drawer",
+      "description": "idle inside open drawer"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 11.984126984126984,
+      "t_end_sec": 14.980158730158731,
+      "control_step_start": 604,
+      "control_step_end": 754,
+      "policy_step_start": 100,
+      "policy_step_end": 125,
+      "phrase": "idle inside open drawer",
+      "description": "idle inside open drawer"
+    },
+    {
+      "segment_index": 5,
+      "t_start_sec": 14.980158730158731,
+      "t_end_sec": 17.916666666666668,
+      "control_step_start": 755,
+      "control_step_end": 902,
+      "policy_step_start": 125,
+      "policy_step_end": 150,
+      "phrase": "idle inside open drawer",
+      "description": "idle inside open drawer"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "stuck_or_no_progress",
+    "vlm_failure_onset_type": "operator_intervention",
+    "vlm_failure_onset_seconds": 7.182539682539683,
+    "vlm_failure_onset_timestamp": "00:07.2",
+    "vlm_failure_onset_frame": 362,
+    "vlm_failure_onset_step": 362,
+    "vlm_coarse_failure_onset_seconds": 8.5,
+    "vlm_failure_window_start_seconds": 8.5,
+    "vlm_failure_window_end_seconds": 17.9,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot arm descended into the drawer and closed its end effector on the pineapple, but then froze completely and ceased moving for the remainder of the trajectory without lifting the object out.",
+    "vlm_failure_reason_summary": "robot arm froze after gripping pineapple; no attempt to lift",
+    "vlm_recovery_action": "Resume arm control, grasp the pineapple securely, and pull it upward and out of the open drawer.",
+    "vlm_recovery_action_summary": "lift gripper upward; remove pineapple from drawer",
+    "vlm_prevention_action": "Ensure the policy outputs a continuous sequence of upward motion control commands immediately following the grasp phase.",
+    "vlm_prevention_action_summary": "command post-grasp lifting motion; avoid policy freeze",
+    "vlm_justification": null,
+    "vlm_temporal_justification": "The robot finishes closing its gripper around the pineapple at second 01:24 (frame 84/85 of clip) and then freezes completely without attempting to lift the object out for the rest of the clip.",
+    "vlm_video_fps": 50.4,
+    "vlm_video_num_frames": 903,
+    "vlm_video_duration_seconds": 17.8968253968254,
+    "vlm_usage": {
+      "input_tokens": 22096,
+      "output_tokens": 318,
+      "cached_tokens": 1225
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"stuck_or_no_progress\",\n \"onset_type\": \"operator_intervention\",\n \"failure_onset_seconds\": 8.5,\n \"failure_window_start_seconds\": 8.5,\n \"failure_window_end_seconds\": 17.9,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot arm descended into the drawer and closed its end effector on the pineapple, but then froze completely and ceased moving for the remainder of the trajectory without lifting the object out.\",\n \"failure_reason_summary\": \"robot arm froze after gripping pineapple; no attempt to lift\",\n \"recovery_action\": \"Resume arm control, grasp the pineapple securely, and pull it upward and out of the open drawer.\",\n \"recovery_action_summary\": \"lift gripper upward; remove pineapple from drawer\",\n \"prevention_action\": \"Ensure the policy outputs a continuous sequence of upward motion control commands immediately following the grasp phase.\",\n \"prevention_action_summary\": \"command post-grasp lifting motion; avoid policy freeze\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 362,
+    "failure_sim_step": 362,
+    "failure_policy_step": 60,
+    "failure_chunk_index": 2,
+    "first_post_failure_policy_step": 61,
+    "failure_type": "stuck_or_no_progress",
+    "correction_action": "Resume arm control, grasp the pineapple securely, and pull it upward and out of the open drawer."
+  }
+}
\ No newline at end of file
diff --git a/pineapple_out_of_drawer/ep033/labels.npz b/pineapple_out_of_drawer/ep033/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..ae92f178ab015ec69e6da2f148da2ce06ffef285
--- /dev/null
+++ b/pineapple_out_of_drawer/ep033/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4f650e108fee27e80e877ab9b355bb6b17daa2866186cb6d0d66aa4213a6fe6b
+size 3909
diff --git a/pineapple_out_of_drawer/ep034/example.md b/pineapple_out_of_drawer/ep034/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..9988a1992cfc43ad76a821f950e62c28039f7a54
--- /dev/null
+++ b/pineapple_out_of_drawer/ep034/example.md
@@ -0,0 +1,37 @@
+# t007__ep034__pineapple_out_of_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the pineapple out of the drawer
+- **success:** False
+- **control steps:** 944  (18.7s at 50 Hz)
+- **policy inferences:** 163  (replan_steps=6)
+- **video:** `pineapple_out_of_drawer/ep034/rollout.mp4`
+- **features:** `163` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | moving toward open drawer |
+| 3.0–6.0 | 151–301 | 25–50 | descending and closing gripper |
+| 6.0–9.0 | 302–452 | 50–75 | idle above pineapple |
+| 9.0–12.0 | 453–603 | 75–100 | idle above pineapple |
+| 12.0–15.0 | 604–754 | 100–125 | idle / no progress |
+| 15.0–18.0 | 755–905 | 125–150 | idle / no progress |
+| 18.0–18.7 | 906–943 | 151–157 | idle until rollout end |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 5.992063492063492 s (00:06.0)
+- **frame / control_step:** 302 (policy_step=50, chunk_index=2)
+- **coarse onset:** 6.0 s (refined=False)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the pineapple by closing its fingers around its top tip rather than lowering further into the drawer, failing to secure the object and stalling.
+  Summary: failed grasp attempt; gripper closed too high
+- **recovery:** Reopen the gripper, descend further into the drawer to envelope the pineapple's body, and close the gripper securely before lifting it out.
+  Summary: reopen gripper; lower arm deeper; regrasp pineapple
+- **prevention:** Ensure the gripper descends sufficiently deep into the drawer around the center mass of the pineapple before closing the fingers.
+  Summary: descend lower into drawer; grasp pineapple body securely
+- **justification:** At around 6.0 seconds into the rollout, the robot closes its gripper on the top tip of the pineapple without positioning low enough around its body. It fails to establish a hold and remains frozen in that position for the remainder of the episode.
diff --git a/pineapple_out_of_drawer/ep034/labels.json b/pineapple_out_of_drawer/ep034/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..9997de23f8976fe4f27f0a9e246520ba43f7917c
--- /dev/null
+++ b/pineapple_out_of_drawer/ep034/labels.json
@@ -0,0 +1,142 @@
+{
+  "rollout_id": "t007__ep034__pineapple_out_of_drawer",
+  "scene_variant": "",
+  "task_id": 7,
+  "task_file": "",
+  "instruction": "Take the pineapple out of the drawer",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:29:55+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.996031746031746,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "moving toward open drawer",
+      "description": "moving toward open drawer"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.996031746031746,
+      "t_end_sec": 5.992063492063492,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "descending and closing gripper",
+      "description": "descending and closing gripper"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.992063492063492,
+      "t_end_sec": 8.988095238095239,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "idle above pineapple",
+      "description": "idle above pineapple"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.988095238095239,
+      "t_end_sec": 11.984126984126984,
+      "control_step_start": 453,
+      "control_step_end": 603,
+      "policy_step_start": 75,
+      "policy_step_end": 100,
+      "phrase": "idle above pineapple",
+      "description": "idle above pineapple"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 11.984126984126984,
+      "t_end_sec": 14.980158730158731,
+      "control_step_start": 604,
+      "control_step_end": 754,
+      "policy_step_start": 100,
+      "policy_step_end": 125,
+      "phrase": "idle / no progress",
+      "description": "idle / no progress"
+    },
+    {
+      "segment_index": 5,
+      "t_start_sec": 14.980158730158731,
+      "t_end_sec": 17.976190476190478,
+      "control_step_start": 755,
+      "control_step_end": 905,
+      "policy_step_start": 125,
+      "policy_step_end": 150,
+      "phrase": "idle / no progress",
+      "description": "idle / no progress"
+    },
+    {
+      "segment_index": 6,
+      "t_start_sec": 17.976190476190478,
+      "t_end_sec": 18.73015873015873,
+      "control_step_start": 906,
+      "control_step_end": 943,
+      "policy_step_start": 151,
+      "policy_step_end": 157,
+      "phrase": "idle until rollout end",
+      "description": "idle until rollout end"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 5.992063492063492,
+    "vlm_failure_onset_timestamp": "00:06.0",
+    "vlm_failure_onset_frame": 302,
+    "vlm_failure_onset_step": 302,
+    "vlm_coarse_failure_onset_seconds": 6.0,
+    "vlm_failure_window_start_seconds": 5.5,
+    "vlm_failure_window_end_seconds": 7.0,
+    "vlm_temporal_refined": false,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the pineapple by closing its fingers around its top tip rather than lowering further into the drawer, failing to secure the object and stalling.",
+    "vlm_failure_reason_summary": "failed grasp attempt; gripper closed too high",
+    "vlm_recovery_action": "Reopen the gripper, descend further into the drawer to envelope the pineapple's body, and close the gripper securely before lifting it out.",
+    "vlm_recovery_action_summary": "reopen gripper; lower arm deeper; regrasp pineapple",
+    "vlm_prevention_action": "Ensure the gripper descends sufficiently deep into the drawer around the center mass of the pineapple before closing the fingers.",
+    "vlm_prevention_action_summary": "descend lower into drawer; grasp pineapple body securely",
+    "vlm_justification": "At around 6.0 seconds into the rollout, the robot closes its gripper on the top tip of the pineapple without positioning low enough around its body. It fails to establish a hold and remains frozen in that position for the remainder of the episode.",
+    "vlm_temporal_justification": "At second 02:08, the robot gripper stops descending and begins closing its fingers around the top tip of the pineapple rather than lowering further into the drawer, initiating the bad grasp attempt that leads to stall.",
+    "vlm_video_fps": 50.4,
+    "vlm_video_num_frames": 944,
+    "vlm_video_duration_seconds": 18.710317460317462,
+    "vlm_usage": {
+      "input_tokens": 22152,
+      "output_tokens": 378,
+      "cached_tokens": 1293
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 6.0,\n \"failure_window_start_seconds\": 5.5,\n \"failure_window_end_seconds\": 7.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the pineapple by closing its fingers around its top tip rather than lowering further into the drawer, failing to secure the object and stalling.\",\n \"failure_reason_summary\": \"failed grasp attempt; gripper closed too high\",\n \"recovery_action\": \"Reopen the gripper, descend further into the drawer to envelope the pineapple's body, and close the gripper securely before lifting it out.\",\n \"recovery_action_summary\": \"reopen gripper; lower arm deeper; regrasp pineapple\",\n \"prevention_action\": \"Ensure the gripper descends sufficiently deep into the drawer around the center mass of the pineapple before closing the fingers.\",\n \"prevention_action_summary\": \"descend lower into drawer; grasp pineapple body securely\",\n \"justification\": \"At around 6.0 seconds into the rollout, the robot closes its gripper on the top tip of the pineapple without positioning low enough around its body. It fails to establish a hold and remains frozen in that position for the remainder of the episode.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 302,
+    "failure_sim_step": 302,
+    "failure_policy_step": 50,
+    "failure_chunk_index": 2,
+    "first_post_failure_policy_step": 51,
+    "failure_type": "grasp_failure",
+    "correction_action": "Reopen the gripper, descend further into the drawer to envelope the pineapple's body, and close the gripper securely before lifting it out."
+  }
+}
\ No newline at end of file
diff --git a/pineapple_out_of_drawer/ep034/labels.npz b/pineapple_out_of_drawer/ep034/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..f01a0fc57198e9fa8ca0eaedb6313ede12982e64
--- /dev/null
+++ b/pineapple_out_of_drawer/ep034/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5839d44302f483f53916d35e5aa679b834f0544504e07a70e6842113dc522608
+size 3953
diff --git a/pineapple_out_of_drawer/ep035/example.md b/pineapple_out_of_drawer/ep035/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..0bf659e786b1cc9daa54f65c694693ffd37aafee
--- /dev/null
+++ b/pineapple_out_of_drawer/ep035/example.md
@@ -0,0 +1,38 @@
+# t007__ep035__pineapple_out_of_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the pineapple out of the drawer
+- **success:** False
+- **control steps:** 1184  (23.3s at 51 Hz)
+- **policy inferences:** 205  (replan_steps=6)
+- **video:** `pineapple_out_of_drawer/ep035/rollout.mp4`
+- **features:** `205` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–152 | 0–25 | moving arm toward drawer |
+| 3.0–6.0 | 153–305 | 25–50 | approaching pineapple in drawer |
+| 6.0–9.0 | 306–458 | 51–76 | closing gripper on pineapple |
+| 9.0–12.0 | 459–611 | 76–101 | stuck hovering over object |
+| 12.0–15.0 | 612–764 | 102–127 | idle near pineapple |
+| 15.0–18.0 | 765–917 | 127–152 | idle / no progress |
+| 18.0–21.0 | 918–1070 | 153–178 | idle / no progress |
+| 21.0–23.3 | 1071–1183 | 178–197 | idle / no progress |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 7.93713163064833 s (00:07.9)
+- **frame / control_step:** 404 (policy_step=67, chunk_index=2)
+- **coarse onset:** 8.0 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the top of the pineapple inside the drawer, but failed to secure a firm grip around the main body and could not lift it out.
+  Summary: failed grasp on pineapple; unable to lift object
+- **recovery:** Open the gripper, reposition the end-effector lower around the main body of the pineapple, grasp firmly, and pull it out of the drawer.
+  Summary: open gripper; reposition around object; grasp and lift out
+- **prevention:** Align the gripper deeper around the center of the pineapple rather than pinching the top stem before closing fingers to ensure a secure grip.
+  Summary: align gripper around pineapple body; grasp securely
+- **justification:** At around 8.0 seconds, the robot's gripper closes on the top stem of the pineapple instead of surrounding its main body. The robot fails to secure a grip and ceases making forward progress for the rest of the execution.
diff --git a/pineapple_out_of_drawer/ep035/labels.json b/pineapple_out_of_drawer/ep035/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..f99108d883f2918bda592e10bea4d08b21fe9259
--- /dev/null
+++ b/pineapple_out_of_drawer/ep035/labels.json
@@ -0,0 +1,153 @@
+{
+  "rollout_id": "t007__ep035__pineapple_out_of_drawer",
+  "scene_variant": "",
+  "task_id": 7,
+  "task_file": "",
+  "instruction": "Take the pineapple out of the drawer",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:30:32+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.005893909626719,
+      "control_step_start": 0,
+      "control_step_end": 152,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "moving arm toward drawer",
+      "description": "moving arm toward drawer"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.005893909626719,
+      "t_end_sec": 6.011787819253438,
+      "control_step_start": 153,
+      "control_step_end": 305,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "approaching pineapple in drawer",
+      "description": "approaching pineapple in drawer"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.011787819253438,
+      "t_end_sec": 9.017681728880158,
+      "control_step_start": 306,
+      "control_step_end": 458,
+      "policy_step_start": 51,
+      "policy_step_end": 76,
+      "phrase": "closing gripper on pineapple",
+      "description": "closing gripper on pineapple"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.017681728880158,
+      "t_end_sec": 12.023575638506877,
+      "control_step_start": 459,
+      "control_step_end": 611,
+      "policy_step_start": 76,
+      "policy_step_end": 101,
+      "phrase": "stuck hovering over object",
+      "description": "stuck hovering over object"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 12.023575638506877,
+      "t_end_sec": 15.029469548133596,
+      "control_step_start": 612,
+      "control_step_end": 764,
+      "policy_step_start": 102,
+      "policy_step_end": 127,
+      "phrase": "idle near pineapple",
+      "description": "idle near pineapple"
+    },
+    {
+      "segment_index": 5,
+      "t_start_sec": 15.029469548133596,
+      "t_end_sec": 18.035363457760315,
+      "control_step_start": 765,
+      "control_step_end": 917,
+      "policy_step_start": 127,
+      "policy_step_end": 152,
+      "phrase": "idle / no progress",
+      "description": "idle / no progress"
+    },
+    {
+      "segment_index": 6,
+      "t_start_sec": 18.035363457760315,
+      "t_end_sec": 21.041257367387033,
+      "control_step_start": 918,
+      "control_step_end": 1070,
+      "policy_step_start": 153,
+      "policy_step_end": 178,
+      "phrase": "idle / no progress",
+      "description": "idle / no progress"
+    },
+    {
+      "segment_index": 7,
+      "t_start_sec": 21.041257367387033,
+      "t_end_sec": 23.26129666011788,
+      "control_step_start": 1071,
+      "control_step_end": 1183,
+      "policy_step_start": 178,
+      "policy_step_end": 197,
+      "phrase": "idle / no progress",
+      "description": "idle / no progress"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 7.93713163064833,
+    "vlm_failure_onset_timestamp": "00:07.9",
+    "vlm_failure_onset_frame": 404,
+    "vlm_failure_onset_step": 404,
+    "vlm_coarse_failure_onset_seconds": 8.0,
+    "vlm_failure_window_start_seconds": 7.5,
+    "vlm_failure_window_end_seconds": 12.0,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the top of the pineapple inside the drawer, but failed to secure a firm grip around the main body and could not lift it out.",
+    "vlm_failure_reason_summary": "failed grasp on pineapple; unable to lift object",
+    "vlm_recovery_action": "Open the gripper, reposition the end-effector lower around the main body of the pineapple, grasp firmly, and pull it out of the drawer.",
+    "vlm_recovery_action_summary": "open gripper; reposition around object; grasp and lift out",
+    "vlm_prevention_action": "Align the gripper deeper around the center of the pineapple rather than pinching the top stem before closing fingers to ensure a secure grip.",
+    "vlm_prevention_action_summary": "align gripper around pineapple body; grasp securely",
+    "vlm_justification": "At around 8.0 seconds, the robot's gripper closes on the top stem of the pineapple instead of surrounding its main body. The robot fails to secure a grip and ceases making forward progress for the rest of the execution.",
+    "vlm_temporal_justification": "At second 150 (02:30 in the clip), the robot gripper finishes closing its fingers, securing only the top leaves of the pineapple rather than reaching down around the main body, leading directly to the grasp failure.",
+    "vlm_video_fps": 50.9,
+    "vlm_video_num_frames": 1184,
+    "vlm_video_duration_seconds": 23.241650294695482,
+    "vlm_usage": {
+      "input_tokens": 22657,
+      "output_tokens": 383,
+      "cached_tokens": 1565
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 8.0,\n \"failure_window_start_seconds\": 7.5,\n \"failure_window_end_seconds\": 12.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the top of the pineapple inside the drawer, but failed to secure a firm grip around the main body and could not lift it out.\",\n \"failure_reason_summary\": \"failed grasp on pineapple; unable to lift object\",\n \"recovery_action\": \"Open the gripper, reposition the end-effector lower around the main body of the pineapple, grasp firmly, and pull it out of the drawer.\",\n \"recovery_action_summary\": \"open gripper; reposition around object; grasp and lift out\",\n \"prevention_action\": \"Align the gripper deeper around the center of the pineapple rather than pinching the top stem before closing fingers to ensure a secure grip.\",\n \"prevention_action_summary\": \"align gripper around pineapple body; grasp securely\",\n \"justification\": \"At around 8.0 seconds, the robot's gripper closes on the top stem of the pineapple instead of surrounding its main body. The robot fails to secure a grip and ceases making forward progress for the rest of the execution.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 404,
+    "failure_sim_step": 404,
+    "failure_policy_step": 67,
+    "failure_chunk_index": 2,
+    "first_post_failure_policy_step": 68,
+    "failure_type": "grasp_failure",
+    "correction_action": "Open the gripper, reposition the end-effector lower around the main body of the pineapple, grasp firmly, and pull it out of the drawer."
+  }
+}
\ No newline at end of file
diff --git a/pineapple_out_of_drawer/ep035/labels.npz b/pineapple_out_of_drawer/ep035/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..cd7fd3ad215d3588a26a5a01060df23c94887863
--- /dev/null
+++ b/pineapple_out_of_drawer/ep035/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0cd99705bb173c7c5395d802a36e0137372d7ebbf6a52dba70bc89f4c0cad732
+size 3992
diff --git a/pineapple_out_of_drawer/ep036/example.md b/pineapple_out_of_drawer/ep036/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..82f14dde9586db9ce5a5d264fec0cd36c230b20d
--- /dev/null
+++ b/pineapple_out_of_drawer/ep036/example.md
@@ -0,0 +1,34 @@
+# t007__ep036__pineapple_out_of_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the pineapple out of the drawer
+- **success:** False
+- **control steps:** 507  (10.0s at 51 Hz)
+- **policy inferences:** 87  (replan_steps=6)
+- **video:** `pineapple_out_of_drawer/ep036/rollout.mp4`
+- **features:** `87` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching toward open drawer |
+| 3.0–6.0 | 152–303 | 25–50 | lowering gripper over pineapple |
+| 6.0–9.0 | 304–455 | 50–75 | attempting grasp on pineapple |
+| 9.0–10.0 | 456–506 | 76–84 | holding position above pineapple |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 7.790927021696252 s (00:07.8)
+- **frame / control_step:** 395 (policy_step=65, chunk_index=5)
+- **coarse onset:** 6.3 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot aligned its gripper over the pineapple stem and closed its fingers, but failed to secure a grasp on the object, leaving it in the drawer.
+  Summary: failed grasp on pineapple; object remaining in drawer
+- **recovery:** Re-open the gripper, lower the end-effector further to envelop the pineapple, close the fingers firmly around it, and pull it out of the drawer.
+  Summary: re-align gripper; secure grasp on pineapple; lift object out
+- **prevention:** Position the gripper lower and more centered over the pineapple before closing the fingers to guarantee a reliable grip.
+  Summary: center gripper over object; achieve firm grip
+- **justification:** Around 6.3 seconds, the robot attempts to close its gripper around the top of the pineapple but fails to establish a firm grip, leaving the pineapple untouched in the drawer for the remainder of the episode.
diff --git a/pineapple_out_of_drawer/ep036/labels.json b/pineapple_out_of_drawer/ep036/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..4cc72562f446dd61eb87eae6bed98accfdd7e5ed
--- /dev/null
+++ b/pineapple_out_of_drawer/ep036/labels.json
@@ -0,0 +1,109 @@
+{
+  "rollout_id": "t007__ep036__pineapple_out_of_drawer",
+  "scene_variant": "",
+  "task_id": 7,
+  "task_file": "",
+  "instruction": "Take the pineapple out of the drawer",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:31:07+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9980276134122286,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching toward open drawer",
+      "description": "reaching toward open drawer"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9980276134122286,
+      "t_end_sec": 5.996055226824457,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "lowering gripper over pineapple",
+      "description": "lowering gripper over pineapple"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.996055226824457,
+      "t_end_sec": 8.994082840236686,
+      "control_step_start": 304,
+      "control_step_end": 455,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "attempting grasp on pineapple",
+      "description": "attempting grasp on pineapple"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.994082840236686,
+      "t_end_sec": 10.0,
+      "control_step_start": 456,
+      "control_step_end": 506,
+      "policy_step_start": 76,
+      "policy_step_end": 84,
+      "phrase": "holding position above pineapple",
+      "description": "holding position above pineapple"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 7.790927021696252,
+    "vlm_failure_onset_timestamp": "00:07.8",
+    "vlm_failure_onset_frame": 395,
+    "vlm_failure_onset_step": 395,
+    "vlm_coarse_failure_onset_seconds": 6.3,
+    "vlm_failure_window_start_seconds": 5.3,
+    "vlm_failure_window_end_seconds": 7.3,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot aligned its gripper over the pineapple stem and closed its fingers, but failed to secure a grasp on the object, leaving it in the drawer.",
+    "vlm_failure_reason_summary": "failed grasp on pineapple; object remaining in drawer",
+    "vlm_recovery_action": "Re-open the gripper, lower the end-effector further to envelop the pineapple, close the fingers firmly around it, and pull it out of the drawer.",
+    "vlm_recovery_action_summary": "re-align gripper; secure grasp on pineapple; lift object out",
+    "vlm_prevention_action": "Position the gripper lower and more centered over the pineapple before closing the fingers to guarantee a reliable grip.",
+    "vlm_prevention_action_summary": "center gripper over object; achieve firm grip",
+    "vlm_justification": "Around 6.3 seconds, the robot attempts to close its gripper around the top of the pineapple but fails to establish a firm grip, leaving the pineapple untouched in the drawer for the remainder of the episode.",
+    "vlm_temporal_justification": "At timestamp 03:48 (clip second 228), the robot's gripper fingers finish closing without securing a hold on the pineapple, leaving the object lying ungrasped in the drawer.",
+    "vlm_video_fps": 50.7,
+    "vlm_video_num_frames": 507,
+    "vlm_video_duration_seconds": 9.980276134122287,
+    "vlm_usage": {
+      "input_tokens": 21707,
+      "output_tokens": 370,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 6.3,\n \"failure_window_start_seconds\": 5.3,\n \"failure_window_end_seconds\": 7.3,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot aligned its gripper over the pineapple stem and closed its fingers, but failed to secure a grasp on the object, leaving it in the drawer.\",\n \"failure_reason_summary\": \"failed grasp on pineapple; object remaining in drawer\",\n \"recovery_action\": \"Re-open the gripper, lower the end-effector further to envelop the pineapple, close the fingers firmly around it, and pull it out of the drawer.\",\n \"recovery_action_summary\": \"re-align gripper; secure grasp on pineapple; lift object out\",\n \"prevention_action\": \"Position the gripper lower and more centered over the pineapple before closing the fingers to guarantee a reliable grip.\",\n \"prevention_action_summary\": \"center gripper over object; achieve firm grip\",\n \"justification\": \"Around 6.3 seconds, the robot attempts to close its gripper around the top of the pineapple but fails to establish a firm grip, leaving the pineapple untouched in the drawer for the remainder of the episode.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 395,
+    "failure_sim_step": 395,
+    "failure_policy_step": 65,
+    "failure_chunk_index": 5,
+    "first_post_failure_policy_step": 66,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-open the gripper, lower the end-effector further to envelop the pineapple, close the fingers firmly around it, and pull it out of the drawer."
+  }
+}
\ No newline at end of file
diff --git a/pineapple_out_of_drawer/ep036/labels.npz b/pineapple_out_of_drawer/ep036/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..2c1bdff30157c37250b0ec7ea45a09916e090b14
--- /dev/null
+++ b/pineapple_out_of_drawer/ep036/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:db3d567240291e3939ee1ce9dc6ca00f9607a9923fa20a92eafb9fb33187f5c1
+size 3854
diff --git a/pineapple_out_of_drawer/ep037/example.md b/pineapple_out_of_drawer/ep037/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..31e96e551031efd45f3ffca78ce83b6fcc6a3912
--- /dev/null
+++ b/pineapple_out_of_drawer/ep037/example.md
@@ -0,0 +1,34 @@
+# t007__ep037__pineapple_out_of_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the pineapple out of the drawer
+- **success:** False
+- **control steps:** 495  (9.8s at 51 Hz)
+- **policy inferences:** 85  (replan_steps=6)
+- **video:** `pineapple_out_of_drawer/ep037/rollout.mp4`
+- **features:** `85` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | lowering gripper toward drawer handle |
+| 3.0–6.0 | 152–303 | 25–50 | hovering near drawer handle |
+| 6.0–9.0 | 304–455 | 50–75 | oscillating near drawer handle |
+| 9.0–9.8 | 456–494 | 76–82 | remaining stationary near handle |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 3.0632411067193677 s (00:03.1)
+- **frame / control_step:** 155 (policy_step=25, chunk_index=5)
+- **coarse onset:** 4.0 s (refined=True)
+- **mode:** wrong_object
+- **reason:** The robot moved toward and targeted the drawer handle instead of reaching inside the open drawer to grasp the pineapple.
+  Summary: targeted drawer handle instead of pineapple
+- **recovery:** Move the gripper away from the drawer handle, align above the pineapple inside the open drawer, and grasp the pineapple.
+  Summary: reposition over pineapple; grasp pineapple
+- **prevention:** Navigate directly to the pineapple inside the drawer rather than attempting to manipulate the drawer handle.
+  Summary: reach directly for pineapple inside drawer
+- **justification:** The drawer is already open at the start of the rollout. At around 4.0 seconds, the robot reaches for the front left edge/handle of the drawer rather than targeting the pineapple resting inside the open container.
diff --git a/pineapple_out_of_drawer/ep037/labels.json b/pineapple_out_of_drawer/ep037/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..6dda5b04dc9de9fa1ef76bcca5dabdd8d8536d31
--- /dev/null
+++ b/pineapple_out_of_drawer/ep037/labels.json
@@ -0,0 +1,109 @@
+{
+  "rollout_id": "t007__ep037__pineapple_out_of_drawer",
+  "scene_variant": "",
+  "task_id": 7,
+  "task_file": "",
+  "instruction": "Take the pineapple out of the drawer",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:31:40+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0039525691699605,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "lowering gripper toward drawer handle",
+      "description": "lowering gripper toward drawer handle"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0039525691699605,
+      "t_end_sec": 6.007905138339921,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "hovering near drawer handle",
+      "description": "hovering near drawer handle"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.007905138339921,
+      "t_end_sec": 9.011857707509881,
+      "control_step_start": 304,
+      "control_step_end": 455,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "oscillating near drawer handle",
+      "description": "oscillating near drawer handle"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.011857707509881,
+      "t_end_sec": 9.782608695652174,
+      "control_step_start": 456,
+      "control_step_end": 494,
+      "policy_step_start": 76,
+      "policy_step_end": 82,
+      "phrase": "remaining stationary near handle",
+      "description": "remaining stationary near handle"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "wrong_object",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 3.0632411067193677,
+    "vlm_failure_onset_timestamp": "00:03.1",
+    "vlm_failure_onset_frame": 155,
+    "vlm_failure_onset_step": 155,
+    "vlm_coarse_failure_onset_seconds": 4.0,
+    "vlm_failure_window_start_seconds": 3.5,
+    "vlm_failure_window_end_seconds": 4.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot moved toward and targeted the drawer handle instead of reaching inside the open drawer to grasp the pineapple.",
+    "vlm_failure_reason_summary": "targeted drawer handle instead of pineapple",
+    "vlm_recovery_action": "Move the gripper away from the drawer handle, align above the pineapple inside the open drawer, and grasp the pineapple.",
+    "vlm_recovery_action_summary": "reposition over pineapple; grasp pineapple",
+    "vlm_prevention_action": "Navigate directly to the pineapple inside the drawer rather than attempting to manipulate the drawer handle.",
+    "vlm_prevention_action_summary": "reach directly for pineapple inside drawer",
+    "vlm_justification": "The drawer is already open at the start of the rollout. At around 4.0 seconds, the robot reaches for the front left edge/handle of the drawer rather than targeting the pineapple resting inside the open container.",
+    "vlm_temporal_justification": "At 01:45 (second 105 of the clip), the robot gripper visibly lowers and aligns directly around the front handle/lip of the drawer rather than entering the drawer cavity to grasp the pineapple.",
+    "vlm_video_fps": 50.6,
+    "vlm_video_num_frames": 495,
+    "vlm_video_duration_seconds": 9.762845849802371,
+    "vlm_usage": {
+      "input_tokens": 21695,
+      "output_tokens": 342,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"wrong_object\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 4.0,\n \"failure_window_start_seconds\": 3.5,\n \"failure_window_end_seconds\": 4.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot moved toward and targeted the drawer handle instead of reaching inside the open drawer to grasp the pineapple.\",\n \"failure_reason_summary\": \"targeted drawer handle instead of pineapple\",\n \"recovery_action\": \"Move the gripper away from the drawer handle, align above the pineapple inside the open drawer, and grasp the pineapple.\",\n \"recovery_action_summary\": \"reposition over pineapple; grasp pineapple\",\n \"prevention_action\": \"Navigate directly to the pineapple inside the drawer rather than attempting to manipulate the drawer handle.\",\n \"prevention_action_summary\": \"reach directly for pineapple inside drawer\",\n \"justification\": \"The drawer is already open at the start of the rollout. At around 4.0 seconds, the robot reaches for the front left edge/handle of the drawer rather than targeting the pineapple resting inside the open container.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 155,
+    "failure_sim_step": 155,
+    "failure_policy_step": 25,
+    "failure_chunk_index": 5,
+    "first_post_failure_policy_step": 26,
+    "failure_type": "wrong_object",
+    "correction_action": "Move the gripper away from the drawer handle, align above the pineapple inside the open drawer, and grasp the pineapple."
+  }
+}
\ No newline at end of file
diff --git a/pineapple_out_of_drawer/ep037/labels.npz b/pineapple_out_of_drawer/ep037/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..7224353ed68a8fde2e0e30020b388e97497cf27a
--- /dev/null
+++ b/pineapple_out_of_drawer/ep037/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8234a262b5e62b23935fc3504126b534f916ddd083499afe5c8587dc0e029073
+size 3703
diff --git a/pineapple_out_of_drawer/ep038/example.md b/pineapple_out_of_drawer/ep038/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..91ad6c2c715660ed5ac5dada4a77fbaf33801698
--- /dev/null
+++ b/pineapple_out_of_drawer/ep038/example.md
@@ -0,0 +1,34 @@
+# t007__ep038__pineapple_out_of_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the pineapple out of the drawer
+- **success:** False
+- **control steps:** 570  (11.4s at 50 Hz)
+- **policy inferences:** 98  (replan_steps=6)
+- **video:** `pineapple_out_of_drawer/ep038/rollout.mp4`
+- **features:** `98` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | reaching toward open drawer |
+| 3.0–6.0 | 151–301 | 25–50 | positioning gripper over pineapple |
+| 6.0–9.0 | 302–452 | 50–75 | closing gripper near pineapple |
+| 9.0–11.4 | 453–569 | 75–94 | hovering idle near drawer |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 6.792828685258963 s (00:06.8)
+- **frame / control_step:** 341 (policy_step=56, chunk_index=5)
+- **coarse onset:** 6.8 s (refined=False)
+- **mode:** grasp_failure
+- **reason:** The robot lowered its end-effector toward the pineapple inside the open drawer, but misaligned its grasp, causing the gripper to close without securing the object.
+  Summary: failed grasp on pineapple; misaligned gripper
+- **recovery:** Re-open the gripper, adjust the end-effector position directly over the middle of the pineapple, and close the gripper to securely grasp it.
+  Summary: re-open gripper; reposition over pineapple; grasp securely
+- **prevention:** Ensure the gripper is properly aligned and positioned directly around the body of the pineapple before attempting to close the fingers.
+  Summary: align gripper over object; close fingers firmly
+- **justification:** At around 6.8 seconds, the robot attempts to grasp the pineapple in the drawer, but the gripper closes without properly surrounding or clamping onto the object, resulting in a grasp failure.
diff --git a/pineapple_out_of_drawer/ep038/labels.json b/pineapple_out_of_drawer/ep038/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..6ce68bb51b0acb7a3919c2576d5e745473f7f846
--- /dev/null
+++ b/pineapple_out_of_drawer/ep038/labels.json
@@ -0,0 +1,109 @@
+{
+  "rollout_id": "t007__ep038__pineapple_out_of_drawer",
+  "scene_variant": "",
+  "task_id": 7,
+  "task_file": "",
+  "instruction": "Take the pineapple out of the drawer",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:32:10+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00796812749004,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching toward open drawer",
+      "description": "reaching toward open drawer"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00796812749004,
+      "t_end_sec": 6.01593625498008,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "positioning gripper over pineapple",
+      "description": "positioning gripper over pineapple"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01593625498008,
+      "t_end_sec": 9.02390438247012,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "closing gripper near pineapple",
+      "description": "closing gripper near pineapple"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.02390438247012,
+      "t_end_sec": 11.354581673306772,
+      "control_step_start": 453,
+      "control_step_end": 569,
+      "policy_step_start": 75,
+      "policy_step_end": 94,
+      "phrase": "hovering idle near drawer",
+      "description": "hovering idle near drawer"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 6.792828685258963,
+    "vlm_failure_onset_timestamp": "00:06.8",
+    "vlm_failure_onset_frame": 341,
+    "vlm_failure_onset_step": 341,
+    "vlm_coarse_failure_onset_seconds": 6.8,
+    "vlm_failure_window_start_seconds": 6.0,
+    "vlm_failure_window_end_seconds": 8.0,
+    "vlm_temporal_refined": false,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot lowered its end-effector toward the pineapple inside the open drawer, but misaligned its grasp, causing the gripper to close without securing the object.",
+    "vlm_failure_reason_summary": "failed grasp on pineapple; misaligned gripper",
+    "vlm_recovery_action": "Re-open the gripper, adjust the end-effector position directly over the middle of the pineapple, and close the gripper to securely grasp it.",
+    "vlm_recovery_action_summary": "re-open gripper; reposition over pineapple; grasp securely",
+    "vlm_prevention_action": "Ensure the gripper is properly aligned and positioned directly around the body of the pineapple before attempting to close the fingers.",
+    "vlm_prevention_action_summary": "align gripper over object; close fingers firmly",
+    "vlm_justification": "At around 6.8 seconds, the robot attempts to grasp the pineapple in the drawer, but the gripper closes without properly surrounding or clamping onto the object, resulting in a grasp failure.",
+    "vlm_temporal_justification": "At second 03:46, the robot arm reaches its lowest point over the drawer and begins closing its gripper fingers while misaligned with the pineapple, leading directly to the failed grasp.",
+    "vlm_video_fps": 50.2,
+    "vlm_video_num_frames": 570,
+    "vlm_video_duration_seconds": 11.334661354581673,
+    "vlm_usage": {
+      "input_tokens": 21647,
+      "output_tokens": 361,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 6.8,\n \"failure_window_start_seconds\": 6.0,\n \"failure_window_end_seconds\": 8.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot lowered its end-effector toward the pineapple inside the open drawer, but misaligned its grasp, causing the gripper to close without securing the object.\",\n \"failure_reason_summary\": \"failed grasp on pineapple; misaligned gripper\",\n \"recovery_action\": \"Re-open the gripper, adjust the end-effector position directly over the middle of the pineapple, and close the gripper to securely grasp it.\",\n \"recovery_action_summary\": \"re-open gripper; reposition over pineapple; grasp securely\",\n \"prevention_action\": \"Ensure the gripper is properly aligned and positioned directly around the body of the pineapple before attempting to close the fingers.\",\n \"prevention_action_summary\": \"align gripper over object; close fingers firmly\",\n \"justification\": \"At around 6.8 seconds, the robot attempts to grasp the pineapple in the drawer, but the gripper closes without properly surrounding or clamping onto the object, resulting in a grasp failure.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 341,
+    "failure_sim_step": 341,
+    "failure_policy_step": 56,
+    "failure_chunk_index": 5,
+    "first_post_failure_policy_step": 57,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-open the gripper, adjust the end-effector position directly over the middle of the pineapple, and close the gripper to securely grasp it."
+  }
+}
\ No newline at end of file
diff --git a/pineapple_out_of_drawer/ep038/labels.npz b/pineapple_out_of_drawer/ep038/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..caf7bec9487397a9fbbb1bfe9e3674a915eff0bd
--- /dev/null
+++ b/pineapple_out_of_drawer/ep038/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4dd1b3eeb03cf71d11cf0e0ea88d576656841dec43f688d34111f0519aaaf786
+size 3876
diff --git a/pineapple_out_of_drawer/ep039/example.md b/pineapple_out_of_drawer/ep039/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..c5852d032a74f418051f70721d6993e069a4405e
--- /dev/null
+++ b/pineapple_out_of_drawer/ep039/example.md
@@ -0,0 +1,34 @@
+# t007__ep039__pineapple_out_of_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the pineapple out of the drawer
+- **success:** False
+- **control steps:** 594  (11.8s at 50 Hz)
+- **policy inferences:** 103  (replan_steps=6)
+- **video:** `pineapple_out_of_drawer/ep039/rollout.mp4`
+- **features:** `103` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | descending toward open drawer |
+| 3.0–6.0 | 152–303 | 25–50 | attempting grasp on pineapple |
+| 6.0–9.0 | 304–455 | 50–75 | lifting empty gripper; repositioning arm |
+| 9.0–11.8 | 456–593 | 76–98 | attempting second grasp in drawer |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 8.0 s (00:08.0)
+- **frame / control_step:** 404 (policy_step=67, chunk_index=2)
+- **coarse onset:** 5.0 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the pineapple inside the drawer, but the gripper failed to establish a secure hold, causing the arm to lift up empty.
+  Summary: failed grasp on pineapple; lifting empty gripper
+- **recovery:** Lower the end-effector back into the open drawer, reposition the fingers around the pineapple, secure a firm grasp, and lift it out.
+  Summary: re-align gripper; grasp pineapple firmly; lift out of drawer
+- **prevention:** Center the end-effector directly over the pineapple and ensure the gripper fully closes around its body before initiating the upward motion.
+  Summary: center gripper over pineapple; secure firm grip before lifting
+- **justification:** Around 5.0 seconds into the rollout, the robot closes its gripper on the pineapple in the drawer but misses a secure grip. At 6.0 seconds, the arm retracts with an empty gripper, leaving the pineapple in the drawer.
diff --git a/pineapple_out_of_drawer/ep039/labels.json b/pineapple_out_of_drawer/ep039/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..d23814ff95df452da6d6b105759f1e97d20247ec
--- /dev/null
+++ b/pineapple_out_of_drawer/ep039/labels.json
@@ -0,0 +1,109 @@
+{
+  "rollout_id": "t007__ep039__pineapple_out_of_drawer",
+  "scene_variant": "",
+  "task_id": 7,
+  "task_file": "",
+  "instruction": "Take the pineapple out of the drawer",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:32:43+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00990099009901,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "descending toward open drawer",
+      "description": "descending toward open drawer"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00990099009901,
+      "t_end_sec": 6.01980198019802,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "attempting grasp on pineapple",
+      "description": "attempting grasp on pineapple"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01980198019802,
+      "t_end_sec": 9.029702970297029,
+      "control_step_start": 304,
+      "control_step_end": 455,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "lifting empty gripper; repositioning arm",
+      "description": "lifting empty gripper; repositioning arm"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.029702970297029,
+      "t_end_sec": 11.762376237623762,
+      "control_step_start": 456,
+      "control_step_end": 593,
+      "policy_step_start": 76,
+      "policy_step_end": 98,
+      "phrase": "attempting second grasp in drawer",
+      "description": "attempting second grasp in drawer"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 8.0,
+    "vlm_failure_onset_timestamp": "00:08.0",
+    "vlm_failure_onset_frame": 404,
+    "vlm_failure_onset_step": 404,
+    "vlm_coarse_failure_onset_seconds": 5.0,
+    "vlm_failure_window_start_seconds": 5.0,
+    "vlm_failure_window_end_seconds": 6.0,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the pineapple inside the drawer, but the gripper failed to establish a secure hold, causing the arm to lift up empty.",
+    "vlm_failure_reason_summary": "failed grasp on pineapple; lifting empty gripper",
+    "vlm_recovery_action": "Lower the end-effector back into the open drawer, reposition the fingers around the pineapple, secure a firm grasp, and lift it out.",
+    "vlm_recovery_action_summary": "re-align gripper; grasp pineapple firmly; lift out of drawer",
+    "vlm_prevention_action": "Center the end-effector directly over the pineapple and ensure the gripper fully closes around its body before initiating the upward motion.",
+    "vlm_prevention_action_summary": "center gripper over pineapple; secure firm grip before lifting",
+    "vlm_justification": "Around 5.0 seconds into the rollout, the robot closes its gripper on the pineapple in the drawer but misses a secure grip. At 6.0 seconds, the arm retracts with an empty gripper, leaving the pineapple in the drawer.",
+    "vlm_temporal_justification": "At second 03:20 (corresponding to original rollout frame 320), the robotic gripper completes its closing motion inside the drawer, but the fingers slip and close fully without securing a hold on the pineapple.",
+    "vlm_video_fps": 50.5,
+    "vlm_video_num_frames": 594,
+    "vlm_video_duration_seconds": 11.742574257425742,
+    "vlm_usage": {
+      "input_tokens": 21834,
+      "output_tokens": 380,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 5.0,\n \"failure_window_start_seconds\": 5.0,\n \"failure_window_end_seconds\": 6.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the pineapple inside the drawer, but the gripper failed to establish a secure hold, causing the arm to lift up empty.\",\n \"failure_reason_summary\": \"failed grasp on pineapple; lifting empty gripper\",\n \"recovery_action\": \"Lower the end-effector back into the open drawer, reposition the fingers around the pineapple, secure a firm grasp, and lift it out.\",\n \"recovery_action_summary\": \"re-align gripper; grasp pineapple firmly; lift out of drawer\",\n \"prevention_action\": \"Center the end-effector directly over the pineapple and ensure the gripper fully closes around its body before initiating the upward motion.\",\n \"prevention_action_summary\": \"center gripper over pineapple; secure firm grip before lifting\",\n \"justification\": \"Around 5.0 seconds into the rollout, the robot closes its gripper on the pineapple in the drawer but misses a secure grip. At 6.0 seconds, the arm retracts with an empty gripper, leaving the pineapple in the drawer.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 404,
+    "failure_sim_step": 404,
+    "failure_policy_step": 67,
+    "failure_chunk_index": 2,
+    "first_post_failure_policy_step": 68,
+    "failure_type": "grasp_failure",
+    "correction_action": "Lower the end-effector back into the open drawer, reposition the fingers around the pineapple, secure a firm grasp, and lift it out."
+  }
+}
\ No newline at end of file
diff --git a/pineapple_out_of_drawer/ep039/labels.npz b/pineapple_out_of_drawer/ep039/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..f8a59f5d6c0ce4ffb54952ffc4abe914fed334af
--- /dev/null
+++ b/pineapple_out_of_drawer/ep039/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:96136c636dbeaadb3d82e1428bbff984994863d5390dc0380366c28daba657ae
+size 3905
diff --git a/pineapple_out_of_drawer/ep040/example.md b/pineapple_out_of_drawer/ep040/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..04bd958397dfbd7669c5ffdec6eebe459ee62508
--- /dev/null
+++ b/pineapple_out_of_drawer/ep040/example.md
@@ -0,0 +1,36 @@
+# t007__ep040__pineapple_out_of_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the pineapple out of the drawer
+- **success:** False
+- **control steps:** 849  (16.8s at 51 Hz)
+- **policy inferences:** 146  (replan_steps=6)
+- **video:** `pineapple_out_of_drawer/ep040/rollout.mp4`
+- **features:** `146` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | idle above workspace |
+| 3.0–6.0 | 152–303 | 25–50 | approaching open drawer |
+| 6.0–9.0 | 304–455 | 50–75 | positioning gripper over pineapple |
+| 9.0–12.0 | 456–607 | 76–101 | attempting grasp on pineapple |
+| 12.0–15.0 | 608–759 | 101–126 | lifting empty gripper |
+| 15.0–16.8 | 760–848 | 126–141 | hovering above open drawer |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 12.490118577075098 s (00:12.5)
+- **frame / control_step:** 632 (policy_step=105, chunk_index=2)
+- **coarse onset:** 10.0 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the pineapple inside the open drawer, but its fingers closed loosely around the top of the object without securing a grip.
+  Summary: failed grasp attempt; pineapple left in drawer
+- **recovery:** Reopen the gripper, descend deeper into the drawer to align the fingers around the body of the pineapple, and close the gripper securely before lifting.
+  Summary: reopen gripper; align fingers; grasp pineapple securely
+- **prevention:** Ensure the gripper descends sufficiently far into the drawer to surround the center of mass of the pineapple before closing the fingers.
+  Summary: descend deeper into drawer; center grasp around pineapple
+- **justification:** At around 10 seconds, the robot gripper closes over the pineapple without establishing a firm hold. The robot then fails to remove the pineapple, leaving it inside the drawer for the rest of the execution.
diff --git a/pineapple_out_of_drawer/ep040/labels.json b/pineapple_out_of_drawer/ep040/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..52eb9251423f4620215898f1b9f416357b7aeddb
--- /dev/null
+++ b/pineapple_out_of_drawer/ep040/labels.json
@@ -0,0 +1,131 @@
+{
+  "rollout_id": "t007__ep040__pineapple_out_of_drawer",
+  "scene_variant": "",
+  "task_id": 7,
+  "task_file": "",
+  "instruction": "Take the pineapple out of the drawer",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:33:17+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0039525691699605,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "idle above workspace",
+      "description": "idle above workspace"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0039525691699605,
+      "t_end_sec": 6.007905138339921,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "approaching open drawer",
+      "description": "approaching open drawer"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.007905138339921,
+      "t_end_sec": 9.011857707509881,
+      "control_step_start": 304,
+      "control_step_end": 455,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "positioning gripper over pineapple",
+      "description": "positioning gripper over pineapple"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.011857707509881,
+      "t_end_sec": 12.015810276679842,
+      "control_step_start": 456,
+      "control_step_end": 607,
+      "policy_step_start": 76,
+      "policy_step_end": 101,
+      "phrase": "attempting grasp on pineapple",
+      "description": "attempting grasp on pineapple"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 12.015810276679842,
+      "t_end_sec": 15.019762845849803,
+      "control_step_start": 608,
+      "control_step_end": 759,
+      "policy_step_start": 101,
+      "policy_step_end": 126,
+      "phrase": "lifting empty gripper",
+      "description": "lifting empty gripper"
+    },
+    {
+      "segment_index": 5,
+      "t_start_sec": 15.019762845849803,
+      "t_end_sec": 16.778656126482215,
+      "control_step_start": 760,
+      "control_step_end": 848,
+      "policy_step_start": 126,
+      "policy_step_end": 141,
+      "phrase": "hovering above open drawer",
+      "description": "hovering above open drawer"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 12.490118577075098,
+    "vlm_failure_onset_timestamp": "00:12.5",
+    "vlm_failure_onset_frame": 632,
+    "vlm_failure_onset_step": 632,
+    "vlm_coarse_failure_onset_seconds": 10.0,
+    "vlm_failure_window_start_seconds": 8.5,
+    "vlm_failure_window_end_seconds": 11.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the pineapple inside the open drawer, but its fingers closed loosely around the top of the object without securing a grip.",
+    "vlm_failure_reason_summary": "failed grasp attempt; pineapple left in drawer",
+    "vlm_recovery_action": "Reopen the gripper, descend deeper into the drawer to align the fingers around the body of the pineapple, and close the gripper securely before lifting.",
+    "vlm_recovery_action_summary": "reopen gripper; align fingers; grasp pineapple securely",
+    "vlm_prevention_action": "Ensure the gripper descends sufficiently far into the drawer to surround the center of mass of the pineapple before closing the fingers.",
+    "vlm_prevention_action_summary": "descend deeper into drawer; center grasp around pineapple",
+    "vlm_justification": "At around 10 seconds, the robot gripper closes over the pineapple without establishing a firm hold. The robot then fails to remove the pineapple, leaving it inside the drawer for the rest of the execution.",
+    "vlm_temporal_justification": "At timestamp 04:38 (second 278 of the clip), the robot's gripper finishes closing loosely around the top of the pineapple inside the open drawer, failing to secure a proper grasp on the object.",
+    "vlm_video_fps": 50.6,
+    "vlm_video_num_frames": 849,
+    "vlm_video_duration_seconds": 16.75889328063241,
+    "vlm_usage": {
+      "input_tokens": 22150,
+      "output_tokens": 371,
+      "cached_tokens": 1157
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 10.0,\n \"failure_window_start_seconds\": 8.5,\n \"failure_window_end_seconds\": 11.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the pineapple inside the open drawer, but its fingers closed loosely around the top of the object without securing a grip.\",\n \"failure_reason_summary\": \"failed grasp attempt; pineapple left in drawer\",\n \"recovery_action\": \"Reopen the gripper, descend deeper into the drawer to align the fingers around the body of the pineapple, and close the gripper securely before lifting.\",\n \"recovery_action_summary\": \"reopen gripper; align fingers; grasp pineapple securely\",\n \"prevention_action\": \"Ensure the gripper descends sufficiently far into the drawer to surround the center of mass of the pineapple before closing the fingers.\",\n \"prevention_action_summary\": \"descend deeper into drawer; center grasp around pineapple\",\n \"justification\": \"At around 10 seconds, the robot gripper closes over the pineapple without establishing a firm hold. The robot then fails to remove the pineapple, leaving it inside the drawer for the rest of the execution.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 632,
+    "failure_sim_step": 632,
+    "failure_policy_step": 105,
+    "failure_chunk_index": 2,
+    "first_post_failure_policy_step": 106,
+    "failure_type": "grasp_failure",
+    "correction_action": "Reopen the gripper, descend deeper into the drawer to align the fingers around the body of the pineapple, and close the gripper securely before lifting."
+  }
+}
\ No newline at end of file
diff --git a/pineapple_out_of_drawer/ep040/labels.npz b/pineapple_out_of_drawer/ep040/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..fb63304dc0e9adbb7fc2b3531eb9afc42efa88bf
--- /dev/null
+++ b/pineapple_out_of_drawer/ep040/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:24aa058ec390e628dc18cfc165f09e3deaa12a799902c9b57d0bc103b6162fa7
+size 3958
diff --git a/pineapple_out_of_drawer/ep041/example.md b/pineapple_out_of_drawer/ep041/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..733b885bd27f1ee1e8e44942dc39c37e4c9f522b
--- /dev/null
+++ b/pineapple_out_of_drawer/ep041/example.md
@@ -0,0 +1,37 @@
+# t007__ep041__pineapple_out_of_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the pineapple out of the drawer
+- **success:** False
+- **control steps:** 998  (19.5s at 51 Hz)
+- **policy inferences:** 171  (replan_steps=6)
+- **video:** `pineapple_out_of_drawer/ep041/rollout.mp4`
+- **features:** `171` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–152 | 0–25 | idle above workspace |
+| 3.0–6.0 | 153–305 | 25–50 | moving toward open drawer |
+| 6.0–9.0 | 306–458 | 51–76 | lowering toward pineapple |
+| 9.0–12.0 | 459–611 | 76–101 | attempting grasp on pineapple |
+| 12.0–15.0 | 612–764 | 102–127 | lifting empty gripper |
+| 15.0–18.0 | 765–917 | 127–152 | hovering above drawer |
+| 18.0–19.5 | 918–997 | 153–166 | idle above drawer |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 11.291585127201564 s (00:11.3)
+- **frame / control_step:** 577 (policy_step=96, chunk_index=1)
+- **coarse onset:** 11.5 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot descended to grasp the pineapple inside the open drawer, but failed to align its fingers properly, closing the gripper without establishing a hold on the object.
+  Summary: failed grasp attempt; gripper missed pineapple
+- **recovery:** Lower the end-effector back into the drawer, re-center the gripper fingers over the pineapple, close the fingers firmly to secure it, and lift the object out.
+  Summary: re-align over pineapple; grasp firmly; lift out of drawer
+- **prevention:** Ensure precise alignment and sufficient depth when approaching the pineapple in the drawer before closing the gripper fingers.
+  Summary: align gripper centered over object; close firmly before lifting
+- **justification:** Around 11.5 seconds, the robot attempts to grasp the pineapple inside the drawer. The fingers close around empty space/the edge, and at 13.0 seconds the robot lifts up without holding the pineapple.
diff --git a/pineapple_out_of_drawer/ep041/labels.json b/pineapple_out_of_drawer/ep041/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..0a6c18ef53c0827421954ee3e53e83a70b969369
--- /dev/null
+++ b/pineapple_out_of_drawer/ep041/labels.json
@@ -0,0 +1,142 @@
+{
+  "rollout_id": "t007__ep041__pineapple_out_of_drawer",
+  "scene_variant": "",
+  "task_id": 7,
+  "task_file": "",
+  "instruction": "Take the pineapple out of the drawer",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:33:47+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.99412915851272,
+      "control_step_start": 0,
+      "control_step_end": 152,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "idle above workspace",
+      "description": "idle above workspace"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.99412915851272,
+      "t_end_sec": 5.98825831702544,
+      "control_step_start": 153,
+      "control_step_end": 305,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "moving toward open drawer",
+      "description": "moving toward open drawer"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.98825831702544,
+      "t_end_sec": 8.982387475538161,
+      "control_step_start": 306,
+      "control_step_end": 458,
+      "policy_step_start": 51,
+      "policy_step_end": 76,
+      "phrase": "lowering toward pineapple",
+      "description": "lowering toward pineapple"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.982387475538161,
+      "t_end_sec": 11.97651663405088,
+      "control_step_start": 459,
+      "control_step_end": 611,
+      "policy_step_start": 76,
+      "policy_step_end": 101,
+      "phrase": "attempting grasp on pineapple",
+      "description": "attempting grasp on pineapple"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 11.97651663405088,
+      "t_end_sec": 14.9706457925636,
+      "control_step_start": 612,
+      "control_step_end": 764,
+      "policy_step_start": 102,
+      "policy_step_end": 127,
+      "phrase": "lifting empty gripper",
+      "description": "lifting empty gripper"
+    },
+    {
+      "segment_index": 5,
+      "t_start_sec": 14.9706457925636,
+      "t_end_sec": 17.964774951076322,
+      "control_step_start": 765,
+      "control_step_end": 917,
+      "policy_step_start": 127,
+      "policy_step_end": 152,
+      "phrase": "hovering above drawer",
+      "description": "hovering above drawer"
+    },
+    {
+      "segment_index": 6,
+      "t_start_sec": 17.964774951076322,
+      "t_end_sec": 19.53033268101761,
+      "control_step_start": 918,
+      "control_step_end": 997,
+      "policy_step_start": 153,
+      "policy_step_end": 166,
+      "phrase": "idle above drawer",
+      "description": "idle above drawer"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 11.291585127201564,
+    "vlm_failure_onset_timestamp": "00:11.3",
+    "vlm_failure_onset_frame": 577,
+    "vlm_failure_onset_step": 577,
+    "vlm_coarse_failure_onset_seconds": 11.5,
+    "vlm_failure_window_start_seconds": 10.5,
+    "vlm_failure_window_end_seconds": 13.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot descended to grasp the pineapple inside the open drawer, but failed to align its fingers properly, closing the gripper without establishing a hold on the object.",
+    "vlm_failure_reason_summary": "failed grasp attempt; gripper missed pineapple",
+    "vlm_recovery_action": "Lower the end-effector back into the drawer, re-center the gripper fingers over the pineapple, close the fingers firmly to secure it, and lift the object out.",
+    "vlm_recovery_action_summary": "re-align over pineapple; grasp firmly; lift out of drawer",
+    "vlm_prevention_action": "Ensure precise alignment and sufficient depth when approaching the pineapple in the drawer before closing the gripper fingers.",
+    "vlm_prevention_action_summary": "align gripper centered over object; close firmly before lifting",
+    "vlm_justification": "Around 11.5 seconds, the robot attempts to grasp the pineapple inside the drawer. The fingers close around empty space/the edge, and at 13.0 seconds the robot lifts up without holding the pineapple.",
+    "vlm_temporal_justification": "At timestamp 02:22 (142 seconds into the clip), the robot begins closing its gripper without properly aligning around the pineapple, causing the fingers to close without securing the object.",
+    "vlm_video_fps": 51.1,
+    "vlm_video_num_frames": 998,
+    "vlm_video_duration_seconds": 19.510763209393346,
+    "vlm_usage": {
+      "input_tokens": 22468,
+      "output_tokens": 376,
+      "cached_tokens": 1361
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 11.5,\n \"failure_window_start_seconds\": 10.5,\n \"failure_window_end_seconds\": 13.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot descended to grasp the pineapple inside the open drawer, but failed to align its fingers properly, closing the gripper without establishing a hold on the object.\",\n \"failure_reason_summary\": \"failed grasp attempt; gripper missed pineapple\",\n \"recovery_action\": \"Lower the end-effector back into the drawer, re-center the gripper fingers over the pineapple, close the fingers firmly to secure it, and lift the object out.\",\n \"recovery_action_summary\": \"re-align over pineapple; grasp firmly; lift out of drawer\",\n \"prevention_action\": \"Ensure precise alignment and sufficient depth when approaching the pineapple in the drawer before closing the gripper fingers.\",\n \"prevention_action_summary\": \"align gripper centered over object; close firmly before lifting\",\n \"justification\": \"Around 11.5 seconds, the robot attempts to grasp the pineapple inside the drawer. The fingers close around empty space/the edge, and at 13.0 seconds the robot lifts up without holding the pineapple.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 577,
+    "failure_sim_step": 577,
+    "failure_policy_step": 96,
+    "failure_chunk_index": 1,
+    "first_post_failure_policy_step": 97,
+    "failure_type": "grasp_failure",
+    "correction_action": "Lower the end-effector back into the drawer, re-center the gripper fingers over the pineapple, close the fingers firmly to secure it, and lift the object out."
+  }
+}
\ No newline at end of file
diff --git a/pineapple_out_of_drawer/ep041/labels.npz b/pineapple_out_of_drawer/ep041/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..26f56b98f338aec3b1bddd158f9fe1b1e18aadca
--- /dev/null
+++ b/pineapple_out_of_drawer/ep041/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9ac98aa71bc2d8b79d018b3a25bccae9c8aeeca8f991a89b2f77fcab03f0775d
+size 4003
diff --git a/pineapple_out_of_drawer/ep042/example.md b/pineapple_out_of_drawer/ep042/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..db65d4295f193775c829e1ee9ab19d940be52f87
--- /dev/null
+++ b/pineapple_out_of_drawer/ep042/example.md
@@ -0,0 +1,37 @@
+# t007__ep042__pineapple_out_of_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the pineapple out of the drawer
+- **success:** False
+- **control steps:** 1016  (20.2s at 50 Hz)
+- **policy inferences:** 175  (replan_steps=6)
+- **video:** `pineapple_out_of_drawer/ep042/rollout.mp4`
+- **features:** `175` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | arm stationary above desk |
+| 3.0–6.0 | 151–301 | 25–50 | descending toward drawer |
+| 6.0–9.0 | 302–452 | 50–75 | reaching toward drawer |
+| 9.0–12.0 | 453–603 | 75–100 | approaching pineapple in drawer |
+| 12.0–15.0 | 604–754 | 100–125 | lowering gripper into drawer |
+| 15.0–18.0 | 755–905 | 125–150 | attempting grasp on pineapple |
+| 18.0–20.2 | 906–1015 | 151–169 | idle inside drawer |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 14.900793650793652 s (00:14.9)
+- **frame / control_step:** 751 (policy_step=125, chunk_index=1)
+- **coarse onset:** 16.0 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot lowered its end-effector into the drawer but failed to align the gripper fingers around the pineapple, closing prematurely without securing the object.
+  Summary: improper gripper alignment; missed object grasp
+- **recovery:** Open the gripper, reposition the end-effector directly over the center of the pineapple inside the drawer, and attempt the grasp again.
+  Summary: reopen gripper; center end-effector; re-grasp pineapple
+- **prevention:** Ensure precise visual alignment with the pineapple's center of mass before closing the gripper fingers inside the drawer.
+  Summary: align gripper over object; grasp firmly
+- **justification:** At 16.0 seconds, the robot attempts to close its gripper around the pineapple in the drawer, but because the fingers are misaligned, it fails to clamp onto the pineapple and cannot pick it up.
diff --git a/pineapple_out_of_drawer/ep042/labels.json b/pineapple_out_of_drawer/ep042/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..b905ca992436954f434f6d9c642bb0580ed1b2d6
--- /dev/null
+++ b/pineapple_out_of_drawer/ep042/labels.json
@@ -0,0 +1,142 @@
+{
+  "rollout_id": "t007__ep042__pineapple_out_of_drawer",
+  "scene_variant": "",
+  "task_id": 7,
+  "task_file": "",
+  "instruction": "Take the pineapple out of the drawer",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:34:19+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.996031746031746,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "arm stationary above desk",
+      "description": "arm stationary above desk"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.996031746031746,
+      "t_end_sec": 5.992063492063492,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "descending toward drawer",
+      "description": "descending toward drawer"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.992063492063492,
+      "t_end_sec": 8.988095238095239,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "reaching toward drawer",
+      "description": "reaching toward drawer"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.988095238095239,
+      "t_end_sec": 11.984126984126984,
+      "control_step_start": 453,
+      "control_step_end": 603,
+      "policy_step_start": 75,
+      "policy_step_end": 100,
+      "phrase": "approaching pineapple in drawer",
+      "description": "approaching pineapple in drawer"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 11.984126984126984,
+      "t_end_sec": 14.980158730158731,
+      "control_step_start": 604,
+      "control_step_end": 754,
+      "policy_step_start": 100,
+      "policy_step_end": 125,
+      "phrase": "lowering gripper into drawer",
+      "description": "lowering gripper into drawer"
+    },
+    {
+      "segment_index": 5,
+      "t_start_sec": 14.980158730158731,
+      "t_end_sec": 17.976190476190478,
+      "control_step_start": 755,
+      "control_step_end": 905,
+      "policy_step_start": 125,
+      "policy_step_end": 150,
+      "phrase": "attempting grasp on pineapple",
+      "description": "attempting grasp on pineapple"
+    },
+    {
+      "segment_index": 6,
+      "t_start_sec": 17.976190476190478,
+      "t_end_sec": 20.158730158730158,
+      "control_step_start": 906,
+      "control_step_end": 1015,
+      "policy_step_start": 151,
+      "policy_step_end": 169,
+      "phrase": "idle inside drawer",
+      "description": "idle inside drawer"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 14.900793650793652,
+    "vlm_failure_onset_timestamp": "00:14.9",
+    "vlm_failure_onset_frame": 751,
+    "vlm_failure_onset_step": 751,
+    "vlm_coarse_failure_onset_seconds": 16.0,
+    "vlm_failure_window_start_seconds": 15.0,
+    "vlm_failure_window_end_seconds": 17.0,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot lowered its end-effector into the drawer but failed to align the gripper fingers around the pineapple, closing prematurely without securing the object.",
+    "vlm_failure_reason_summary": "improper gripper alignment; missed object grasp",
+    "vlm_recovery_action": "Open the gripper, reposition the end-effector directly over the center of the pineapple inside the drawer, and attempt the grasp again.",
+    "vlm_recovery_action_summary": "reopen gripper; center end-effector; re-grasp pineapple",
+    "vlm_prevention_action": "Ensure precise visual alignment with the pineapple's center of mass before closing the gripper fingers inside the drawer.",
+    "vlm_prevention_action_summary": "align gripper over object; grasp firmly",
+    "vlm_justification": "At 16.0 seconds, the robot attempts to close its gripper around the pineapple in the drawer, but because the fingers are misaligned, it fails to clamp onto the pineapple and cannot pick it up.",
+    "vlm_temporal_justification": "At timestamp 01:36 (96 seconds into the clip), the robot begins closing its gripper fingers prematurely without properly aligning around the pineapple, leading directly to the failed grasp.",
+    "vlm_video_fps": 50.4,
+    "vlm_video_num_frames": 1016,
+    "vlm_video_duration_seconds": 20.13888888888889,
+    "vlm_usage": {
+      "input_tokens": 22213,
+      "output_tokens": 359,
+      "cached_tokens": 1361
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 16.0,\n \"failure_window_start_seconds\": 15.0,\n \"failure_window_end_seconds\": 17.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot lowered its end-effector into the drawer but failed to align the gripper fingers around the pineapple, closing prematurely without securing the object.\",\n \"failure_reason_summary\": \"improper gripper alignment; missed object grasp\",\n \"recovery_action\": \"Open the gripper, reposition the end-effector directly over the center of the pineapple inside the drawer, and attempt the grasp again.\",\n \"recovery_action_summary\": \"reopen gripper; center end-effector; re-grasp pineapple\",\n \"prevention_action\": \"Ensure precise visual alignment with the pineapple's center of mass before closing the gripper fingers inside the drawer.\",\n \"prevention_action_summary\": \"align gripper over object; grasp firmly\",\n \"justification\": \"At 16.0 seconds, the robot attempts to close its gripper around the pineapple in the drawer, but because the fingers are misaligned, it fails to clamp onto the pineapple and cannot pick it up.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 751,
+    "failure_sim_step": 751,
+    "failure_policy_step": 125,
+    "failure_chunk_index": 1,
+    "first_post_failure_policy_step": 126,
+    "failure_type": "grasp_failure",
+    "correction_action": "Open the gripper, reposition the end-effector directly over the center of the pineapple inside the drawer, and attempt the grasp again."
+  }
+}
\ No newline at end of file
diff --git a/pineapple_out_of_drawer/ep042/labels.npz b/pineapple_out_of_drawer/ep042/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..9cdd734515ee08a66e6ac3f9a0146d9815056c89
--- /dev/null
+++ b/pineapple_out_of_drawer/ep042/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:97836287f3bd4a63aa63dd7876959d95716c5e972ae515f5dbc850c5d321fafa
+size 3960
diff --git a/pineapple_out_of_drawer/ep043/example.md b/pineapple_out_of_drawer/ep043/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..d37131bfd6a53200b910af154a4205b5c47c0500
--- /dev/null
+++ b/pineapple_out_of_drawer/ep043/example.md
@@ -0,0 +1,34 @@
+# t007__ep043__pineapple_out_of_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the pineapple out of the drawer
+- **success:** False
+- **control steps:** 522  (10.3s at 51 Hz)
+- **policy inferences:** 90  (replan_steps=6)
+- **video:** `pineapple_out_of_drawer/ep043/rollout.mp4`
+- **features:** `90` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | hovering above table |
+| 3.0–6.0 | 152–303 | 25–50 | lowering arm toward pineapple |
+| 6.0–9.0 | 304–455 | 50–75 | attempting grasp on pineapple |
+| 9.0–10.3 | 456–521 | 76–86 | hovering above open drawer |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 8.60236220472441 s (00:08.6)
+- **frame / control_step:** 437 (policy_step=72, chunk_index=5)
+- **coarse onset:** 8.2 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the top leaves of the pineapple, but failed to establish a secure hold, causing the gripper to slip off without lifting the pineapple.
+  Summary: failed grasp on pineapple; gripper slipped off
+- **recovery:** Re-align the end-effector over the body of the pineapple, close the gripper firmly, and lift it out of the drawer.
+  Summary: re-align gripper on pineapple; grasp firmly; lift out of drawer
+- **prevention:** Position the gripper lower around the main body of the pineapple rather than the top leaves before closing to ensure a stable grip.
+  Summary: position gripper around pineapple body; secure firm grasp
+- **justification:** At around 8.2 seconds, the end-effector attempts to close on the top leaves of the pineapple inside the open drawer. The fingers slip off the object during the grasp attempt, leaving the pineapple in the drawer and rendering the robot unable to complete the task.
diff --git a/pineapple_out_of_drawer/ep043/labels.json b/pineapple_out_of_drawer/ep043/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..cc8d2b49304931cba1a7036fa26e32f3259ce2d3
--- /dev/null
+++ b/pineapple_out_of_drawer/ep043/labels.json
@@ -0,0 +1,109 @@
+{
+  "rollout_id": "t007__ep043__pineapple_out_of_drawer",
+  "scene_variant": "",
+  "task_id": 7,
+  "task_file": "",
+  "instruction": "Take the pineapple out of the drawer",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:34:52+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9921259842519685,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "hovering above table",
+      "description": "hovering above table"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9921259842519685,
+      "t_end_sec": 5.984251968503937,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "lowering arm toward pineapple",
+      "description": "lowering arm toward pineapple"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.984251968503937,
+      "t_end_sec": 8.976377952755906,
+      "control_step_start": 304,
+      "control_step_end": 455,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "attempting grasp on pineapple",
+      "description": "attempting grasp on pineapple"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.976377952755906,
+      "t_end_sec": 10.275590551181104,
+      "control_step_start": 456,
+      "control_step_end": 521,
+      "policy_step_start": 76,
+      "policy_step_end": 86,
+      "phrase": "hovering above open drawer",
+      "description": "hovering above open drawer"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 8.60236220472441,
+    "vlm_failure_onset_timestamp": "00:08.6",
+    "vlm_failure_onset_frame": 437,
+    "vlm_failure_onset_step": 437,
+    "vlm_coarse_failure_onset_seconds": 8.2,
+    "vlm_failure_window_start_seconds": 7.7,
+    "vlm_failure_window_end_seconds": 9.2,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the top leaves of the pineapple, but failed to establish a secure hold, causing the gripper to slip off without lifting the pineapple.",
+    "vlm_failure_reason_summary": "failed grasp on pineapple; gripper slipped off",
+    "vlm_recovery_action": "Re-align the end-effector over the body of the pineapple, close the gripper firmly, and lift it out of the drawer.",
+    "vlm_recovery_action_summary": "re-align gripper on pineapple; grasp firmly; lift out of drawer",
+    "vlm_prevention_action": "Position the gripper lower around the main body of the pineapple rather than the top leaves before closing to ensure a stable grip.",
+    "vlm_prevention_action_summary": "position gripper around pineapple body; secure firm grasp",
+    "vlm_justification": "At around 8.2 seconds, the end-effector attempts to close on the top leaves of the pineapple inside the open drawer. The fingers slip off the object during the grasp attempt, leaving the pineapple in the drawer and rendering the robot unable to complete the task.",
+    "vlm_temporal_justification": "At timestamp 02:52 (172 seconds into the magnified clip), the robot closes its gripper onto the upper tips of the pineapple leaves. This improper grasp positioning fails to secure the object, causing the gripper to slip off during the lift attempt.",
+    "vlm_video_fps": 50.8,
+    "vlm_video_num_frames": 522,
+    "vlm_video_duration_seconds": 10.255905511811024,
+    "vlm_usage": {
+      "input_tokens": 18686,
+      "output_tokens": 393,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 8.2,\n \"failure_window_start_seconds\": 7.7,\n \"failure_window_end_seconds\": 9.2,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the top leaves of the pineapple, but failed to establish a secure hold, causing the gripper to slip off without lifting the pineapple.\",\n \"failure_reason_summary\": \"failed grasp on pineapple; gripper slipped off\",\n \"recovery_action\": \"Re-align the end-effector over the body of the pineapple, close the gripper firmly, and lift it out of the drawer.\",\n \"recovery_action_summary\": \"re-align gripper on pineapple; grasp firmly; lift out of drawer\",\n \"prevention_action\": \"Position the gripper lower around the main body of the pineapple rather than the top leaves before closing to ensure a stable grip.\",\n \"prevention_action_summary\": \"position gripper around pineapple body; secure firm grasp\",\n \"justification\": \"At around 8.2 seconds, the end-effector attempts to close on the top leaves of the pineapple inside the open drawer. The fingers slip off the object during the grasp attempt, leaving the pineapple in the drawer and rendering the robot unable to complete the task.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 437,
+    "failure_sim_step": 437,
+    "failure_policy_step": 72,
+    "failure_chunk_index": 5,
+    "first_post_failure_policy_step": 73,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-align the end-effector over the body of the pineapple, close the gripper firmly, and lift it out of the drawer."
+  }
+}
\ No newline at end of file
diff --git a/pineapple_out_of_drawer/ep043/labels.npz b/pineapple_out_of_drawer/ep043/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..7f68c096f75778706b23ff360da5f5156669c261
--- /dev/null
+++ b/pineapple_out_of_drawer/ep043/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:54512a58f4793f3dc3b87e379d5cccf343d0170f4eb70df4bc0b615d9170ba53
+size 3838
diff --git a/pineapple_out_of_drawer/ep044/example.md b/pineapple_out_of_drawer/ep044/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..773848a61d6576e35538736f428bc4d5cb4ebb9a
--- /dev/null
+++ b/pineapple_out_of_drawer/ep044/example.md
@@ -0,0 +1,35 @@
+# t007__ep044__pineapple_out_of_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the pineapple out of the drawer
+- **success:** False
+- **control steps:** 671  (13.2s at 51 Hz)
+- **policy inferences:** 116  (replan_steps=6)
+- **video:** `pineapple_out_of_drawer/ep044/rollout.mp4`
+- **features:** `116` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–152 | 0–25 | hovering above table |
+| 3.0–6.0 | 153–305 | 25–50 | moving toward open drawer |
+| 6.0–9.0 | 306–458 | 51–76 | reaching inside drawer for pineapple |
+| 9.0–12.0 | 459–611 | 76–101 | attempting to grasp pineapple |
+| 12.0–13.2 | 612–670 | 102–111 | remaining stationary in drawer |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 8.094302554027506 s (00:08.1)
+- **frame / control_step:** 412 (policy_step=68, chunk_index=4)
+- **coarse onset:** 8.0 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot lowered its end-effector over the pineapple in the drawer and closed its gripper, but failed to establish a secure grasp on the object.
+  Summary: failed to grasp pineapple; improper gripper closure
+- **recovery:** Re-open the gripper, realign the end-effector directly over the center of the pineapple, lower the arm further into the drawer, and close the gripper securely before lifting.
+  Summary: re-open gripper; align and grasp pineapple
+- **prevention:** Ensure the gripper descends deep enough into the drawer and centers around the pineapple before closing the fingers to secure a firm hold.
+  Summary: align gripper over pineapple; close fingers securely
+- **justification:** Around 8.0 seconds into the rollout, the robot attempts to pick up the pineapple inside the drawer. However, the gripper closes without properly surrounding or clamping the object, resulting in a grasp failure.
diff --git a/pineapple_out_of_drawer/ep044/labels.json b/pineapple_out_of_drawer/ep044/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..3b91c232aee629957aa3e360d64552f7c76e8df1
--- /dev/null
+++ b/pineapple_out_of_drawer/ep044/labels.json
@@ -0,0 +1,120 @@
+{
+  "rollout_id": "t007__ep044__pineapple_out_of_drawer",
+  "scene_variant": "",
+  "task_id": 7,
+  "task_file": "",
+  "instruction": "Take the pineapple out of the drawer",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:35:22+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.005893909626719,
+      "control_step_start": 0,
+      "control_step_end": 152,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "hovering above table",
+      "description": "hovering above table"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.005893909626719,
+      "t_end_sec": 6.011787819253438,
+      "control_step_start": 153,
+      "control_step_end": 305,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "moving toward open drawer",
+      "description": "moving toward open drawer"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.011787819253438,
+      "t_end_sec": 9.017681728880158,
+      "control_step_start": 306,
+      "control_step_end": 458,
+      "policy_step_start": 51,
+      "policy_step_end": 76,
+      "phrase": "reaching inside drawer for pineapple",
+      "description": "reaching inside drawer for pineapple"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.017681728880158,
+      "t_end_sec": 12.023575638506877,
+      "control_step_start": 459,
+      "control_step_end": 611,
+      "policy_step_start": 76,
+      "policy_step_end": 101,
+      "phrase": "attempting to grasp pineapple",
+      "description": "attempting to grasp pineapple"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 12.023575638506877,
+      "t_end_sec": 13.182711198428292,
+      "control_step_start": 612,
+      "control_step_end": 670,
+      "policy_step_start": 102,
+      "policy_step_end": 111,
+      "phrase": "remaining stationary in drawer",
+      "description": "remaining stationary in drawer"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 8.094302554027506,
+    "vlm_failure_onset_timestamp": "00:08.1",
+    "vlm_failure_onset_frame": 412,
+    "vlm_failure_onset_step": 412,
+    "vlm_coarse_failure_onset_seconds": 8.0,
+    "vlm_failure_window_start_seconds": 8.0,
+    "vlm_failure_window_end_seconds": 11.0,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot lowered its end-effector over the pineapple in the drawer and closed its gripper, but failed to establish a secure grasp on the object.",
+    "vlm_failure_reason_summary": "failed to grasp pineapple; improper gripper closure",
+    "vlm_recovery_action": "Re-open the gripper, realign the end-effector directly over the center of the pineapple, lower the arm further into the drawer, and close the gripper securely before lifting.",
+    "vlm_recovery_action_summary": "re-open gripper; align and grasp pineapple",
+    "vlm_prevention_action": "Ensure the gripper descends deep enough into the drawer and centers around the pineapple before closing the fingers to secure a firm hold.",
+    "vlm_prevention_action_summary": "align gripper over pineapple; close fingers securely",
+    "vlm_justification": "Around 8.0 seconds into the rollout, the robot attempts to pick up the pineapple inside the drawer. However, the gripper closes without properly surrounding or clamping the object, resulting in a grasp failure.",
+    "vlm_temporal_justification": "At clip second 02:38 (frame 158 of the clip), the robot gripper begins closing its fingers over the pineapple, but fails to position properly around the object's body, resulting in a failed grasp when fully closed at 02:40.",
+    "vlm_video_fps": 50.9,
+    "vlm_video_num_frames": 671,
+    "vlm_video_duration_seconds": 13.163064833005894,
+    "vlm_usage": {
+      "input_tokens": 22023,
+      "output_tokens": 387,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 8.0,\n \"failure_window_start_seconds\": 8.0,\n \"failure_window_end_seconds\": 11.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot lowered its end-effector over the pineapple in the drawer and closed its gripper, but failed to establish a secure grasp on the object.\",\n \"failure_reason_summary\": \"failed to grasp pineapple; improper gripper closure\",\n \"recovery_action\": \"Re-open the gripper, realign the end-effector directly over the center of the pineapple, lower the arm further into the drawer, and close the gripper securely before lifting.\",\n \"recovery_action_summary\": \"re-open gripper; align and grasp pineapple\",\n \"prevention_action\": \"Ensure the gripper descends deep enough into the drawer and centers around the pineapple before closing the fingers to secure a firm hold.\",\n \"prevention_action_summary\": \"align gripper over pineapple; close fingers securely\",\n \"justification\": \"Around 8.0 seconds into the rollout, the robot attempts to pick up the pineapple inside the drawer. However, the gripper closes without properly surrounding or clamping the object, resulting in a grasp failure.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 412,
+    "failure_sim_step": 412,
+    "failure_policy_step": 68,
+    "failure_chunk_index": 4,
+    "first_post_failure_policy_step": 69,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-open the gripper, realign the end-effector directly over the center of the pineapple, lower the arm further into the drawer, and close the gripper securely before lifting."
+  }
+}
\ No newline at end of file
diff --git a/pineapple_out_of_drawer/ep044/labels.npz b/pineapple_out_of_drawer/ep044/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..01f1cd61dbdd8c61c6d3af7c3d838fe81d8d3874
--- /dev/null
+++ b/pineapple_out_of_drawer/ep044/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a7686ebf2db497bae93fbce993812cec512f7a3281d444fbe15654e53b3cab3c
+size 3926
diff --git a/pineapple_out_of_drawer/ep045/example.md b/pineapple_out_of_drawer/ep045/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..19835f28f5ebdee3e9045efd6ac68889970ff1be
--- /dev/null
+++ b/pineapple_out_of_drawer/ep045/example.md
@@ -0,0 +1,23 @@
+# t007__ep045__pineapple_out_of_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the pineapple out of the drawer
+- **success:** True
+- **control steps:** 650  (12.7s at 51 Hz)
+- **policy inferences:** 111  (replan_steps=6)
+- **video:** `pineapple_out_of_drawer/ep045/rollout.mp4`
+- **features:** `111` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–152 | 0–25 | reaching for pineapple |
+| 3.0–6.0 | 153–305 | 25–50 | grasping the pineapple |
+| 6.0–9.0 | 306–458 | 51–76 | lifting pineapple from drawer |
+| 9.0–12.0 | 459–611 | 76–101 | placing pineapple on table |
+| 12.0–12.7 | 612–649 | 102–108 | releasing pineapple on table |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/pineapple_out_of_drawer/ep045/labels.json b/pineapple_out_of_drawer/ep045/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..b3088cb07140550900ed89ce210ae342e36bce02
--- /dev/null
+++ b/pineapple_out_of_drawer/ep045/labels.json
@@ -0,0 +1,90 @@
+{
+  "rollout_id": "t007__ep045__pineapple_out_of_drawer",
+  "scene_variant": "",
+  "task_id": 7,
+  "task_file": "",
+  "instruction": "Take the pineapple out of the drawer",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:35:30+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0,
+      "control_step_start": 0,
+      "control_step_end": 152,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for pineapple",
+      "description": "reaching for pineapple"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0,
+      "t_end_sec": 6.0,
+      "control_step_start": 153,
+      "control_step_end": 305,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "grasping the pineapple",
+      "description": "grasping the pineapple"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.0,
+      "t_end_sec": 9.0,
+      "control_step_start": 306,
+      "control_step_end": 458,
+      "policy_step_start": 51,
+      "policy_step_end": 76,
+      "phrase": "lifting pineapple from drawer",
+      "description": "lifting pineapple from drawer"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.0,
+      "t_end_sec": 12.0,
+      "control_step_start": 459,
+      "control_step_end": 611,
+      "policy_step_start": 76,
+      "policy_step_end": 101,
+      "phrase": "placing pineapple on table",
+      "description": "placing pineapple on table"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 12.0,
+      "t_end_sec": 12.745098039215685,
+      "control_step_start": 612,
+      "control_step_end": 649,
+      "policy_step_start": 102,
+      "policy_step_end": 108,
+      "phrase": "releasing pineapple on table",
+      "description": "releasing pineapple on table"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/pineapple_out_of_drawer/ep045/labels.npz b/pineapple_out_of_drawer/ep045/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..93f52bd3f7729fc52b4365acda6aab1de350855f
--- /dev/null
+++ b/pineapple_out_of_drawer/ep045/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f2d1e394d0e9582facbd4b7e222e0a72b79594d6f0c051e3925f8e2c5d6e16ed
+size 3134
diff --git a/pineapple_out_of_drawer/ep046/example.md b/pineapple_out_of_drawer/ep046/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..d3a1981a119070f5510dbee2456ab3de506f267d
--- /dev/null
+++ b/pineapple_out_of_drawer/ep046/example.md
@@ -0,0 +1,35 @@
+# t007__ep046__pineapple_out_of_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the pineapple out of the drawer
+- **success:** False
+- **control steps:** 654  (13.1s at 50 Hz)
+- **policy inferences:** 113  (replan_steps=6)
+- **video:** `pineapple_out_of_drawer/ep046/rollout.mp4`
+- **features:** `113` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–149 | 0–24 | approaching drawer with pineapple |
+| 3.0–6.0 | 150–299 | 25–49 | lowering gripper into drawer |
+| 6.0–9.0 | 300–449 | 50–74 | attempting grasp on pineapple |
+| 9.0–12.0 | 450–599 | 75–99 | hovering above drawer empty-handed |
+| 12.0–13.1 | 600–653 | 100–108 | idle above open drawer |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 6.107784431137724 s (00:06.1)
+- **frame / control_step:** 306 (policy_step=51, chunk_index=0)
+- **coarse onset:** 6.0 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the pineapple inside the open drawer, but failed to secure a firm grip, leaving the pineapple behind when lifting the end-effector.
+  Summary: failed grasp on pineapple; object remains in drawer
+- **recovery:** Re-align the gripper over the pineapple in the drawer, descend to surround the object, close the gripper securely, and lift the pineapple out of the drawer.
+  Summary: re-align gripper; firmly grasp pineapple; lift out
+- **prevention:** Accurately position the gripper around the pineapple's center of mass before closing fingers to ensure a successful hold.
+  Summary: center gripper on pineapple; secure grasp before lifting
+- **justification:** At around 6.0 seconds into the rollout, the robot attempts to grasp the pineapple inside the open drawer. However, the gripper fails to enclose the object properly, and as the arm moves away, the pineapple remains in the drawer.
diff --git a/pineapple_out_of_drawer/ep046/labels.json b/pineapple_out_of_drawer/ep046/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..ece8503ca543ba9abbd6309bab4ad1c6231cec72
--- /dev/null
+++ b/pineapple_out_of_drawer/ep046/labels.json
@@ -0,0 +1,120 @@
+{
+  "rollout_id": "t007__ep046__pineapple_out_of_drawer",
+  "scene_variant": "",
+  "task_id": 7,
+  "task_file": "",
+  "instruction": "Take the pineapple out of the drawer",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:36:01+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9940119760479043,
+      "control_step_start": 0,
+      "control_step_end": 149,
+      "policy_step_start": 0,
+      "policy_step_end": 24,
+      "phrase": "approaching drawer with pineapple",
+      "description": "approaching drawer with pineapple"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9940119760479043,
+      "t_end_sec": 5.9880239520958085,
+      "control_step_start": 150,
+      "control_step_end": 299,
+      "policy_step_start": 25,
+      "policy_step_end": 49,
+      "phrase": "lowering gripper into drawer",
+      "description": "lowering gripper into drawer"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.9880239520958085,
+      "t_end_sec": 8.982035928143713,
+      "control_step_start": 300,
+      "control_step_end": 449,
+      "policy_step_start": 50,
+      "policy_step_end": 74,
+      "phrase": "attempting grasp on pineapple",
+      "description": "attempting grasp on pineapple"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.982035928143713,
+      "t_end_sec": 11.976047904191617,
+      "control_step_start": 450,
+      "control_step_end": 599,
+      "policy_step_start": 75,
+      "policy_step_end": 99,
+      "phrase": "hovering above drawer empty-handed",
+      "description": "hovering above drawer empty-handed"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 11.976047904191617,
+      "t_end_sec": 13.053892215568862,
+      "control_step_start": 600,
+      "control_step_end": 653,
+      "policy_step_start": 100,
+      "policy_step_end": 108,
+      "phrase": "idle above open drawer",
+      "description": "idle above open drawer"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 6.107784431137724,
+    "vlm_failure_onset_timestamp": "00:06.1",
+    "vlm_failure_onset_frame": 306,
+    "vlm_failure_onset_step": 306,
+    "vlm_coarse_failure_onset_seconds": 6.0,
+    "vlm_failure_window_start_seconds": 5.0,
+    "vlm_failure_window_end_seconds": 8.0,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the pineapple inside the open drawer, but failed to secure a firm grip, leaving the pineapple behind when lifting the end-effector.",
+    "vlm_failure_reason_summary": "failed grasp on pineapple; object remains in drawer",
+    "vlm_recovery_action": "Re-align the gripper over the pineapple in the drawer, descend to surround the object, close the gripper securely, and lift the pineapple out of the drawer.",
+    "vlm_recovery_action_summary": "re-align gripper; firmly grasp pineapple; lift out",
+    "vlm_prevention_action": "Accurately position the gripper around the pineapple's center of mass before closing fingers to ensure a successful hold.",
+    "vlm_prevention_action_summary": "center gripper on pineapple; secure grasp before lifting",
+    "vlm_justification": "At around 6.0 seconds into the rollout, the robot attempts to grasp the pineapple inside the open drawer. However, the gripper fails to enclose the object properly, and as the arm moves away, the pineapple remains in the drawer.",
+    "vlm_temporal_justification": "At 02:35 in this clip, the end-effector begins lifting upward away from the drawer, leaving the pineapple behind inside the drawer due to an ineffective grasp.",
+    "vlm_video_fps": 50.1,
+    "vlm_video_num_frames": 654,
+    "vlm_video_duration_seconds": 13.033932135728543,
+    "vlm_usage": {
+      "input_tokens": 21647,
+      "output_tokens": 371,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 6.0,\n \"failure_window_start_seconds\": 5.0,\n \"failure_window_end_seconds\": 8.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the pineapple inside the open drawer, but failed to secure a firm grip, leaving the pineapple behind when lifting the end-effector.\",\n \"failure_reason_summary\": \"failed grasp on pineapple; object remains in drawer\",\n \"recovery_action\": \"Re-align the gripper over the pineapple in the drawer, descend to surround the object, close the gripper securely, and lift the pineapple out of the drawer.\",\n \"recovery_action_summary\": \"re-align gripper; firmly grasp pineapple; lift out\",\n \"prevention_action\": \"Accurately position the gripper around the pineapple's center of mass before closing fingers to ensure a successful hold.\",\n \"prevention_action_summary\": \"center gripper on pineapple; secure grasp before lifting\",\n \"justification\": \"At around 6.0 seconds into the rollout, the robot attempts to grasp the pineapple inside the open drawer. However, the gripper fails to enclose the object properly, and as the arm moves away, the pineapple remains in the drawer.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 306,
+    "failure_sim_step": 306,
+    "failure_policy_step": 51,
+    "failure_chunk_index": 0,
+    "first_post_failure_policy_step": 52,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-align the gripper over the pineapple in the drawer, descend to surround the object, close the gripper securely, and lift the pineapple out of the drawer."
+  }
+}
\ No newline at end of file
diff --git a/pineapple_out_of_drawer/ep046/labels.npz b/pineapple_out_of_drawer/ep046/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..718ff87f3010448a3b2092e07d0d2ee4bf4ddf44
--- /dev/null
+++ b/pineapple_out_of_drawer/ep046/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4c9928379f6b642ba87eea3102ff4f3926fff1f2362ede9c0aa513b091174090
+size 3922
diff --git a/pineapple_out_of_drawer/ep047/example.md b/pineapple_out_of_drawer/ep047/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..d5bd70b25fcbdc345eb5a6f681c55cbbf97d3955
--- /dev/null
+++ b/pineapple_out_of_drawer/ep047/example.md
@@ -0,0 +1,34 @@
+# t007__ep047__pineapple_out_of_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the pineapple out of the drawer
+- **success:** False
+- **control steps:** 548  (10.9s at 50 Hz)
+- **policy inferences:** 94  (replan_steps=6)
+- **video:** `pineapple_out_of_drawer/ep047/rollout.mp4`
+- **features:** `94` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | moving arm toward open drawer |
+| 3.0–6.0 | 151–301 | 25–50 | lowering gripper over pineapple |
+| 6.0–9.0 | 302–452 | 50–75 | attempting to grasp pineapple |
+| 9.0–10.9 | 453–547 | 75–91 | hovering with empty gripper |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 7.495029821073559 s (00:07.5)
+- **frame / control_step:** 377 (policy_step=62, chunk_index=5)
+- **coarse onset:** 7.5 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot reaches into the open drawer to grab the pineapple, but fails to position its gripper correctly, resulting in the fingers closing without securing the object.
+  Summary: improper gripper alignment; missed grasp on pineapple
+- **recovery:** Re-open the gripper, center the end-effector directly over the pineapple, and execute a firm grasp before lifting it out of the drawer.
+  Summary: open gripper; align with pineapple; grasp securely
+- **prevention:** Position the gripper directly over the center of the pineapple at a suitable depth before closing the fingers to ensure a secure grip.
+  Summary: align gripper over object; adjust grasp depth
+- **justification:** Around timestamp 00:07 to 00:08, the robot lowers its end-effector into the drawer and closes its fingers, but fails to enclose or grasp the pineapple, resulting in a grasp failure.
diff --git a/pineapple_out_of_drawer/ep047/labels.json b/pineapple_out_of_drawer/ep047/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..eae8071c5ba30ff3166010c41b21b16f70992f3c
--- /dev/null
+++ b/pineapple_out_of_drawer/ep047/labels.json
@@ -0,0 +1,109 @@
+{
+  "rollout_id": "t007__ep047__pineapple_out_of_drawer",
+  "scene_variant": "",
+  "task_id": 7,
+  "task_file": "",
+  "instruction": "Take the pineapple out of the drawer",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:36:33+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.001988071570577,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "moving arm toward open drawer",
+      "description": "moving arm toward open drawer"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.001988071570577,
+      "t_end_sec": 6.003976143141154,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "lowering gripper over pineapple",
+      "description": "lowering gripper over pineapple"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.003976143141154,
+      "t_end_sec": 9.00596421471173,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "attempting to grasp pineapple",
+      "description": "attempting to grasp pineapple"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.00596421471173,
+      "t_end_sec": 10.894632206759445,
+      "control_step_start": 453,
+      "control_step_end": 547,
+      "policy_step_start": 75,
+      "policy_step_end": 91,
+      "phrase": "hovering with empty gripper",
+      "description": "hovering with empty gripper"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 7.495029821073559,
+    "vlm_failure_onset_timestamp": "00:07.5",
+    "vlm_failure_onset_frame": 377,
+    "vlm_failure_onset_step": 377,
+    "vlm_coarse_failure_onset_seconds": 7.5,
+    "vlm_failure_window_start_seconds": 7.0,
+    "vlm_failure_window_end_seconds": 8.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot reaches into the open drawer to grab the pineapple, but fails to position its gripper correctly, resulting in the fingers closing without securing the object.",
+    "vlm_failure_reason_summary": "improper gripper alignment; missed grasp on pineapple",
+    "vlm_recovery_action": "Re-open the gripper, center the end-effector directly over the pineapple, and execute a firm grasp before lifting it out of the drawer.",
+    "vlm_recovery_action_summary": "open gripper; align with pineapple; grasp securely",
+    "vlm_prevention_action": "Position the gripper directly over the center of the pineapple at a suitable depth before closing the fingers to ensure a secure grip.",
+    "vlm_prevention_action_summary": "align gripper over object; adjust grasp depth",
+    "vlm_justification": "Around timestamp 00:07 to 00:08, the robot lowers its end-effector into the drawer and closes its fingers, but fails to enclose or grasp the pineapple, resulting in a grasp failure.",
+    "vlm_temporal_justification": "At second 02:31 (151 seconds into the clip), the robot gripper begins closing its fingers while misaligned inside the drawer tray, failing to grasp the pineapple and closing empty.",
+    "vlm_video_fps": 50.3,
+    "vlm_video_num_frames": 548,
+    "vlm_video_duration_seconds": 10.87475149105368,
+    "vlm_usage": {
+      "input_tokens": 21646,
+      "output_tokens": 369,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 7.5,\n \"failure_window_start_seconds\": 7.0,\n \"failure_window_end_seconds\": 8.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot reaches into the open drawer to grab the pineapple, but fails to position its gripper correctly, resulting in the fingers closing without securing the object.\",\n \"failure_reason_summary\": \"improper gripper alignment; missed grasp on pineapple\",\n \"recovery_action\": \"Re-open the gripper, center the end-effector directly over the pineapple, and execute a firm grasp before lifting it out of the drawer.\",\n \"recovery_action_summary\": \"open gripper; align with pineapple; grasp securely\",\n \"prevention_action\": \"Position the gripper directly over the center of the pineapple at a suitable depth before closing the fingers to ensure a secure grip.\",\n \"prevention_action_summary\": \"align gripper over object; adjust grasp depth\",\n \"justification\": \"Around timestamp 00:07 to 00:08, the robot lowers its end-effector into the drawer and closes its fingers, but fails to enclose or grasp the pineapple, resulting in a grasp failure.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 377,
+    "failure_sim_step": 377,
+    "failure_policy_step": 62,
+    "failure_chunk_index": 5,
+    "first_post_failure_policy_step": 63,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-open the gripper, center the end-effector directly over the pineapple, and execute a firm grasp before lifting it out of the drawer."
+  }
+}
\ No newline at end of file
diff --git a/pineapple_out_of_drawer/ep047/labels.npz b/pineapple_out_of_drawer/ep047/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..56951dd7df2d8cfcb66092304dde55b319b1683d
--- /dev/null
+++ b/pineapple_out_of_drawer/ep047/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4dfe04082eeecd96e273f2cc01c3755ab265c66f49823bb3872ea3a88d50f3c1
+size 3867
diff --git a/pineapple_out_of_drawer/ep048/example.md b/pineapple_out_of_drawer/ep048/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..974be60a781666ebf71c0416de638a63c0c7fe85
--- /dev/null
+++ b/pineapple_out_of_drawer/ep048/example.md
@@ -0,0 +1,36 @@
+# t007__ep048__pineapple_out_of_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the pineapple out of the drawer
+- **success:** False
+- **control steps:** 778  (15.3s at 51 Hz)
+- **policy inferences:** 135  (replan_steps=6)
+- **video:** `pineapple_out_of_drawer/ep048/rollout.mp4`
+- **features:** `135` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching toward open drawer |
+| 3.0–6.0 | 152–303 | 25–50 | attempting grasp on pineapple |
+| 6.0–9.0 | 304–455 | 50–75 | retracting arm empty-handed |
+| 9.0–12.0 | 456–607 | 76–101 | hovering above drawer idle |
+| 12.0–15.0 | 608–759 | 101–126 | hovering above drawer idle |
+| 15.0–15.3 | 760–777 | 126–129 | remaining idle above table |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 3.826429980276134 s (00:03.8)
+- **frame / control_step:** 194 (policy_step=32, chunk_index=2)
+- **coarse onset:** 5.0 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the pineapple inside the open drawer, but failed to secure its grip, retracting its gripper empty-handed.
+  Summary: failed grasp attempt on pineapple; gripper missed hold
+- **recovery:** Re-align the end-effector over the pineapple inside the drawer, close the fingers firmly around it, and lift the pineapple completely out of the drawer.
+  Summary: re-align gripper over pineapple; grasp and remove object
+- **prevention:** Ensure the gripper is properly centered and lowered around the pineapple before closing the fingers to achieve a secure grip.
+  Summary: center gripper over pineapple; secure full grip
+- **justification:** Around 5.0 seconds, the robot's gripper reaches into the open drawer and attempts to grasp the pineapple, but fails to close securely on it. The end-effector then pulls back empty and hovers above the table, resulting in task failure.
diff --git a/pineapple_out_of_drawer/ep048/labels.json b/pineapple_out_of_drawer/ep048/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..9988413e1bbdab1fa2cb60b47f57757fabdd4b68
--- /dev/null
+++ b/pineapple_out_of_drawer/ep048/labels.json
@@ -0,0 +1,131 @@
+{
+  "rollout_id": "t007__ep048__pineapple_out_of_drawer",
+  "scene_variant": "",
+  "task_id": 7,
+  "task_file": "",
+  "instruction": "Take the pineapple out of the drawer",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:37:04+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9980276134122286,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching toward open drawer",
+      "description": "reaching toward open drawer"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9980276134122286,
+      "t_end_sec": 5.996055226824457,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "attempting grasp on pineapple",
+      "description": "attempting grasp on pineapple"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.996055226824457,
+      "t_end_sec": 8.994082840236686,
+      "control_step_start": 304,
+      "control_step_end": 455,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "retracting arm empty-handed",
+      "description": "retracting arm empty-handed"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.994082840236686,
+      "t_end_sec": 11.992110453648914,
+      "control_step_start": 456,
+      "control_step_end": 607,
+      "policy_step_start": 76,
+      "policy_step_end": 101,
+      "phrase": "hovering above drawer idle",
+      "description": "hovering above drawer idle"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 11.992110453648914,
+      "t_end_sec": 14.990138067061142,
+      "control_step_start": 608,
+      "control_step_end": 759,
+      "policy_step_start": 101,
+      "policy_step_end": 126,
+      "phrase": "hovering above drawer idle",
+      "description": "hovering above drawer idle"
+    },
+    {
+      "segment_index": 5,
+      "t_start_sec": 14.990138067061142,
+      "t_end_sec": 15.34516765285996,
+      "control_step_start": 760,
+      "control_step_end": 777,
+      "policy_step_start": 126,
+      "policy_step_end": 129,
+      "phrase": "remaining idle above table",
+      "description": "remaining idle above table"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 3.826429980276134,
+    "vlm_failure_onset_timestamp": "00:03.8",
+    "vlm_failure_onset_frame": 194,
+    "vlm_failure_onset_step": 194,
+    "vlm_coarse_failure_onset_seconds": 5.0,
+    "vlm_failure_window_start_seconds": 4.5,
+    "vlm_failure_window_end_seconds": 5.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the pineapple inside the open drawer, but failed to secure its grip, retracting its gripper empty-handed.",
+    "vlm_failure_reason_summary": "failed grasp attempt on pineapple; gripper missed hold",
+    "vlm_recovery_action": "Re-align the end-effector over the pineapple inside the drawer, close the fingers firmly around it, and lift the pineapple completely out of the drawer.",
+    "vlm_recovery_action_summary": "re-align gripper over pineapple; grasp and remove object",
+    "vlm_prevention_action": "Ensure the gripper is properly centered and lowered around the pineapple before closing the fingers to achieve a secure grip.",
+    "vlm_prevention_action_summary": "center gripper over pineapple; secure full grip",
+    "vlm_justification": "Around 5.0 seconds, the robot's gripper reaches into the open drawer and attempts to grasp the pineapple, but fails to close securely on it. The end-effector then pulls back empty and hovers above the table, resulting in task failure.",
+    "vlm_temporal_justification": "At clip second 01:32 (corresponding to 01:32 in the video timestamp), the robot gripper fingers close completely without securing the pineapple, failing to establish a valid grasp before retracting empty-handed.",
+    "vlm_video_fps": 50.7,
+    "vlm_video_num_frames": 778,
+    "vlm_video_duration_seconds": 15.325443786982248,
+    "vlm_usage": {
+      "input_tokens": 22021,
+      "output_tokens": 373,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 5.0,\n \"failure_window_start_seconds\": 4.5,\n \"failure_window_end_seconds\": 5.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the pineapple inside the open drawer, but failed to secure its grip, retracting its gripper empty-handed.\",\n \"failure_reason_summary\": \"failed grasp attempt on pineapple; gripper missed hold\",\n \"recovery_action\": \"Re-align the end-effector over the pineapple inside the drawer, close the fingers firmly around it, and lift the pineapple completely out of the drawer.\",\n \"recovery_action_summary\": \"re-align gripper over pineapple; grasp and remove object\",\n \"prevention_action\": \"Ensure the gripper is properly centered and lowered around the pineapple before closing the fingers to achieve a secure grip.\",\n \"prevention_action_summary\": \"center gripper over pineapple; secure full grip\",\n \"justification\": \"Around 5.0 seconds, the robot's gripper reaches into the open drawer and attempts to grasp the pineapple, but fails to close securely on it. The end-effector then pulls back empty and hovers above the table, resulting in task failure.\"\n}"
+  },
+  "failure_annotation": {
+    "failure_control_step": 194,
+    "failure_sim_step": 194,
+    "failure_policy_step": 32,
+    "failure_chunk_index": 2,
+    "first_post_failure_policy_step": 33,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-align the end-effector over the pineapple inside the drawer, close the fingers firmly around it, and lift the pineapple completely out of the drawer."
+  }
+}
\ No newline at end of file
diff --git a/pineapple_out_of_drawer/ep048/labels.npz b/pineapple_out_of_drawer/ep048/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..876e7a252eb03f67ad966f74f36c0c2526ca01a8
--- /dev/null
+++ b/pineapple_out_of_drawer/ep048/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f08aa4b227dd55430907f878150309d2ee6ac1551c79f99dd2537ca222c5e59a
+size 3906
diff --git a/pineapple_out_of_drawer/ep049/example.md b/pineapple_out_of_drawer/ep049/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..84d9c8fe81b044336f8c2ad941c7d7fd27a7fbe7
--- /dev/null
+++ b/pineapple_out_of_drawer/ep049/example.md
@@ -0,0 +1,25 @@
+# t007__ep049__pineapple_out_of_drawer
+
+- **suite / variant:** `` / ``
+- **instruction:** Take the pineapple out of the drawer
+- **success:** True
+- **control steps:** 942  (18.7s at 50 Hz)
+- **policy inferences:** 163  (replan_steps=6)
+- **video:** `pineapple_out_of_drawer/ep049/rollout.mp4`
+- **features:** `163` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | approaching pineapple in drawer |
+| 3.0–6.0 | 151–301 | 25–50 | grasping pineapple in drawer |
+| 6.0–9.0 | 302–452 | 50–75 | lifting pineapple out of drawer |
+| 9.0–12.0 | 453–603 | 75–100 | moving pineapple across table |
+| 12.0–15.0 | 604–754 | 100–125 | placing pineapple on table |
+| 15.0–18.0 | 755–905 | 125–150 | releasing pineapple and retracting |
+| 18.0–18.7 | 906–941 | 151–156 | idle after task completion |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/pineapple_out_of_drawer/ep049/labels.json b/pineapple_out_of_drawer/ep049/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..e20c9a5b9bf8e3f746535f76b62eaa1a3896f23e
--- /dev/null
+++ b/pineapple_out_of_drawer/ep049/labels.json
@@ -0,0 +1,112 @@
+{
+  "rollout_id": "t007__ep049__pineapple_out_of_drawer",
+  "scene_variant": "",
+  "task_id": 7,
+  "task_file": "",
+  "instruction": "Take the pineapple out of the drawer",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:37:13+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.001988071570577,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "approaching pineapple in drawer",
+      "description": "approaching pineapple in drawer"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.001988071570577,
+      "t_end_sec": 6.003976143141154,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "grasping pineapple in drawer",
+      "description": "grasping pineapple in drawer"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.003976143141154,
+      "t_end_sec": 9.00596421471173,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "lifting pineapple out of drawer",
+      "description": "lifting pineapple out of drawer"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.00596421471173,
+      "t_end_sec": 12.007952286282308,
+      "control_step_start": 453,
+      "control_step_end": 603,
+      "policy_step_start": 75,
+      "policy_step_end": 100,
+      "phrase": "moving pineapple across table",
+      "description": "moving pineapple across table"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 12.007952286282308,
+      "t_end_sec": 15.009940357852884,
+      "control_step_start": 604,
+      "control_step_end": 754,
+      "policy_step_start": 100,
+      "policy_step_end": 125,
+      "phrase": "placing pineapple on table",
+      "description": "placing pineapple on table"
+    },
+    {
+      "segment_index": 5,
+      "t_start_sec": 15.009940357852884,
+      "t_end_sec": 18.01192842942346,
+      "control_step_start": 755,
+      "control_step_end": 905,
+      "policy_step_start": 125,
+      "policy_step_end": 150,
+      "phrase": "releasing pineapple and retracting",
+      "description": "releasing pineapple and retracting"
+    },
+    {
+      "segment_index": 6,
+      "t_start_sec": 18.01192842942346,
+      "t_end_sec": 18.727634194831015,
+      "control_step_start": 906,
+      "control_step_end": 941,
+      "policy_step_start": 151,
+      "policy_step_end": 156,
+      "phrase": "idle after task completion",
+      "description": "idle after task completion"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/pineapple_out_of_drawer/ep049/labels.npz b/pineapple_out_of_drawer/ep049/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..d67af46c33310d895d7911aaf9d526c140516e65
--- /dev/null
+++ b/pineapple_out_of_drawer/ep049/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7b9ad0c7d5cb8b4d98c0ed78c54ec5a4ab10a2423d617eaa756d1670e791654b
+size 3233
diff --git a/push_pineapple/ep000/example.md b/push_pineapple/ep000/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..a60556f006a9e8c523b847e6887e7f719dc60923
--- /dev/null
+++ b/push_pineapple/ep000/example.md
@@ -0,0 +1,20 @@
+# t005__ep000__push_pineapple
+
+- **suite / variant:** `` / ``
+- **instruction:** Push the pineapple off the stove
+- **success:** True
+- **control steps:** 288  (5.7s at 50 Hz)
+- **policy inferences:** 49  (replan_steps=6)
+- **video:** `push_pineapple/ep000/rollout.mp4`
+- **features:** `49` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | reaching for pineapple |
+| 3.0–5.7 | 151–287 | 25–47 | pushing pineapple off stove |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/push_pineapple/ep000/labels.json b/push_pineapple/ep000/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..81629825296157b81fdabefab70e43ad520f6f93
--- /dev/null
+++ b/push_pineapple/ep000/labels.json
@@ -0,0 +1,57 @@
+{
+  "rollout_id": "t005__ep000__push_pineapple",
+  "scene_variant": "",
+  "task_id": 5,
+  "task_file": "",
+  "instruction": "Push the pineapple off the stove",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:37:21+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00796812749004,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for pineapple",
+      "description": "reaching for pineapple"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00796812749004,
+      "t_end_sec": 5.737051792828685,
+      "control_step_start": 151,
+      "control_step_end": 287,
+      "policy_step_start": 25,
+      "policy_step_end": 47,
+      "phrase": "pushing pineapple off stove",
+      "description": "pushing pineapple off stove"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/push_pineapple/ep000/labels.npz b/push_pineapple/ep000/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..6c41f5788a169e158abe8f063266ee8ec6dc8159
--- /dev/null
+++ b/push_pineapple/ep000/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:690b0364efa5af60bcea75eb1c94889cbde7a977da24fcc1abe3c25ec715d376
+size 3029
diff --git a/push_pineapple/ep001/example.md b/push_pineapple/ep001/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..41c3c696873a67e7f8e425ce3ea3245411577d1c
--- /dev/null
+++ b/push_pineapple/ep001/example.md
@@ -0,0 +1,21 @@
+# t005__ep001__push_pineapple
+
+- **suite / variant:** `` / ``
+- **instruction:** Push the pineapple off the stove
+- **success:** True
+- **control steps:** 361  (7.2s at 50 Hz)
+- **policy inferences:** 62  (replan_steps=6)
+- **video:** `push_pineapple/ep001/rollout.mp4`
+- **features:** `62` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | reaching for pineapple |
+| 3.0–6.0 | 151–301 | 25–50 | pushing pineapple off stove |
+| 6.0–7.2 | 302–360 | 50–60 | retracting arm |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/push_pineapple/ep001/labels.json b/push_pineapple/ep001/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..1fcdcee0e54a0159de27e91831defa8ee9d0d527
--- /dev/null
+++ b/push_pineapple/ep001/labels.json
@@ -0,0 +1,68 @@
+{
+  "rollout_id": "t005__ep001__push_pineapple",
+  "scene_variant": "",
+  "task_id": 5,
+  "task_file": "",
+  "instruction": "Push the pineapple off the stove",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:37:29+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.996031746031746,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for pineapple",
+      "description": "reaching for pineapple"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.996031746031746,
+      "t_end_sec": 5.992063492063492,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "pushing pineapple off stove",
+      "description": "pushing pineapple off stove"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.992063492063492,
+      "t_end_sec": 7.162698412698413,
+      "control_step_start": 302,
+      "control_step_end": 360,
+      "policy_step_start": 50,
+      "policy_step_end": 60,
+      "phrase": "retracting arm",
+      "description": "retracting arm"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/push_pineapple/ep001/labels.npz b/push_pineapple/ep001/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..31528d022784a2af3f2cbe76ffaa8ce5d7391bbe
--- /dev/null
+++ b/push_pineapple/ep001/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a653755edfa41b879e772780d4f1133e390339cf76dc5eacdf674de49801b3f1
+size 3070
diff --git a/push_pineapple/ep002/example.md b/push_pineapple/ep002/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..1474a0539d2d7ab54b37bd658f91c63113bc1df4
--- /dev/null
+++ b/push_pineapple/ep002/example.md
@@ -0,0 +1,21 @@
+# t005__ep002__push_pineapple
+
+- **suite / variant:** `` / ``
+- **instruction:** Push the pineapple off the stove
+- **success:** True
+- **control steps:** 319  (6.3s at 51 Hz)
+- **policy inferences:** 56  (replan_steps=6)
+- **video:** `push_pineapple/ep002/rollout.mp4`
+- **features:** `56` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | pushing pineapple off the stove |
+| 3.0–6.0 | 152–303 | 25–50 | retracting arm back |
+| 6.0–6.3 | 304–318 | 50–53 | holding position |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/push_pineapple/ep002/labels.json b/push_pineapple/ep002/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..2c96c7311ed09b497344bfaf8c272531b0527957
--- /dev/null
+++ b/push_pineapple/ep002/labels.json
@@ -0,0 +1,68 @@
+{
+  "rollout_id": "t005__ep002__push_pineapple",
+  "scene_variant": "",
+  "task_id": 5,
+  "task_file": "",
+  "instruction": "Push the pineapple off the stove",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:37:37+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9980276134122286,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "pushing pineapple off the stove",
+      "description": "pushing pineapple off the stove"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9980276134122286,
+      "t_end_sec": 5.996055226824457,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "retracting arm back",
+      "description": "retracting arm back"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.996055226824457,
+      "t_end_sec": 6.291913214990138,
+      "control_step_start": 304,
+      "control_step_end": 318,
+      "policy_step_start": 50,
+      "policy_step_end": 53,
+      "phrase": "holding position",
+      "description": "holding position"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/push_pineapple/ep002/labels.npz b/push_pineapple/ep002/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..dc8561a24c4e7a088f381b6f9ea70642ab232d16
--- /dev/null
+++ b/push_pineapple/ep002/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1765dfa7781f4ae0a970194d6266c039234499185ece4e3f58a949e4668fd9f5
+size 3082
diff --git a/push_pineapple/ep003/example.md b/push_pineapple/ep003/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..9ffae4710ffb036455296f3a61c923a83defa533
--- /dev/null
+++ b/push_pineapple/ep003/example.md
@@ -0,0 +1,24 @@
+# t005__ep003__push_pineapple
+
+- **suite / variant:** `` / ``
+- **instruction:** Push the pineapple off the stove
+- **success:** True
+- **control steps:** 820  (16.2s at 51 Hz)
+- **policy inferences:** 141  (replan_steps=6)
+- **video:** `push_pineapple/ep003/rollout.mp4`
+- **features:** `141` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching towards pineapple |
+| 3.0–6.0 | 152–303 | 25–50 | contacting pineapple on stove |
+| 6.0–9.0 | 304–455 | 50–75 | pushing pineapple across stove |
+| 9.0–12.0 | 456–607 | 76–101 | pushing pineapple onto table |
+| 12.0–15.0 | 608–759 | 101–126 | retracting arm away |
+| 15.0–16.2 | 760–819 | 126–136 | holding final position |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/push_pineapple/ep003/labels.json b/push_pineapple/ep003/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..b9289ab8ecef5de15103dfa0d494b87fb738dccd
--- /dev/null
+++ b/push_pineapple/ep003/labels.json
@@ -0,0 +1,101 @@
+{
+  "rollout_id": "t005__ep003__push_pineapple",
+  "scene_variant": "",
+  "task_id": 5,
+  "task_file": "",
+  "instruction": "Push the pineapple off the stove",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:37:48+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0039525691699605,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching towards pineapple",
+      "description": "reaching towards pineapple"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0039525691699605,
+      "t_end_sec": 6.007905138339921,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "contacting pineapple on stove",
+      "description": "contacting pineapple on stove"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.007905138339921,
+      "t_end_sec": 9.011857707509881,
+      "control_step_start": 304,
+      "control_step_end": 455,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "pushing pineapple across stove",
+      "description": "pushing pineapple across stove"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.011857707509881,
+      "t_end_sec": 12.015810276679842,
+      "control_step_start": 456,
+      "control_step_end": 607,
+      "policy_step_start": 76,
+      "policy_step_end": 101,
+      "phrase": "pushing pineapple onto table",
+      "description": "pushing pineapple onto table"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 12.015810276679842,
+      "t_end_sec": 15.019762845849803,
+      "control_step_start": 608,
+      "control_step_end": 759,
+      "policy_step_start": 101,
+      "policy_step_end": 126,
+      "phrase": "retracting arm away",
+      "description": "retracting arm away"
+    },
+    {
+      "segment_index": 5,
+      "t_start_sec": 15.019762845849803,
+      "t_end_sec": 16.205533596837945,
+      "control_step_start": 760,
+      "control_step_end": 819,
+      "policy_step_start": 126,
+      "policy_step_end": 136,
+      "phrase": "holding final position",
+      "description": "holding final position"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/push_pineapple/ep003/labels.npz b/push_pineapple/ep003/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..81d8ebbbb2ad72d8888f9ead362e727aec02866a
--- /dev/null
+++ b/push_pineapple/ep003/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9e4f8908158ab8f6336d1bbe8b06d4dba8faac74189c787258bd24a12c30608e
+size 3208
diff --git a/push_pineapple/ep004/example.md b/push_pineapple/ep004/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..14047f1cf08281cfa99b38362b4ac48e8f4038f6
--- /dev/null
+++ b/push_pineapple/ep004/example.md
@@ -0,0 +1,22 @@
+# t005__ep004__push_pineapple
+
+- **suite / variant:** `` / ``
+- **instruction:** Push the pineapple off the stove
+- **success:** True
+- **control steps:** 577  (11.3s at 51 Hz)
+- **policy inferences:** 99  (replan_steps=6)
+- **video:** `push_pineapple/ep004/rollout.mp4`
+- **features:** `99` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–152 | 0–25 | reaching toward pineapple |
+| 3.0–6.0 | 153–305 | 25–50 | pushing pineapple across stove |
+| 6.0–9.0 | 306–458 | 51–76 | pushing pineapple on stove |
+| 9.0–11.3 | 459–576 | 76–96 | retracting robotic arm |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/push_pineapple/ep004/labels.json b/push_pineapple/ep004/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..96f2346619e1b616becad4624fc2d2d1de40be1f
--- /dev/null
+++ b/push_pineapple/ep004/labels.json
@@ -0,0 +1,79 @@
+{
+  "rollout_id": "t005__ep004__push_pineapple",
+  "scene_variant": "",
+  "task_id": 5,
+  "task_file": "",
+  "instruction": "Push the pineapple off the stove",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:37:56+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0,
+      "control_step_start": 0,
+      "control_step_end": 152,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching toward pineapple",
+      "description": "reaching toward pineapple"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0,
+      "t_end_sec": 6.0,
+      "control_step_start": 153,
+      "control_step_end": 305,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "pushing pineapple across stove",
+      "description": "pushing pineapple across stove"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.0,
+      "t_end_sec": 9.0,
+      "control_step_start": 306,
+      "control_step_end": 458,
+      "policy_step_start": 51,
+      "policy_step_end": 76,
+      "phrase": "pushing pineapple on stove",
+      "description": "pushing pineapple on stove"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.0,
+      "t_end_sec": 11.313725490196079,
+      "control_step_start": 459,
+      "control_step_end": 576,
+      "policy_step_start": 76,
+      "policy_step_end": 96,
+      "phrase": "retracting robotic arm",
+      "description": "retracting robotic arm"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/push_pineapple/ep004/labels.npz b/push_pineapple/ep004/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..b4fe0c3db769cb54114090bdc77e82baa6c558ee
--- /dev/null
+++ b/push_pineapple/ep004/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ccde868c39d1a1ee01e173df423a21b0bd79f8a0c910dc982c10a51eb5219a3a
+size 3109
diff --git a/push_pineapple/ep005/example.md b/push_pineapple/ep005/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..3e6ea11c6cee1d332136894f24a3a0dd11a91721
--- /dev/null
+++ b/push_pineapple/ep005/example.md
@@ -0,0 +1,21 @@
+# t005__ep005__push_pineapple
+
+- **suite / variant:** `` / ``
+- **instruction:** Push the pineapple off the stove
+- **success:** True
+- **control steps:** 336  (6.6s at 51 Hz)
+- **policy inferences:** 57  (replan_steps=6)
+- **video:** `push_pineapple/ep005/rollout.mp4`
+- **features:** `57` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | pushing pineapple off the stove |
+| 3.0–6.0 | 152–303 | 25–50 | retracting arm upwards |
+| 6.0–6.6 | 304–335 | 50–55 | idle / no progress |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/push_pineapple/ep005/labels.json b/push_pineapple/ep005/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..57ca172d81e7c4baab652ff7576d1317f6285402
--- /dev/null
+++ b/push_pineapple/ep005/labels.json
@@ -0,0 +1,68 @@
+{
+  "rollout_id": "t005__ep005__push_pineapple",
+  "scene_variant": "",
+  "task_id": 5,
+  "task_file": "",
+  "instruction": "Push the pineapple off the stove",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:38:04+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0039525691699605,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "pushing pineapple off the stove",
+      "description": "pushing pineapple off the stove"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0039525691699605,
+      "t_end_sec": 6.007905138339921,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "retracting arm upwards",
+      "description": "retracting arm upwards"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.007905138339921,
+      "t_end_sec": 6.640316205533597,
+      "control_step_start": 304,
+      "control_step_end": 335,
+      "policy_step_start": 50,
+      "policy_step_end": 55,
+      "phrase": "idle / no progress",
+      "description": "idle / no progress"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/push_pineapple/ep005/labels.npz b/push_pineapple/ep005/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..327f3b27850d732c6c2af70d6bfacb63ce2f5bc9
--- /dev/null
+++ b/push_pineapple/ep005/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3325ac8d24190fddb369cc042cf0dcbd5389405b8abda333efc22ab75410ddaf
+size 3090
diff --git a/push_pineapple/ep006/example.md b/push_pineapple/ep006/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..9f5f598356abaeb09eda1922d6a5ab93b770963c
--- /dev/null
+++ b/push_pineapple/ep006/example.md
@@ -0,0 +1,21 @@
+# t005__ep006__push_pineapple
+
+- **suite / variant:** `` / ``
+- **instruction:** Push the pineapple off the stove
+- **success:** True
+- **control steps:** 323  (6.4s at 50 Hz)
+- **policy inferences:** 56  (replan_steps=6)
+- **video:** `push_pineapple/ep006/rollout.mp4`
+- **features:** `56` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | pushing pineapple off stove |
+| 3.0–6.0 | 152–303 | 25–50 | retracting robot arm |
+| 6.0–6.4 | 304–322 | 50–53 | remaining stationary |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/push_pineapple/ep006/labels.json b/push_pineapple/ep006/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..7f5da4d7eaaa7f0d036c4d03b54da91f51346f84
--- /dev/null
+++ b/push_pineapple/ep006/labels.json
@@ -0,0 +1,68 @@
+{
+  "rollout_id": "t005__ep006__push_pineapple",
+  "scene_variant": "",
+  "task_id": 5,
+  "task_file": "",
+  "instruction": "Push the pineapple off the stove",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:38:12+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00990099009901,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "pushing pineapple off stove",
+      "description": "pushing pineapple off stove"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00990099009901,
+      "t_end_sec": 6.01980198019802,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "retracting robot arm",
+      "description": "retracting robot arm"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01980198019802,
+      "t_end_sec": 6.396039603960396,
+      "control_step_start": 304,
+      "control_step_end": 322,
+      "policy_step_start": 50,
+      "policy_step_end": 53,
+      "phrase": "remaining stationary",
+      "description": "remaining stationary"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/push_pineapple/ep006/labels.npz b/push_pineapple/ep006/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..41816dcfca888d68e746d5e42d95cd075f6dab9c
--- /dev/null
+++ b/push_pineapple/ep006/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4cf2e413f4373a1e32f8fb534dd25b49f50b6839fdbd3293206f752bc0a79a04
+size 3086
diff --git a/push_pineapple/ep007/example.md b/push_pineapple/ep007/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..c164277f3f96387bdd3ed7979c6ac4c4c96626df
--- /dev/null
+++ b/push_pineapple/ep007/example.md
@@ -0,0 +1,21 @@
+# t005__ep007__push_pineapple
+
+- **suite / variant:** `` / ``
+- **instruction:** Push the pineapple off the stove
+- **success:** True
+- **control steps:** 387  (7.6s at 51 Hz)
+- **policy inferences:** 66  (replan_steps=6)
+- **video:** `push_pineapple/ep007/rollout.mp4`
+- **features:** `66` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching for pineapple |
+| 3.0–6.0 | 152–303 | 25–50 | pushing pineapple off stove |
+| 6.0–7.6 | 304–386 | 50–64 | retracting robot arm |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/push_pineapple/ep007/labels.json b/push_pineapple/ep007/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..78df1bdf754e97414dffa54aee144dbc6a10f88e
--- /dev/null
+++ b/push_pineapple/ep007/labels.json
@@ -0,0 +1,68 @@
+{
+  "rollout_id": "t005__ep007__push_pineapple",
+  "scene_variant": "",
+  "task_id": 5,
+  "task_file": "",
+  "instruction": "Push the pineapple off the stove",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:38:18+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9980276134122286,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for pineapple",
+      "description": "reaching for pineapple"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9980276134122286,
+      "t_end_sec": 5.996055226824457,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "pushing pineapple off stove",
+      "description": "pushing pineapple off stove"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.996055226824457,
+      "t_end_sec": 7.633136094674556,
+      "control_step_start": 304,
+      "control_step_end": 386,
+      "policy_step_start": 50,
+      "policy_step_end": 64,
+      "phrase": "retracting robot arm",
+      "description": "retracting robot arm"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/push_pineapple/ep007/labels.npz b/push_pineapple/ep007/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..0eb9c2660ddc08164d39f2c3138d06bf5cfddb66
--- /dev/null
+++ b/push_pineapple/ep007/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a958fe6cc18fac90d51da92f27701054f3ca05e466974baaafd559ba48c1a281
+size 3078
diff --git a/push_pineapple/ep008/example.md b/push_pineapple/ep008/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..2fe119ecd93fa28ae9a400b3ee4648ed5493126b
--- /dev/null
+++ b/push_pineapple/ep008/example.md
@@ -0,0 +1,21 @@
+# t005__ep008__push_pineapple
+
+- **suite / variant:** `` / ``
+- **instruction:** Push the pineapple off the stove
+- **success:** True
+- **control steps:** 411  (8.2s at 50 Hz)
+- **policy inferences:** 71  (replan_steps=6)
+- **video:** `push_pineapple/ep008/rollout.mp4`
+- **features:** `71` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–149 | 0–24 | reaching for pineapple |
+| 3.0–6.0 | 150–299 | 25–49 | pushing pineapple off stove |
+| 6.0–8.2 | 300–410 | 50–68 | retracting robotic arm |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/push_pineapple/ep008/labels.json b/push_pineapple/ep008/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..54232f43ae992b85ca11bcd22ca218055c647a32
--- /dev/null
+++ b/push_pineapple/ep008/labels.json
@@ -0,0 +1,68 @@
+{
+  "rollout_id": "t005__ep008__push_pineapple",
+  "scene_variant": "",
+  "task_id": 5,
+  "task_file": "",
+  "instruction": "Push the pineapple off the stove",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:38:26+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9940119760479043,
+      "control_step_start": 0,
+      "control_step_end": 149,
+      "policy_step_start": 0,
+      "policy_step_end": 24,
+      "phrase": "reaching for pineapple",
+      "description": "reaching for pineapple"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9940119760479043,
+      "t_end_sec": 5.9880239520958085,
+      "control_step_start": 150,
+      "control_step_end": 299,
+      "policy_step_start": 25,
+      "policy_step_end": 49,
+      "phrase": "pushing pineapple off stove",
+      "description": "pushing pineapple off stove"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.9880239520958085,
+      "t_end_sec": 8.203592814371257,
+      "control_step_start": 300,
+      "control_step_end": 410,
+      "policy_step_start": 50,
+      "policy_step_end": 68,
+      "phrase": "retracting robotic arm",
+      "description": "retracting robotic arm"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/push_pineapple/ep008/labels.npz b/push_pineapple/ep008/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..871bdc364c207d8a4bda160811a4e7a4b9857e8b
--- /dev/null
+++ b/push_pineapple/ep008/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:359e5ca04c46348893e1205aab56cd84654c19fc11efd39abe8eb525dbeed706
+size 3080
diff --git a/push_pineapple/ep009/example.md b/push_pineapple/ep009/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..2c08992682490fb2a696646ed690735bb0435dfc
--- /dev/null
+++ b/push_pineapple/ep009/example.md
@@ -0,0 +1,21 @@
+# t005__ep009__push_pineapple
+
+- **suite / variant:** `` / ``
+- **instruction:** Push the pineapple off the stove
+- **success:** True
+- **control steps:** 416  (8.2s at 50 Hz)
+- **policy inferences:** 72  (replan_steps=6)
+- **video:** `push_pineapple/ep009/rollout.mp4`
+- **features:** `72` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching for pineapple |
+| 3.0–6.0 | 152–303 | 25–50 | pushing pineapple off stove |
+| 6.0–8.2 | 304–415 | 50–69 | idle / holding position |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/push_pineapple/ep009/labels.json b/push_pineapple/ep009/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..f20c4802fedbaced00dbecfc944fc339dd066eba
--- /dev/null
+++ b/push_pineapple/ep009/labels.json
@@ -0,0 +1,68 @@
+{
+  "rollout_id": "t005__ep009__push_pineapple",
+  "scene_variant": "",
+  "task_id": 5,
+  "task_file": "",
+  "instruction": "Push the pineapple off the stove",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:38:34+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00990099009901,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for pineapple",
+      "description": "reaching for pineapple"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00990099009901,
+      "t_end_sec": 6.01980198019802,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "pushing pineapple off stove",
+      "description": "pushing pineapple off stove"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01980198019802,
+      "t_end_sec": 8.237623762376238,
+      "control_step_start": 304,
+      "control_step_end": 415,
+      "policy_step_start": 50,
+      "policy_step_end": 69,
+      "phrase": "idle / holding position",
+      "description": "idle / holding position"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/push_pineapple/ep009/labels.npz b/push_pineapple/ep009/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..053e5eb8ac4c68c32435a4d1f206d66960f1d62a
--- /dev/null
+++ b/push_pineapple/ep009/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1a1d029a3dc823dd04ea263a64c759fc3c3ee515b60555df51a8b3149640f658
+size 3075
diff --git a/push_pineapple/ep010/example.md b/push_pineapple/ep010/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..7039df8a7a2069e334f2322257678a9b972bbb45
--- /dev/null
+++ b/push_pineapple/ep010/example.md
@@ -0,0 +1,23 @@
+# t005__ep010__push_pineapple
+
+- **suite / variant:** `` / ``
+- **instruction:** Push the pineapple off the stove
+- **success:** True
+- **control steps:** 639  (12.6s at 51 Hz)
+- **policy inferences:** 110  (replan_steps=6)
+- **video:** `push_pineapple/ep010/rollout.mp4`
+- **features:** `110` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching towards pineapple |
+| 3.0–6.0 | 152–303 | 25–50 | pushing pineapple across stove |
+| 6.0–9.0 | 304–455 | 50–75 | pushing pineapple off stove |
+| 9.0–12.0 | 456–607 | 76–101 | retracting arm upwards |
+| 12.0–12.6 | 608–638 | 101–106 | idle in home position |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/push_pineapple/ep010/labels.json b/push_pineapple/ep010/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..a3cbfc63404fcab3f6d3f11f1bb9b993295c6675
--- /dev/null
+++ b/push_pineapple/ep010/labels.json
@@ -0,0 +1,90 @@
+{
+  "rollout_id": "t005__ep010__push_pineapple",
+  "scene_variant": "",
+  "task_id": 5,
+  "task_file": "",
+  "instruction": "Push the pineapple off the stove",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:38:43+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9921259842519685,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching towards pineapple",
+      "description": "reaching towards pineapple"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9921259842519685,
+      "t_end_sec": 5.984251968503937,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "pushing pineapple across stove",
+      "description": "pushing pineapple across stove"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.984251968503937,
+      "t_end_sec": 8.976377952755906,
+      "control_step_start": 304,
+      "control_step_end": 455,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "pushing pineapple off stove",
+      "description": "pushing pineapple off stove"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.976377952755906,
+      "t_end_sec": 11.968503937007874,
+      "control_step_start": 456,
+      "control_step_end": 607,
+      "policy_step_start": 76,
+      "policy_step_end": 101,
+      "phrase": "retracting arm upwards",
+      "description": "retracting arm upwards"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 11.968503937007874,
+      "t_end_sec": 12.578740157480315,
+      "control_step_start": 608,
+      "control_step_end": 638,
+      "policy_step_start": 101,
+      "policy_step_end": 106,
+      "phrase": "idle in home position",
+      "description": "idle in home position"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/push_pineapple/ep010/labels.npz b/push_pineapple/ep010/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..d61c2cafe86639a41d0a4f704b6e0eedcb469194
--- /dev/null
+++ b/push_pineapple/ep010/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b7e4f8f6590829bb84b6625bc980eaec4a5cc5116b7695f199d9058e65044e04
+size 3158
diff --git a/push_pineapple/ep011/example.md b/push_pineapple/ep011/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..371253c18bcfff08b47064a23b8bc61946cfb213
--- /dev/null
+++ b/push_pineapple/ep011/example.md
@@ -0,0 +1,21 @@
+# t005__ep011__push_pineapple
+
+- **suite / variant:** `` / ``
+- **instruction:** Push the pineapple off the stove
+- **success:** True
+- **control steps:** 392  (7.8s at 50 Hz)
+- **policy inferences:** 68  (replan_steps=6)
+- **video:** `push_pineapple/ep011/rollout.mp4`
+- **features:** `68` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | reaching for the pineapple |
+| 3.0–6.0 | 151–301 | 25–50 | pushing pineapple off stove |
+| 6.0–7.8 | 302–391 | 50–65 | retracting robot arm |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/push_pineapple/ep011/labels.json b/push_pineapple/ep011/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..33a2e972554a7d76513db74c8c0b3d958f3e3397
--- /dev/null
+++ b/push_pineapple/ep011/labels.json
@@ -0,0 +1,68 @@
+{
+  "rollout_id": "t005__ep011__push_pineapple",
+  "scene_variant": "",
+  "task_id": 5,
+  "task_file": "",
+  "instruction": "Push the pineapple off the stove",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:38:50+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.996031746031746,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for the pineapple",
+      "description": "reaching for the pineapple"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.996031746031746,
+      "t_end_sec": 5.992063492063492,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "pushing pineapple off stove",
+      "description": "pushing pineapple off stove"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.992063492063492,
+      "t_end_sec": 7.777777777777778,
+      "control_step_start": 302,
+      "control_step_end": 391,
+      "policy_step_start": 50,
+      "policy_step_end": 65,
+      "phrase": "retracting robot arm",
+      "description": "retracting robot arm"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/push_pineapple/ep011/labels.npz b/push_pineapple/ep011/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..bf9142e98c9b086c65b9832c54f1d170465617f7
--- /dev/null
+++ b/push_pineapple/ep011/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:de09e9692c1cf855d65eead6e97ae9104e5b2e256d6ed1cf08fd29cb7eb3ca6b
+size 3082
diff --git a/push_pineapple/ep012/example.md b/push_pineapple/ep012/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..3adec58799f5a26cb8135db8e11cd9a39f79fe83
--- /dev/null
+++ b/push_pineapple/ep012/example.md
@@ -0,0 +1,21 @@
+# t005__ep012__push_pineapple
+
+- **suite / variant:** `` / ``
+- **instruction:** Push the pineapple off the stove
+- **success:** True
+- **control steps:** 445  (8.7s at 51 Hz)
+- **policy inferences:** 78  (replan_steps=6)
+- **video:** `push_pineapple/ep012/rollout.mp4`
+- **features:** `78` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–152 | 0–25 | reaching toward pineapple |
+| 3.0–6.0 | 153–305 | 25–50 | pushing pineapple off stove |
+| 6.0–8.7 | 306–444 | 51–74 | retracting robotic arm |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/push_pineapple/ep012/labels.json b/push_pineapple/ep012/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..6cdac6a51ae65dc793dbdd0e14715b421d230ea3
--- /dev/null
+++ b/push_pineapple/ep012/labels.json
@@ -0,0 +1,68 @@
+{
+  "rollout_id": "t005__ep012__push_pineapple",
+  "scene_variant": "",
+  "task_id": 5,
+  "task_file": "",
+  "instruction": "Push the pineapple off the stove",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:38:57+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.005893909626719,
+      "control_step_start": 0,
+      "control_step_end": 152,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching toward pineapple",
+      "description": "reaching toward pineapple"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.005893909626719,
+      "t_end_sec": 6.011787819253438,
+      "control_step_start": 153,
+      "control_step_end": 305,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "pushing pineapple off stove",
+      "description": "pushing pineapple off stove"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.011787819253438,
+      "t_end_sec": 8.7426326129666,
+      "control_step_start": 306,
+      "control_step_end": 444,
+      "policy_step_start": 51,
+      "policy_step_end": 74,
+      "phrase": "retracting robotic arm",
+      "description": "retracting robotic arm"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/push_pineapple/ep012/labels.npz b/push_pineapple/ep012/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..d8dda9b90adbd499b783d6013e3790866b7ad872
--- /dev/null
+++ b/push_pineapple/ep012/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:28e881a8e43ed5e81d123837c42a7cbe9696d4d7a0a8c7c5641aca7bb3c75985
+size 3085
diff --git a/push_pineapple/ep013/example.md b/push_pineapple/ep013/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..048b1d01f38115609bf48f7162cc2f3bdfdd840d
--- /dev/null
+++ b/push_pineapple/ep013/example.md
@@ -0,0 +1,21 @@
+# t005__ep013__push_pineapple
+
+- **suite / variant:** `` / ``
+- **instruction:** Push the pineapple off the stove
+- **success:** True
+- **control steps:** 318  (6.3s at 50 Hz)
+- **policy inferences:** 55  (replan_steps=6)
+- **video:** `push_pineapple/ep013/rollout.mp4`
+- **features:** `55` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | reaching for the pineapple |
+| 3.0–6.0 | 151–301 | 25–50 | pushing pineapple off stove |
+| 6.0–6.3 | 302–317 | 50–52 | retracting robotic arm |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/push_pineapple/ep013/labels.json b/push_pineapple/ep013/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..86dd2ae33c4cb1ce1d1104270bb4bf2414426bce
--- /dev/null
+++ b/push_pineapple/ep013/labels.json
@@ -0,0 +1,68 @@
+{
+  "rollout_id": "t005__ep013__push_pineapple",
+  "scene_variant": "",
+  "task_id": 5,
+  "task_file": "",
+  "instruction": "Push the pineapple off the stove",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:39:04+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.996031746031746,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for the pineapple",
+      "description": "reaching for the pineapple"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.996031746031746,
+      "t_end_sec": 5.992063492063492,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "pushing pineapple off stove",
+      "description": "pushing pineapple off stove"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.992063492063492,
+      "t_end_sec": 6.309523809523809,
+      "control_step_start": 302,
+      "control_step_end": 317,
+      "policy_step_start": 50,
+      "policy_step_end": 52,
+      "phrase": "retracting robotic arm",
+      "description": "retracting robotic arm"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/push_pineapple/ep013/labels.npz b/push_pineapple/ep013/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..a92090df32e5284bf4b9088c201e30a5549a7121
--- /dev/null
+++ b/push_pineapple/ep013/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9f71c9cbda5eea6807bde1f1347721308dafac64731658fa6ccdf16ce18b4133
+size 3084
diff --git a/push_pineapple/ep014/example.md b/push_pineapple/ep014/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..33ac806d0a8c27300e16fb2f99fc3148af5f6445
--- /dev/null
+++ b/push_pineapple/ep014/example.md
@@ -0,0 +1,21 @@
+# t005__ep014__push_pineapple
+
+- **suite / variant:** `` / ``
+- **instruction:** Push the pineapple off the stove
+- **success:** True
+- **control steps:** 334  (6.6s at 50 Hz)
+- **policy inferences:** 58  (replan_steps=6)
+- **video:** `push_pineapple/ep014/rollout.mp4`
+- **features:** `58` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching for pineapple |
+| 3.0–6.0 | 152–303 | 25–50 | pushing pineapple off stove |
+| 6.0–6.6 | 304–333 | 50–55 | retracting arm |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/push_pineapple/ep014/labels.json b/push_pineapple/ep014/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..a254d56fb229fac8e4ea7ca0fa4ac887c3a05b54
--- /dev/null
+++ b/push_pineapple/ep014/labels.json
@@ -0,0 +1,68 @@
+{
+  "rollout_id": "t005__ep014__push_pineapple",
+  "scene_variant": "",
+  "task_id": 5,
+  "task_file": "",
+  "instruction": "Push the pineapple off the stove",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:39:12+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00990099009901,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for pineapple",
+      "description": "reaching for pineapple"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00990099009901,
+      "t_end_sec": 6.01980198019802,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "pushing pineapple off stove",
+      "description": "pushing pineapple off stove"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01980198019802,
+      "t_end_sec": 6.6138613861386135,
+      "control_step_start": 304,
+      "control_step_end": 333,
+      "policy_step_start": 50,
+      "policy_step_end": 55,
+      "phrase": "retracting arm",
+      "description": "retracting arm"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/push_pineapple/ep014/labels.npz b/push_pineapple/ep014/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..cdc947d3d8e90d78bcfa13070b57bfb848eaff90
--- /dev/null
+++ b/push_pineapple/ep014/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:01afa150059e57f88e8fe9e80fcc183fc79ae12656799de54a24f6f9f439e839
+size 3070
diff --git a/push_pineapple/ep015/example.md b/push_pineapple/ep015/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..f3df2a1f950c7263f710e28125d80d6483fd8653
--- /dev/null
+++ b/push_pineapple/ep015/example.md
@@ -0,0 +1,22 @@
+# t005__ep015__push_pineapple
+
+- **suite / variant:** `` / ``
+- **instruction:** Push the pineapple off the stove
+- **success:** True
+- **control steps:** 556  (11.0s at 50 Hz)
+- **policy inferences:** 96  (replan_steps=6)
+- **video:** `push_pineapple/ep015/rollout.mp4`
+- **features:** `96` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | reaching for pineapple |
+| 3.0–6.0 | 151–301 | 25–50 | pushing pineapple across stove |
+| 6.0–9.0 | 302–452 | 50–75 | pushing pineapple off stove |
+| 9.0–11.0 | 453–555 | 75–92 | holding position above pineapple |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/push_pineapple/ep015/labels.json b/push_pineapple/ep015/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..73a84939cec3648bc6281d1900da4d4d7e64d2f4
--- /dev/null
+++ b/push_pineapple/ep015/labels.json
@@ -0,0 +1,79 @@
+{
+  "rollout_id": "t005__ep015__push_pineapple",
+  "scene_variant": "",
+  "task_id": 5,
+  "task_file": "",
+  "instruction": "Push the pineapple off the stove",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:39:22+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.996031746031746,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for pineapple",
+      "description": "reaching for pineapple"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.996031746031746,
+      "t_end_sec": 5.992063492063492,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "pushing pineapple across stove",
+      "description": "pushing pineapple across stove"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.992063492063492,
+      "t_end_sec": 8.988095238095239,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "pushing pineapple off stove",
+      "description": "pushing pineapple off stove"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.988095238095239,
+      "t_end_sec": 11.031746031746032,
+      "control_step_start": 453,
+      "control_step_end": 555,
+      "policy_step_start": 75,
+      "policy_step_end": 92,
+      "phrase": "holding position above pineapple",
+      "description": "holding position above pineapple"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/push_pineapple/ep015/labels.npz b/push_pineapple/ep015/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..1ccb9fe87c7d8a6d15ce9d76d98ebe6902f11b40
--- /dev/null
+++ b/push_pineapple/ep015/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7786b1594f0e3369db83177af6b7ce4fc2c6a1bf52a558513340a0d5fa62f19a
+size 3113
diff --git a/push_pineapple/ep016/example.md b/push_pineapple/ep016/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..9c113ae88c57a74da27fcd57ad7a1ad6500b2e53
--- /dev/null
+++ b/push_pineapple/ep016/example.md
@@ -0,0 +1,20 @@
+# t005__ep016__push_pineapple
+
+- **suite / variant:** `` / ``
+- **instruction:** Push the pineapple off the stove
+- **success:** True
+- **control steps:** 249  (4.9s at 50 Hz)
+- **policy inferences:** 43  (replan_steps=6)
+- **video:** `push_pineapple/ep016/rollout.mp4`
+- **features:** `43` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | reaching towards the pineapple |
+| 3.0–4.9 | 151–248 | 25–41 | pushing pineapple off stove |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/push_pineapple/ep016/labels.json b/push_pineapple/ep016/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..9fc04c8e67d1c14cbcaca2ae1894fb8ccabbd88d
--- /dev/null
+++ b/push_pineapple/ep016/labels.json
@@ -0,0 +1,57 @@
+{
+  "rollout_id": "t005__ep016__push_pineapple",
+  "scene_variant": "",
+  "task_id": 5,
+  "task_file": "",
+  "instruction": "Push the pineapple off the stove",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:39:29+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.996031746031746,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching towards the pineapple",
+      "description": "reaching towards the pineapple"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.996031746031746,
+      "t_end_sec": 4.940476190476191,
+      "control_step_start": 151,
+      "control_step_end": 248,
+      "policy_step_start": 25,
+      "policy_step_end": 41,
+      "phrase": "pushing pineapple off stove",
+      "description": "pushing pineapple off stove"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/push_pineapple/ep016/labels.npz b/push_pineapple/ep016/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..b8b62b497063c0c1cd728eccad1458cc9c82e699
--- /dev/null
+++ b/push_pineapple/ep016/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d6396c7ff9876e32b58afa2499d31dcd39d848864f7538f6e76848c15f432713
+size 3040
diff --git a/push_pineapple/ep017/example.md b/push_pineapple/ep017/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..79bdfb8d87c436e883dadf9ee1e88109112acb25
--- /dev/null
+++ b/push_pineapple/ep017/example.md
@@ -0,0 +1,21 @@
+# t005__ep017__push_pineapple
+
+- **suite / variant:** `` / ``
+- **instruction:** Push the pineapple off the stove
+- **success:** True
+- **control steps:** 401  (8.0s at 50 Hz)
+- **policy inferences:** 69  (replan_steps=6)
+- **video:** `push_pineapple/ep017/rollout.mp4`
+- **features:** `69` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | reaching towards the pineapple |
+| 3.0–6.0 | 151–301 | 25–50 | pushing pineapple off the stove |
+| 6.0–8.0 | 302–400 | 50–66 | retracting robot arm |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/push_pineapple/ep017/labels.json b/push_pineapple/ep017/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..9d074f60a60087ffaecbe0f514434dcf29f57e4f
--- /dev/null
+++ b/push_pineapple/ep017/labels.json
@@ -0,0 +1,68 @@
+{
+  "rollout_id": "t005__ep017__push_pineapple",
+  "scene_variant": "",
+  "task_id": 5,
+  "task_file": "",
+  "instruction": "Push the pineapple off the stove",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:39:36+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00796812749004,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching towards the pineapple",
+      "description": "reaching towards the pineapple"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00796812749004,
+      "t_end_sec": 6.01593625498008,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "pushing pineapple off the stove",
+      "description": "pushing pineapple off the stove"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01593625498008,
+      "t_end_sec": 7.988047808764939,
+      "control_step_start": 302,
+      "control_step_end": 400,
+      "policy_step_start": 50,
+      "policy_step_end": 66,
+      "phrase": "retracting robot arm",
+      "description": "retracting robot arm"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/push_pineapple/ep017/labels.npz b/push_pineapple/ep017/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..356efe9af14777186f0ebef9265eb9220da28103
--- /dev/null
+++ b/push_pineapple/ep017/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b812d275431bd09519a70138dedbae614754bea44a606057ead42495da276b92
+size 3094
diff --git a/push_pineapple/ep018/example.md b/push_pineapple/ep018/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..ae4540cecf91f280a8d943b0542843805bf93639
--- /dev/null
+++ b/push_pineapple/ep018/example.md
@@ -0,0 +1,20 @@
+# t005__ep018__push_pineapple
+
+- **suite / variant:** `` / ``
+- **instruction:** Push the pineapple off the stove
+- **success:** True
+- **control steps:** 243  (4.8s at 50 Hz)
+- **policy inferences:** 42  (replan_steps=5)
+- **video:** `push_pineapple/ep018/rollout.mp4`
+- **features:** `42` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–30 | pushing pineapple off stove |
+| 3.0–4.8 | 151–242 | 30–48 | retracting robot arm |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/push_pineapple/ep018/labels.json b/push_pineapple/ep018/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..109bfdb09f3382418dd83d6a23c30e517a19bc08
--- /dev/null
+++ b/push_pineapple/ep018/labels.json
@@ -0,0 +1,57 @@
+{
+  "rollout_id": "t005__ep018__push_pineapple",
+  "scene_variant": "",
+  "task_id": 5,
+  "task_file": "",
+  "instruction": "Push the pineapple off the stove",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:39:43+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00796812749004,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 30,
+      "phrase": "pushing pineapple off stove",
+      "description": "pushing pineapple off stove"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00796812749004,
+      "t_end_sec": 4.840637450199203,
+      "control_step_start": 151,
+      "control_step_end": 242,
+      "policy_step_start": 30,
+      "policy_step_end": 48,
+      "phrase": "retracting robot arm",
+      "description": "retracting robot arm"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/push_pineapple/ep018/labels.npz b/push_pineapple/ep018/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..3a81623b512b674fe72a6a01c297fe9dd17c8789
--- /dev/null
+++ b/push_pineapple/ep018/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fa0ac84ae6ab1996855d85e4d996d95da1e1c09c48aa2172ca0eb5ae70a45b29
+size 3037
diff --git a/push_pineapple/ep019/example.md b/push_pineapple/ep019/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..0b852d500e222a48df5eeafd6da0bb16d0ffc25f
--- /dev/null
+++ b/push_pineapple/ep019/example.md
@@ -0,0 +1,20 @@
+# t005__ep019__push_pineapple
+
+- **suite / variant:** `` / ``
+- **instruction:** Push the pineapple off the stove
+- **success:** True
+- **control steps:** 271  (5.4s at 51 Hz)
+- **policy inferences:** 47  (replan_steps=6)
+- **video:** `push_pineapple/ep019/rollout.mp4`
+- **features:** `47` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | pushing pineapple off stove |
+| 3.0–5.4 | 152–270 | 25–45 | retracting robot arm |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/push_pineapple/ep019/labels.json b/push_pineapple/ep019/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..7b8477f99ac59aacc91539107e86d206e24c48d3
--- /dev/null
+++ b/push_pineapple/ep019/labels.json
@@ -0,0 +1,57 @@
+{
+  "rollout_id": "t005__ep019__push_pineapple",
+  "scene_variant": "",
+  "task_id": 5,
+  "task_file": "",
+  "instruction": "Push the pineapple off the stove",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:39:50+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0039525691699605,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "pushing pineapple off stove",
+      "description": "pushing pineapple off stove"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0039525691699605,
+      "t_end_sec": 5.355731225296442,
+      "control_step_start": 152,
+      "control_step_end": 270,
+      "policy_step_start": 25,
+      "policy_step_end": 45,
+      "phrase": "retracting robot arm",
+      "description": "retracting robot arm"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/push_pineapple/ep019/labels.npz b/push_pineapple/ep019/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..56907e62a60fa963fa52a3909535b8db40d82905
--- /dev/null
+++ b/push_pineapple/ep019/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c3af1d004fae4849a793bb066482e51637be9ca44235ab204655c007f977d72c
+size 3038
diff --git a/push_pineapple/ep020/example.md b/push_pineapple/ep020/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..bf8b7d360d0113219d023ee172c10853620d240f
--- /dev/null
+++ b/push_pineapple/ep020/example.md
@@ -0,0 +1,33 @@
+# t005__ep020__push_pineapple
+
+- **suite / variant:** `` / ``
+- **instruction:** Push the pineapple off the stove
+- **success:** False
+- **control steps:** 437  (8.6s at 51 Hz)
+- **policy inferences:** 75  (replan_steps=6)
+- **video:** `push_pineapple/ep020/rollout.mp4`
+- **features:** `75` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–152 | 0–25 | reaching toward pineapple on stove |
+| 3.0–6.0 | 153–305 | 25–50 | pushing pineapple and lifting arm |
+| 6.0–8.6 | 306–436 | 51–72 | retracting arm and staying idle |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 3.8627450980392157 s (00:03.9)
+- **frame / control_step:** 197 (policy_step=32, chunk_index=5)
+- **coarse onset:** 3.8 s (refined=True)
+- **mode:** object_displacement
+- **reason:** The robot pushed the pineapple toward the edge of the stove but retracted its arm prematurely, leaving the pineapple's leaves still touching the stove plate.
+  Summary: insufficient push distance; object still touching stove
+- **recovery:** Approach the pineapple again and push it further forward onto the table so that it completely clears the stove surface.
+  Summary: re-engage pineapple; push clear of stove
+- **prevention:** Extend the pushing motion further forward to guarantee the pineapple is fully off the stove before retracting the end-effector.
+  Summary: push pineapple further; ensure total clearance before retracting
+- **justification:** Around 3.8 seconds into the video (frame 00:04), the end-effector stops pushing and begins lifting away while the pineapple is still partially resting on the stove edge. Because the robot retracts to home position for the remainder of the episode without finishing the push, the goal predicate remains unsatisfied.
diff --git a/push_pineapple/ep020/labels.json b/push_pineapple/ep020/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..1f9da3b4023b42b06c490dca659aa956c6520388
--- /dev/null
+++ b/push_pineapple/ep020/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t005__ep020__push_pineapple",
+  "scene_variant": "",
+  "task_id": 5,
+  "task_file": "",
+  "instruction": "Push the pineapple off the stove",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:40:30+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0,
+      "control_step_start": 0,
+      "control_step_end": 152,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching toward pineapple on stove",
+      "description": "reaching toward pineapple on stove"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0,
+      "t_end_sec": 6.0,
+      "control_step_start": 153,
+      "control_step_end": 305,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "pushing pineapple and lifting arm",
+      "description": "pushing pineapple and lifting arm"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.0,
+      "t_end_sec": 8.568627450980392,
+      "control_step_start": 306,
+      "control_step_end": 436,
+      "policy_step_start": 51,
+      "policy_step_end": 72,
+      "phrase": "retracting arm and staying idle",
+      "description": "retracting arm and staying idle"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "object_displacement",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 3.8627450980392157,
+    "vlm_failure_onset_timestamp": "00:03.9",
+    "vlm_failure_onset_frame": 197,
+    "vlm_failure_onset_step": 197,
+    "vlm_coarse_failure_onset_seconds": 3.8,
+    "vlm_failure_window_start_seconds": 3.3,
+    "vlm_failure_window_end_seconds": 4.3,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot pushed the pineapple toward the edge of the stove but retracted its arm prematurely, leaving the pineapple's leaves still touching the stove plate.",
+    "vlm_failure_reason_summary": "insufficient push distance; object still touching stove",
+    "vlm_recovery_action": "Approach the pineapple again and push it further forward onto the table so that it completely clears the stove surface.",
+    "vlm_recovery_action_summary": "re-engage pineapple; push clear of stove",
+    "vlm_prevention_action": "Extend the pushing motion further forward to guarantee the pineapple is fully off the stove before retracting the end-effector.",
+    "vlm_prevention_action_summary": "push pineapple further; ensure total clearance before retracting",
+    "vlm_justification": "Around 3.8 seconds into the video (frame 00:04), the end-effector stops pushing and begins lifting away while the pineapple is still partially resting on the stove edge. Because the robot retracts to home position for the remainder of the episode without finishing the push, the goal predicate remains unsatisfied.",
+    "vlm_temporal_justification": "At 02:36 (frame 196 of the original rollout), the robot arm stops pushing the pineapple to the right and begins retracting upwards prematurely, leaving the pineapple's leaves still resting on the edge of the stove plate.",
+    "vlm_video_fps": 51.0,
+    "vlm_video_num_frames": 437,
+    "vlm_video_duration_seconds": 8.549019607843137,
+    "vlm_usage": {
+      "input_tokens": 21762,
+      "output_tokens": 390,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"object_displacement\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 3.8,\n \"failure_window_start_seconds\": 3.3,\n \"failure_window_end_seconds\": 4.3,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot pushed the pineapple toward the edge of the stove but retracted its arm prematurely, leaving the pineapple's leaves still touching the stove plate.\",\n \"failure_reason_summary\": \"insufficient push distance; object still touching stove\",\n \"recovery_action\": \"Approach the pineapple again and push it further forward onto the table so that it completely clears the stove surface.\",\n \"recovery_action_summary\": \"re-engage pineapple; push clear of stove\",\n \"prevention_action\": \"Extend the pushing motion further forward to guarantee the pineapple is fully off the stove before retracting the end-effector.\",\n \"prevention_action_summary\": \"push pineapple further; ensure total clearance before retracting\",\n \"justification\": \"Around 3.8 seconds into the video (frame 00:04), the end-effector stops pushing and begins lifting away while the pineapple is still partially resting on the stove edge. Because the robot retracts to home position for the remainder of the episode without finishing the push, the goal predicate remains unsatisfied.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 197,
+    "failure_sim_step": 197,
+    "failure_policy_step": 32,
+    "failure_chunk_index": 5,
+    "first_post_failure_policy_step": 33,
+    "failure_type": "object_displacement",
+    "correction_action": "Approach the pineapple again and push it further forward onto the table so that it completely clears the stove surface."
+  }
+}
\ No newline at end of file
diff --git a/push_pineapple/ep020/labels.npz b/push_pineapple/ep020/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..16afd1093acbcb387309decc0a03ce92d6f2114b
--- /dev/null
+++ b/push_pineapple/ep020/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6c4f575567a142c656db995cea1234eeefbdb226dd534fb6832e770eeeac2833
+size 3812
diff --git a/push_pineapple/ep021/example.md b/push_pineapple/ep021/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..8e6c4b8196351ccb38ae582632397ce39bcdce58
--- /dev/null
+++ b/push_pineapple/ep021/example.md
@@ -0,0 +1,35 @@
+# t005__ep021__push_pineapple
+
+- **suite / variant:** `` / ``
+- **instruction:** Push the pineapple off the stove
+- **success:** False
+- **control steps:** 752  (14.9s at 51 Hz)
+- **policy inferences:** 128  (replan_steps=6)
+- **video:** `push_pineapple/ep021/rollout.mp4`
+- **features:** `128` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | moving gripper toward pineapple |
+| 3.0–6.0 | 152–303 | 25–50 | pressing downward on pineapple |
+| 6.0–9.0 | 304–455 | 50–75 | nudging pineapple on burner |
+| 9.0–12.0 | 456–607 | 76–101 | hovering and pressing pineapple |
+| 12.0–14.9 | 608–751 | 101–125 | idle while touching pineapple |
+
+## Failure
+
+- **yes**
+- **onset type:** operator_intervention
+- **onset:** 3.0039525691699605 s (00:03.0)
+- **frame / control_step:** 152 (policy_step=25, chunk_index=2)
+- **coarse onset:** 4.0 s (refined=True)
+- **mode:** stuck_or_no_progress
+- **reason:** The robot reaches the pineapple on the stove but continuously presses downward and performs ineffective nudges rather than pushing it horizontally off the stove surface.
+  Summary: ineffective pushing motion; pineapple remains on stove
+- **recovery:** Re-position the gripper beside the pineapple and apply a horizontal pushing force until the pineapple moves completely off the stove surface onto the table.
+  Summary: re-align gripper sideways; push pineapple off stove surface
+- **prevention:** Approach the pineapple from the side or behind with a clear horizontal trajectory to push it off the stove without pressing it downward into the burner.
+  Summary: approach from side; apply horizontal pushing force
+- **justification:** None
diff --git a/push_pineapple/ep021/labels.json b/push_pineapple/ep021/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..f23f915849805336f3ad4cb088b3a541c668f4aa
--- /dev/null
+++ b/push_pineapple/ep021/labels.json
@@ -0,0 +1,120 @@
+{
+  "rollout_id": "t005__ep021__push_pineapple",
+  "scene_variant": "",
+  "task_id": 5,
+  "task_file": "",
+  "instruction": "Push the pineapple off the stove",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:40:59+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0039525691699605,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "moving gripper toward pineapple",
+      "description": "moving gripper toward pineapple"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0039525691699605,
+      "t_end_sec": 6.007905138339921,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "pressing downward on pineapple",
+      "description": "pressing downward on pineapple"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.007905138339921,
+      "t_end_sec": 9.011857707509881,
+      "control_step_start": 304,
+      "control_step_end": 455,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "nudging pineapple on burner",
+      "description": "nudging pineapple on burner"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.011857707509881,
+      "t_end_sec": 12.015810276679842,
+      "control_step_start": 456,
+      "control_step_end": 607,
+      "policy_step_start": 76,
+      "policy_step_end": 101,
+      "phrase": "hovering and pressing pineapple",
+      "description": "hovering and pressing pineapple"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 12.015810276679842,
+      "t_end_sec": 14.861660079051383,
+      "control_step_start": 608,
+      "control_step_end": 751,
+      "policy_step_start": 101,
+      "policy_step_end": 125,
+      "phrase": "idle while touching pineapple",
+      "description": "idle while touching pineapple"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "stuck_or_no_progress",
+    "vlm_failure_onset_type": "operator_intervention",
+    "vlm_failure_onset_seconds": 3.0039525691699605,
+    "vlm_failure_onset_timestamp": "00:03.0",
+    "vlm_failure_onset_frame": 152,
+    "vlm_failure_onset_step": 152,
+    "vlm_coarse_failure_onset_seconds": 4.0,
+    "vlm_failure_window_start_seconds": 3.0,
+    "vlm_failure_window_end_seconds": 5.0,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot reaches the pineapple on the stove but continuously presses downward and performs ineffective nudges rather than pushing it horizontally off the stove surface.",
+    "vlm_failure_reason_summary": "ineffective pushing motion; pineapple remains on stove",
+    "vlm_recovery_action": "Re-position the gripper beside the pineapple and apply a horizontal pushing force until the pineapple moves completely off the stove surface onto the table.",
+    "vlm_recovery_action_summary": "re-align gripper sideways; push pineapple off stove surface",
+    "vlm_prevention_action": "Approach the pineapple from the side or behind with a clear horizontal trajectory to push it off the stove without pressing it downward into the burner.",
+    "vlm_prevention_action_summary": "approach from side; apply horizontal pushing force",
+    "vlm_justification": null,
+    "vlm_temporal_justification": "At second 102 (01:42 in the clip), the robot gripper makes full contact with the stove surface/pineapple and begins pressing vertically downward rather than executing a horizontal trajectory to push the pineapple off the stove.",
+    "vlm_video_fps": 50.6,
+    "vlm_video_num_frames": 752,
+    "vlm_video_duration_seconds": 14.84189723320158,
+    "vlm_usage": {
+      "input_tokens": 22017,
+      "output_tokens": 327,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"stuck_or_no_progress\",\n \"onset_type\": \"operator_intervention\",\n \"failure_onset_seconds\": 4.0,\n \"failure_window_start_seconds\": 3.0,\n \"failure_window_end_seconds\": 5.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot reaches the pineapple on the stove but continuously presses downward and performs ineffective nudges rather than pushing it horizontally off the stove surface.\",\n \"failure_reason_summary\": \"ineffective pushing motion; pineapple remains on stove\",\n \"recovery_action\": \"Re-position the gripper beside the pineapple and apply a horizontal pushing force until the pineapple moves completely off the stove surface onto the table.\",\n \"recovery_action_summary\": \"re-align gripper sideways; push pineapple off stove surface\",\n \"prevention_action\": \"Approach the pineapple from the side or behind with a clear horizontal trajectory to push it off the stove without pressing it downward into the burner.\",\n \"prevention_action_summary\": \"approach from side; apply horizontal pushing force\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 152,
+    "failure_sim_step": 152,
+    "failure_policy_step": 25,
+    "failure_chunk_index": 2,
+    "first_post_failure_policy_step": 26,
+    "failure_type": "stuck_or_no_progress",
+    "correction_action": "Re-position the gripper beside the pineapple and apply a horizontal pushing force until the pineapple moves completely off the stove surface onto the table."
+  }
+}
\ No newline at end of file
diff --git a/push_pineapple/ep021/labels.npz b/push_pineapple/ep021/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..9966183215b3b2090d29ea03a40c38d80a4cefed
--- /dev/null
+++ b/push_pineapple/ep021/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8da516195f02a91c2d533fef7f42d6ebace1e12ea4c5fccb2f518bb44dfaa926
+size 3973
diff --git a/push_pineapple/ep022/example.md b/push_pineapple/ep022/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..1bbf966c0298cb08a65df13d0d8ab7271957cd25
--- /dev/null
+++ b/push_pineapple/ep022/example.md
@@ -0,0 +1,33 @@
+# t005__ep022__push_pineapple
+
+- **suite / variant:** `` / ``
+- **instruction:** Push the pineapple off the stove
+- **success:** False
+- **control steps:** 415  (8.2s at 51 Hz)
+- **policy inferences:** 71  (replan_steps=6)
+- **video:** `push_pineapple/ep022/rollout.mp4`
+- **features:** `71` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | lowering end-effector toward pineapple |
+| 3.0–6.0 | 152–303 | 25–50 | ineffectively pushing pineapple edge |
+| 6.0–8.2 | 304–414 | 50–69 | hovering near stove without progress |
+
+## Failure
+
+- **yes**
+- **onset type:** operator_intervention
+- **onset:** 3.675889328063241 s (00:03.7)
+- **frame / control_step:** 186 (policy_step=31, chunk_index=0)
+- **coarse onset:** 3.8 s (refined=True)
+- **mode:** stuck_or_no_progress
+- **reason:** The robot end-effector failed to position itself firmly behind the pineapple, instead pressing weakly on its edge and becoming stuck in repeated ineffective contact without pushing it off the stove.
+  Summary: ineffective pushing contact; stuck near stove edge
+- **recovery:** Retract the end-effector, reposition it behind the body of the pineapple, and push firmly sideways until the pineapple is completely off the stove surface.
+  Summary: reposition gripper behind pineapple; push off stove
+- **prevention:** Align the end-effector with the main body of the pineapple before making contact, then execute a clear pushing stroke across the stove to slide it off.
+  Summary: align with pineapple body; push firmly across stove
+- **justification:** At around 3.8 seconds, the robot end-effector reaches the pineapple but makes poor contact near its leafy top and stove edge. Instead of sliding the pineapple off the stove, the arm gets stuck repeating unproductive pushing motions without making progress until the episode ends.
diff --git a/push_pineapple/ep022/labels.json b/push_pineapple/ep022/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..b08131660ea84a62b8ffb3e9564577d0316e2ec0
--- /dev/null
+++ b/push_pineapple/ep022/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t005__ep022__push_pineapple",
+  "scene_variant": "",
+  "task_id": 5,
+  "task_file": "",
+  "instruction": "Push the pineapple off the stove",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:41:32+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0039525691699605,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "lowering end-effector toward pineapple",
+      "description": "lowering end-effector toward pineapple"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0039525691699605,
+      "t_end_sec": 6.007905138339921,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "ineffectively pushing pineapple edge",
+      "description": "ineffectively pushing pineapple edge"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.007905138339921,
+      "t_end_sec": 8.201581027667984,
+      "control_step_start": 304,
+      "control_step_end": 414,
+      "policy_step_start": 50,
+      "policy_step_end": 69,
+      "phrase": "hovering near stove without progress",
+      "description": "hovering near stove without progress"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "stuck_or_no_progress",
+    "vlm_failure_onset_type": "operator_intervention",
+    "vlm_failure_onset_seconds": 3.675889328063241,
+    "vlm_failure_onset_timestamp": "00:03.7",
+    "vlm_failure_onset_frame": 186,
+    "vlm_failure_onset_step": 186,
+    "vlm_coarse_failure_onset_seconds": 3.8,
+    "vlm_failure_window_start_seconds": 3.0,
+    "vlm_failure_window_end_seconds": 4.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot end-effector failed to position itself firmly behind the pineapple, instead pressing weakly on its edge and becoming stuck in repeated ineffective contact without pushing it off the stove.",
+    "vlm_failure_reason_summary": "ineffective pushing contact; stuck near stove edge",
+    "vlm_recovery_action": "Retract the end-effector, reposition it behind the body of the pineapple, and push firmly sideways until the pineapple is completely off the stove surface.",
+    "vlm_recovery_action_summary": "reposition gripper behind pineapple; push off stove",
+    "vlm_prevention_action": "Align the end-effector with the main body of the pineapple before making contact, then execute a clear pushing stroke across the stove to slide it off.",
+    "vlm_prevention_action_summary": "align with pineapple body; push firmly across stove",
+    "vlm_justification": "At around 3.8 seconds, the robot end-effector reaches the pineapple but makes poor contact near its leafy top and stove edge. Instead of sliding the pineapple off the stove, the arm gets stuck repeating unproductive pushing motions without making progress until the episode ends.",
+    "vlm_temporal_justification": "At 02:26 (frame 186 of original rollout), the robot end-effector establishes contact with the right edge of the pineapple instead of positioning behind it, immediately becoming stuck in repeated, ineffective contact without moving the pineapple off the stove.",
+    "vlm_video_fps": 50.6,
+    "vlm_video_num_frames": 415,
+    "vlm_video_duration_seconds": 8.181818181818182,
+    "vlm_usage": {
+      "input_tokens": 21581,
+      "output_tokens": 404,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"stuck_or_no_progress\",\n \"onset_type\": \"operator_intervention\",\n \"failure_onset_seconds\": 3.8,\n \"failure_window_start_seconds\": 3.0,\n \"failure_window_end_seconds\": 4.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot end-effector failed to position itself firmly behind the pineapple, instead pressing weakly on its edge and becoming stuck in repeated ineffective contact without pushing it off the stove.\",\n \"failure_reason_summary\": \"ineffective pushing contact; stuck near stove edge\",\n \"recovery_action\": \"Retract the end-effector, reposition it behind the body of the pineapple, and push firmly sideways until the pineapple is completely off the stove surface.\",\n \"recovery_action_summary\": \"reposition gripper behind pineapple; push off stove\",\n \"prevention_action\": \"Align the end-effector with the main body of the pineapple before making contact, then execute a clear pushing stroke across the stove to slide it off.\",\n \"prevention_action_summary\": \"align with pineapple body; push firmly across stove\",\n \"justification\": \"At around 3.8 seconds, the robot end-effector reaches the pineapple but makes poor contact near its leafy top and stove edge. Instead of sliding the pineapple off the stove, the arm gets stuck repeating unproductive pushing motions without making progress until the episode ends.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 186,
+    "failure_sim_step": 186,
+    "failure_policy_step": 31,
+    "failure_chunk_index": 0,
+    "first_post_failure_policy_step": 32,
+    "failure_type": "stuck_or_no_progress",
+    "correction_action": "Retract the end-effector, reposition it behind the body of the pineapple, and push firmly sideways until the pineapple is completely off the stove surface."
+  }
+}
\ No newline at end of file
diff --git a/push_pineapple/ep022/labels.npz b/push_pineapple/ep022/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..e2a692d09338350d26f0bc8607c755578cfe2559
--- /dev/null
+++ b/push_pineapple/ep022/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4fb675f3929682077de67457e4d2a4ef72618181a89cfd82a2083be58838a3b5
+size 3943
diff --git a/push_pineapple/ep023/example.md b/push_pineapple/ep023/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..95d47a31b0b57b53d63bfa831dfc76972e23c708
--- /dev/null
+++ b/push_pineapple/ep023/example.md
@@ -0,0 +1,20 @@
+# t005__ep023__push_pineapple
+
+- **suite / variant:** `` / ``
+- **instruction:** Push the pineapple off the stove
+- **success:** True
+- **control steps:** 246  (4.9s at 50 Hz)
+- **policy inferences:** 42  (replan_steps=6)
+- **video:** `push_pineapple/ep023/rollout.mp4`
+- **features:** `42` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | pushing pineapple off the stove |
+| 3.0–4.9 | 151–245 | 25–40 | retracting robot arm |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/push_pineapple/ep023/labels.json b/push_pineapple/ep023/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..f2e50d002a9f93c78f5b21c8f82d21771d2fa2cc
--- /dev/null
+++ b/push_pineapple/ep023/labels.json
@@ -0,0 +1,57 @@
+{
+  "rollout_id": "t005__ep023__push_pineapple",
+  "scene_variant": "",
+  "task_id": 5,
+  "task_file": "",
+  "instruction": "Push the pineapple off the stove",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:41:39+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.996031746031746,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "pushing pineapple off the stove",
+      "description": "pushing pineapple off the stove"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.996031746031746,
+      "t_end_sec": 4.880952380952381,
+      "control_step_start": 151,
+      "control_step_end": 245,
+      "policy_step_start": 25,
+      "policy_step_end": 40,
+      "phrase": "retracting robot arm",
+      "description": "retracting robot arm"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/push_pineapple/ep023/labels.npz b/push_pineapple/ep023/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..351da8793bdbf08757a0fce94c2f79aed06b9c85
--- /dev/null
+++ b/push_pineapple/ep023/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:da76e6f72e10d160dae7d011f0e777815458a711c4f0da7ebe8bcb88dcf9624a
+size 3043
diff --git a/push_pineapple/ep024/example.md b/push_pineapple/ep024/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..7d63237fc122fbb85eb49cef4ee064b9452f0d94
--- /dev/null
+++ b/push_pineapple/ep024/example.md
@@ -0,0 +1,22 @@
+# t005__ep024__push_pineapple
+
+- **suite / variant:** `` / ``
+- **instruction:** Push the pineapple off the stove
+- **success:** True
+- **control steps:** 478  (9.4s at 51 Hz)
+- **policy inferences:** 82  (replan_steps=6)
+- **video:** `push_pineapple/ep024/rollout.mp4`
+- **features:** `82` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | approaching pineapple on stove |
+| 3.0–6.0 | 152–303 | 25–50 | pushing pineapple across stove |
+| 6.0–9.0 | 304–455 | 50–75 | pushing pineapple off and retracting |
+| 9.0–9.4 | 456–477 | 76–79 | holding position at home |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/push_pineapple/ep024/labels.json b/push_pineapple/ep024/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..3b67cac6dff312a84a25fe214a4b26342ecf3260
--- /dev/null
+++ b/push_pineapple/ep024/labels.json
@@ -0,0 +1,79 @@
+{
+  "rollout_id": "t005__ep024__push_pineapple",
+  "scene_variant": "",
+  "task_id": 5,
+  "task_file": "",
+  "instruction": "Push the pineapple off the stove",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:41:48+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9980276134122286,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "approaching pineapple on stove",
+      "description": "approaching pineapple on stove"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9980276134122286,
+      "t_end_sec": 5.996055226824457,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "pushing pineapple across stove",
+      "description": "pushing pineapple across stove"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.996055226824457,
+      "t_end_sec": 8.994082840236686,
+      "control_step_start": 304,
+      "control_step_end": 455,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "pushing pineapple off and retracting",
+      "description": "pushing pineapple off and retracting"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.994082840236686,
+      "t_end_sec": 9.42800788954635,
+      "control_step_start": 456,
+      "control_step_end": 477,
+      "policy_step_start": 76,
+      "policy_step_end": 79,
+      "phrase": "holding position at home",
+      "description": "holding position at home"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/push_pineapple/ep024/labels.npz b/push_pineapple/ep024/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..426bb09b9ccd97fee12a5a105bd24618dcaa8aad
--- /dev/null
+++ b/push_pineapple/ep024/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7ec6dad48e3a15126bb724a17d08be56e665cc2bd839286dd96c1a950f001d18
+size 3123
diff --git a/push_pineapple/ep025/example.md b/push_pineapple/ep025/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..9f461dc95e7ac56310fe6f6376401910721a6625
--- /dev/null
+++ b/push_pineapple/ep025/example.md
@@ -0,0 +1,20 @@
+# t005__ep025__push_pineapple
+
+- **suite / variant:** `` / ``
+- **instruction:** Push the pineapple off the stove
+- **success:** True
+- **control steps:** 265  (5.3s at 50 Hz)
+- **policy inferences:** 46  (replan_steps=6)
+- **video:** `push_pineapple/ep025/rollout.mp4`
+- **features:** `46` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | pushing pineapple off stove |
+| 3.0–5.3 | 151–264 | 25–44 | retracting arm and remaining idle |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/push_pineapple/ep025/labels.json b/push_pineapple/ep025/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..1ac95cc7995a9f41cb7c9b1930640a820db9f91c
--- /dev/null
+++ b/push_pineapple/ep025/labels.json
@@ -0,0 +1,57 @@
+{
+  "rollout_id": "t005__ep025__push_pineapple",
+  "scene_variant": "",
+  "task_id": 5,
+  "task_file": "",
+  "instruction": "Push the pineapple off the stove",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:41:56+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00796812749004,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "pushing pineapple off stove",
+      "description": "pushing pineapple off stove"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00796812749004,
+      "t_end_sec": 5.278884462151394,
+      "control_step_start": 151,
+      "control_step_end": 264,
+      "policy_step_start": 25,
+      "policy_step_end": 44,
+      "phrase": "retracting arm and remaining idle",
+      "description": "retracting arm and remaining idle"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/push_pineapple/ep025/labels.npz b/push_pineapple/ep025/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..c8c17ee9ba6eae81c3b85103b8aa356aa88e684c
--- /dev/null
+++ b/push_pineapple/ep025/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4526a839a30e697012dd156bb8d9d815a6672af13b289bd179320666a35daa39
+size 3048
diff --git a/push_pineapple/ep026/example.md b/push_pineapple/ep026/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..4261586097decc74f48e3c0993dd3322af154d83
--- /dev/null
+++ b/push_pineapple/ep026/example.md
@@ -0,0 +1,32 @@
+# t005__ep026__push_pineapple
+
+- **suite / variant:** `` / ``
+- **instruction:** Push the pineapple off the stove
+- **success:** False
+- **control steps:** 246  (4.9s at 50 Hz)
+- **policy inferences:** 43  (replan_steps=6)
+- **video:** `push_pineapple/ep026/rollout.mp4`
+- **features:** `43` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | lowering gripper to push pineapple |
+| 3.0–4.9 | 151–245 | 25–40 | stuck against pineapple stem |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 2.3505976095617527 s (00:02.4)
+- **frame / control_step:** 118 (policy_step=19, chunk_index=4)
+- **coarse onset:** 2.8 s (refined=True)
+- **mode:** stuck_or_no_progress
+- **reason:** The robot lowered its end-effector onto the pineapple's stem leaves and the edge of the stove plate, getting stuck without pushing the pineapple off the stove.
+  Summary: end-effector stuck on stove edge; failing to push pineapple
+- **recovery:** Lift the end-effector above the stove, reposition behind the main body of the pineapple, and push it forward off the stove plate.
+  Summary: reposition behind pineapple; push pineapple off stove
+- **prevention:** Approach the pineapple by aligning behind its central body rather than pressing down on its stem and the tray edge.
+  Summary: align behind pineapple body; push off stove
+- **justification:** Around 2.8 seconds into the video, the robot descends directly into the stem of the pineapple and the right edge of the stove plate. It makes contact at an ineffective angle and becomes stuck, making no progress toward pushing the pineapple off the stove for the remainder of the episode.
diff --git a/push_pineapple/ep026/labels.json b/push_pineapple/ep026/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..0c8510208a903cdf6426b948b3a4bcabaeec3042
--- /dev/null
+++ b/push_pineapple/ep026/labels.json
@@ -0,0 +1,87 @@
+{
+  "rollout_id": "t005__ep026__push_pineapple",
+  "scene_variant": "",
+  "task_id": 5,
+  "task_file": "",
+  "instruction": "Push the pineapple off the stove",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:42:29+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00796812749004,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "lowering gripper to push pineapple",
+      "description": "lowering gripper to push pineapple"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00796812749004,
+      "t_end_sec": 4.900398406374502,
+      "control_step_start": 151,
+      "control_step_end": 245,
+      "policy_step_start": 25,
+      "policy_step_end": 40,
+      "phrase": "stuck against pineapple stem",
+      "description": "stuck against pineapple stem"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "stuck_or_no_progress",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 2.3505976095617527,
+    "vlm_failure_onset_timestamp": "00:02.4",
+    "vlm_failure_onset_frame": 118,
+    "vlm_failure_onset_step": 118,
+    "vlm_coarse_failure_onset_seconds": 2.8,
+    "vlm_failure_window_start_seconds": 2.3,
+    "vlm_failure_window_end_seconds": 3.3,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot lowered its end-effector onto the pineapple's stem leaves and the edge of the stove plate, getting stuck without pushing the pineapple off the stove.",
+    "vlm_failure_reason_summary": "end-effector stuck on stove edge; failing to push pineapple",
+    "vlm_recovery_action": "Lift the end-effector above the stove, reposition behind the main body of the pineapple, and push it forward off the stove plate.",
+    "vlm_recovery_action_summary": "reposition behind pineapple; push pineapple off stove",
+    "vlm_prevention_action": "Approach the pineapple by aligning behind its central body rather than pressing down on its stem and the tray edge.",
+    "vlm_prevention_action_summary": "align behind pineapple body; push off stove",
+    "vlm_justification": "Around 2.8 seconds into the video, the robot descends directly into the stem of the pineapple and the right edge of the stove plate. It makes contact at an ineffective angle and becomes stuck, making no progress toward pushing the pineapple off the stove for the remainder of the episode.",
+    "vlm_temporal_justification": "At timestamp 01:58 (frame 118), the robot's gripper makes contact with the pineapple's stem leaves and the edge of the stove top plate instead of positioning properly behind the pineapple to push it. From this moment, the end-effector presses down and gets stuck without moving the pineapple.",
+    "vlm_video_fps": 50.2,
+    "vlm_video_num_frames": 246,
+    "vlm_video_duration_seconds": 4.8804780876494025,
+    "vlm_usage": {
+      "input_tokens": 17672,
+      "output_tokens": 410,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"stuck_or_no_progress\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 2.8,\n \"failure_window_start_seconds\": 2.3,\n \"failure_window_end_seconds\": 3.3,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot lowered its end-effector onto the pineapple's stem leaves and the edge of the stove plate, getting stuck without pushing the pineapple off the stove.\",\n \"failure_reason_summary\": \"end-effector stuck on stove edge; failing to push pineapple\",\n \"recovery_action\": \"Lift the end-effector above the stove, reposition behind the main body of the pineapple, and push it forward off the stove plate.\",\n \"recovery_action_summary\": \"reposition behind pineapple; push pineapple off stove\",\n \"prevention_action\": \"Approach the pineapple by aligning behind its central body rather than pressing down on its stem and the tray edge.\",\n \"prevention_action_summary\": \"align behind pineapple body; push off stove\",\n \"justification\": \"Around 2.8 seconds into the video, the robot descends directly into the stem of the pineapple and the right edge of the stove plate. It makes contact at an ineffective angle and becomes stuck, making no progress toward pushing the pineapple off the stove for the remainder of the episode.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 118,
+    "failure_sim_step": 118,
+    "failure_policy_step": 19,
+    "failure_chunk_index": 4,
+    "first_post_failure_policy_step": 20,
+    "failure_type": "stuck_or_no_progress",
+    "correction_action": "Lift the end-effector above the stove, reposition behind the main body of the pineapple, and push it forward off the stove plate."
+  }
+}
\ No newline at end of file
diff --git a/push_pineapple/ep026/labels.npz b/push_pineapple/ep026/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..c42eb043038933b7e52f91983cb30e682f047862
--- /dev/null
+++ b/push_pineapple/ep026/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c53d5e90d193301d00d430eaf0bf7d703f2d3548d99a86cd802d35ed331b81a6
+size 3755
diff --git a/push_pineapple/ep027/example.md b/push_pineapple/ep027/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..4a7bb1febc684a34222766d0680fa231cc5e6cf6
--- /dev/null
+++ b/push_pineapple/ep027/example.md
@@ -0,0 +1,34 @@
+# t005__ep027__push_pineapple
+
+- **suite / variant:** `` / ``
+- **instruction:** Push the pineapple off the stove
+- **success:** False
+- **control steps:** 527  (10.4s at 51 Hz)
+- **policy inferences:** 91  (replan_steps=6)
+- **video:** `push_pineapple/ep027/rollout.mp4`
+- **features:** `91` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | lowering arm toward pineapple |
+| 3.0–6.0 | 152–303 | 25–50 | contacting pineapple and retracting |
+| 6.0–9.0 | 304–455 | 50–75 | hovering above pineapple |
+| 9.0–10.4 | 456–526 | 76–87 | remaining motionless over stove |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 3.8142292490118574 s (00:03.8)
+- **frame / control_step:** 193 (policy_step=32, chunk_index=1)
+- **coarse onset:** 6.0 s (refined=True)
+- **mode:** stuck_or_no_progress
+- **reason:** The robot arm approached and contacted the pineapple on the stove top but failed to exert sufficient pushing force, retracting upward without removing the pineapple from the stove.
+  Summary: failed to push pineapple; stopped making progress
+- **recovery:** Re-align the gripper behind the pineapple and push it firmly forward until it slides completely off the stove surface.
+  Summary: re-align gripper; push pineapple off stove
+- **prevention:** Approach the pineapple from a lower angle and maintain forward pushing pressure until the object falls off the stove.
+  Summary: maintain contact; push pineapple off stove
+- **justification:** At around 6.0 seconds, the robot arm touches the pineapple but fails to move it off the stove top and pulls away, leaving the pineapple in place for the remainder of the episode.
diff --git a/push_pineapple/ep027/labels.json b/push_pineapple/ep027/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..fd78466861c4d18046c2577e967e7adf75019c9b
--- /dev/null
+++ b/push_pineapple/ep027/labels.json
@@ -0,0 +1,109 @@
+{
+  "rollout_id": "t005__ep027__push_pineapple",
+  "scene_variant": "",
+  "task_id": 5,
+  "task_file": "",
+  "instruction": "Push the pineapple off the stove",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:43:03+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0039525691699605,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "lowering arm toward pineapple",
+      "description": "lowering arm toward pineapple"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0039525691699605,
+      "t_end_sec": 6.007905138339921,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "contacting pineapple and retracting",
+      "description": "contacting pineapple and retracting"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.007905138339921,
+      "t_end_sec": 9.011857707509881,
+      "control_step_start": 304,
+      "control_step_end": 455,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "hovering above pineapple",
+      "description": "hovering above pineapple"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.011857707509881,
+      "t_end_sec": 10.41501976284585,
+      "control_step_start": 456,
+      "control_step_end": 526,
+      "policy_step_start": 76,
+      "policy_step_end": 87,
+      "phrase": "remaining motionless over stove",
+      "description": "remaining motionless over stove"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "stuck_or_no_progress",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 3.8142292490118574,
+    "vlm_failure_onset_timestamp": "00:03.8",
+    "vlm_failure_onset_frame": 193,
+    "vlm_failure_onset_step": 193,
+    "vlm_coarse_failure_onset_seconds": 6.0,
+    "vlm_failure_window_start_seconds": 5.5,
+    "vlm_failure_window_end_seconds": 6.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot arm approached and contacted the pineapple on the stove top but failed to exert sufficient pushing force, retracting upward without removing the pineapple from the stove.",
+    "vlm_failure_reason_summary": "failed to push pineapple; stopped making progress",
+    "vlm_recovery_action": "Re-align the gripper behind the pineapple and push it firmly forward until it slides completely off the stove surface.",
+    "vlm_recovery_action_summary": "re-align gripper; push pineapple off stove",
+    "vlm_prevention_action": "Approach the pineapple from a lower angle and maintain forward pushing pressure until the object falls off the stove.",
+    "vlm_prevention_action_summary": "maintain contact; push pineapple off stove",
+    "vlm_justification": "At around 6.0 seconds, the robot arm touches the pineapple but fails to move it off the stove top and pulls away, leaving the pineapple in place for the remainder of the episode.",
+    "vlm_temporal_justification": "At second 41 (00:41), the robot arm makes contact with the pineapple on the stove top, but instead of applying force to push it off, it immediately fails to progress and begins retracting upward shortly after.",
+    "vlm_video_fps": 50.6,
+    "vlm_video_num_frames": 527,
+    "vlm_video_duration_seconds": 10.395256916996047,
+    "vlm_usage": {
+      "input_tokens": 21711,
+      "output_tokens": 361,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"stuck_or_no_progress\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 6.0,\n \"failure_window_start_seconds\": 5.5,\n \"failure_window_end_seconds\": 6.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot arm approached and contacted the pineapple on the stove top but failed to exert sufficient pushing force, retracting upward without removing the pineapple from the stove.\",\n \"failure_reason_summary\": \"failed to push pineapple; stopped making progress\",\n \"recovery_action\": \"Re-align the gripper behind the pineapple and push it firmly forward until it slides completely off the stove surface.\",\n \"recovery_action_summary\": \"re-align gripper; push pineapple off stove\",\n \"prevention_action\": \"Approach the pineapple from a lower angle and maintain forward pushing pressure until the object falls off the stove.\",\n \"prevention_action_summary\": \"maintain contact; push pineapple off stove\",\n \"justification\": \"At around 6.0 seconds, the robot arm touches the pineapple but fails to move it off the stove top and pulls away, leaving the pineapple in place for the remainder of the episode.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 193,
+    "failure_sim_step": 193,
+    "failure_policy_step": 32,
+    "failure_chunk_index": 1,
+    "first_post_failure_policy_step": 33,
+    "failure_type": "stuck_or_no_progress",
+    "correction_action": "Re-align the gripper behind the pineapple and push it firmly forward until it slides completely off the stove surface."
+  }
+}
\ No newline at end of file
diff --git a/push_pineapple/ep027/labels.npz b/push_pineapple/ep027/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..27be133e8bb91ce254ce13df1d4da50e8915f5c0
--- /dev/null
+++ b/push_pineapple/ep027/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:455f922764d4dd1fd215ab2dd75265d97096d4f70ea373a1b329d2aab1b94f04
+size 3837
diff --git a/push_pineapple/ep028/example.md b/push_pineapple/ep028/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..8cd18c69e2daf0a60bb0dda528780c3378592d2e
--- /dev/null
+++ b/push_pineapple/ep028/example.md
@@ -0,0 +1,22 @@
+# t005__ep028__push_pineapple
+
+- **suite / variant:** `` / ``
+- **instruction:** Push the pineapple off the stove
+- **success:** True
+- **control steps:** 571  (11.2s at 51 Hz)
+- **policy inferences:** 98  (replan_steps=6)
+- **video:** `push_pineapple/ep028/rollout.mp4`
+- **features:** `98` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–153 | 0–25 | approaching the pineapple |
+| 3.0–6.0 | 154–307 | 25–51 | positioning near pineapple |
+| 6.0–9.0 | 308–461 | 51–76 | pushing pineapple off stove |
+| 9.0–11.2 | 462–570 | 77–95 | retracting robot arm |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/push_pineapple/ep028/labels.json b/push_pineapple/ep028/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..27474b50c2a0335b1fdf806392fbd3b87b3ab4f0
--- /dev/null
+++ b/push_pineapple/ep028/labels.json
@@ -0,0 +1,79 @@
+{
+  "rollout_id": "t005__ep028__push_pineapple",
+  "scene_variant": "",
+  "task_id": 5,
+  "task_file": "",
+  "instruction": "Push the pineapple off the stove",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:43:11+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0078125,
+      "control_step_start": 0,
+      "control_step_end": 153,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "approaching the pineapple",
+      "description": "approaching the pineapple"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0078125,
+      "t_end_sec": 6.015625,
+      "control_step_start": 154,
+      "control_step_end": 307,
+      "policy_step_start": 25,
+      "policy_step_end": 51,
+      "phrase": "positioning near pineapple",
+      "description": "positioning near pineapple"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.015625,
+      "t_end_sec": 9.0234375,
+      "control_step_start": 308,
+      "control_step_end": 461,
+      "policy_step_start": 51,
+      "policy_step_end": 76,
+      "phrase": "pushing pineapple off stove",
+      "description": "pushing pineapple off stove"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.0234375,
+      "t_end_sec": 11.15234375,
+      "control_step_start": 462,
+      "control_step_end": 570,
+      "policy_step_start": 77,
+      "policy_step_end": 95,
+      "phrase": "retracting robot arm",
+      "description": "retracting robot arm"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/push_pineapple/ep028/labels.npz b/push_pineapple/ep028/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..32cc2e3419af0e051a09ff15bb5d7e40bf502ebf
--- /dev/null
+++ b/push_pineapple/ep028/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2512c2d67d606619eac0778ab171c0b824c75963fae04f69dcfd9f3cf7a9e343
+size 3122
diff --git a/push_pineapple/ep029/example.md b/push_pineapple/ep029/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..e0462f1b359edd35dd65a70139aa52e5190501d9
--- /dev/null
+++ b/push_pineapple/ep029/example.md
@@ -0,0 +1,32 @@
+# t005__ep029__push_pineapple
+
+- **suite / variant:** `` / ``
+- **instruction:** Push the pineapple off the stove
+- **success:** False
+- **control steps:** 283  (5.5s at 51 Hz)
+- **policy inferences:** 48  (replan_steps=6)
+- **video:** `push_pineapple/ep029/rollout.mp4`
+- **features:** `48` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–152 | 0–25 | lowering arm toward stove |
+| 3.0–5.5 | 153–282 | 25–47 | swiping right and missing pineapple |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 3.2941176470588234 s (00:03.3)
+- **frame / control_step:** 168 (policy_step=28, chunk_index=0)
+- **coarse onset:** 2.8 s (refined=True)
+- **mode:** stuck_or_no_progress
+- **reason:** The robot arm lowered near the stove but swiped past the pineapple without making contact, leaving the pineapple stationary on the stove.
+  Summary: swiping past object; missing contact with pineapple
+- **recovery:** Reposition the gripper behind the pineapple on the stove and push it forward until it falls off the stove surface.
+  Summary: reposition behind pineapple; push off stove
+- **prevention:** Align the gripper directly behind the pineapple before initiating the push trajectory across the stove.
+  Summary: align gripper behind pineapple; execute forward push
+- **justification:** Around 2.8 seconds into the rollout, the end-effector lowers to the table surface next to the stove and moves to the right, completely bypassing contact with the pineapple. Because no contact is made, the pineapple remains on the stove for the rest of the episode.
diff --git a/push_pineapple/ep029/labels.json b/push_pineapple/ep029/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..3bcf15ad1d1831643721bebdd22129627284a4d6
--- /dev/null
+++ b/push_pineapple/ep029/labels.json
@@ -0,0 +1,87 @@
+{
+  "rollout_id": "t005__ep029__push_pineapple",
+  "scene_variant": "",
+  "task_id": 5,
+  "task_file": "",
+  "instruction": "Push the pineapple off the stove",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:43:54+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0,
+      "control_step_start": 0,
+      "control_step_end": 152,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "lowering arm toward stove",
+      "description": "lowering arm toward stove"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0,
+      "t_end_sec": 5.549019607843137,
+      "control_step_start": 153,
+      "control_step_end": 282,
+      "policy_step_start": 25,
+      "policy_step_end": 47,
+      "phrase": "swiping right and missing pineapple",
+      "description": "swiping right and missing pineapple"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "stuck_or_no_progress",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 3.2941176470588234,
+    "vlm_failure_onset_timestamp": "00:03.3",
+    "vlm_failure_onset_frame": 168,
+    "vlm_failure_onset_step": 168,
+    "vlm_coarse_failure_onset_seconds": 2.8,
+    "vlm_failure_window_start_seconds": 2.3,
+    "vlm_failure_window_end_seconds": 3.3,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot arm lowered near the stove but swiped past the pineapple without making contact, leaving the pineapple stationary on the stove.",
+    "vlm_failure_reason_summary": "swiping past object; missing contact with pineapple",
+    "vlm_recovery_action": "Reposition the gripper behind the pineapple on the stove and push it forward until it falls off the stove surface.",
+    "vlm_recovery_action_summary": "reposition behind pineapple; push off stove",
+    "vlm_prevention_action": "Align the gripper directly behind the pineapple before initiating the push trajectory across the stove.",
+    "vlm_prevention_action_summary": "align gripper behind pineapple; execute forward push",
+    "vlm_justification": "Around 2.8 seconds into the rollout, the end-effector lowers to the table surface next to the stove and moves to the right, completely bypassing contact with the pineapple. Because no contact is made, the pineapple remains on the stove for the rest of the episode.",
+    "vlm_temporal_justification": "At timestamp 02:48 (frame 168), the robot end-effector finishes descending near the stove and begins its trajectory to push, but it is positioned off-target behind the pineapple and swipes past without making contact.",
+    "vlm_video_fps": 51.0,
+    "vlm_video_num_frames": 283,
+    "vlm_video_duration_seconds": 5.529411764705882,
+    "vlm_usage": {
+      "input_tokens": 20059,
+      "output_tokens": 370,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"stuck_or_no_progress\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 2.8,\n \"failure_window_start_seconds\": 2.3,\n \"failure_window_end_seconds\": 3.3,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot arm lowered near the stove but swiped past the pineapple without making contact, leaving the pineapple stationary on the stove.\",\n \"failure_reason_summary\": \"swiping past object; missing contact with pineapple\",\n \"recovery_action\": \"Reposition the gripper behind the pineapple on the stove and push it forward until it falls off the stove surface.\",\n \"recovery_action_summary\": \"reposition behind pineapple; push off stove\",\n \"prevention_action\": \"Align the gripper directly behind the pineapple before initiating the push trajectory across the stove.\",\n \"prevention_action_summary\": \"align gripper behind pineapple; execute forward push\",\n \"justification\": \"Around 2.8 seconds into the rollout, the end-effector lowers to the table surface next to the stove and moves to the right, completely bypassing contact with the pineapple. Because no contact is made, the pineapple remains on the stove for the rest of the episode.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 168,
+    "failure_sim_step": 168,
+    "failure_policy_step": 28,
+    "failure_chunk_index": 0,
+    "first_post_failure_policy_step": 29,
+    "failure_type": "stuck_or_no_progress",
+    "correction_action": "Reposition the gripper behind the pineapple on the stove and push it forward until it falls off the stove surface."
+  }
+}
\ No newline at end of file
diff --git a/push_pineapple/ep029/labels.npz b/push_pineapple/ep029/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..d43734e3a88bbabe0fa73e83f6b8d643ab73da0b
--- /dev/null
+++ b/push_pineapple/ep029/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:92afed618506bf2dd2a0dab8b5194bb7de366854c1dfbdc908c36013de578b2e
+size 3698
diff --git a/push_pineapple/ep030/example.md b/push_pineapple/ep030/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..62098c8a5e7c45f139b3e304b16c5ac0f6874bce
--- /dev/null
+++ b/push_pineapple/ep030/example.md
@@ -0,0 +1,32 @@
+# t005__ep030__push_pineapple
+
+- **suite / variant:** `` / ``
+- **instruction:** Push the pineapple off the stove
+- **success:** False
+- **control steps:** 277  (5.5s at 50 Hz)
+- **policy inferences:** 48  (replan_steps=6)
+- **video:** `push_pineapple/ep030/rollout.mp4`
+- **features:** `48` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | moving gripper toward pineapple |
+| 3.0–5.5 | 152–276 | 25–46 | holding position over pineapple |
+
+## Failure
+
+- **yes**
+- **onset type:** operator_intervention
+- **onset:** 2.9702970297029703 s (00:03.0)
+- **frame / control_step:** 150 (policy_step=25, chunk_index=0)
+- **coarse onset:** 3.5 s (refined=True)
+- **mode:** stuck_or_no_progress
+- **reason:** The robot arm lowered its gripper onto the pineapple on the stove but failed to execute a forward pushing motion, becoming stuck in place over the stove tray.
+  Summary: stuck at target; failed to push pineapple off stove
+- **recovery:** Lift and reposition the gripper behind the pineapple, then execute a horizontal forward sweep to push the pineapple off the stove tray.
+  Summary: reposition gripper; push pineapple off stove
+- **prevention:** Approach the pineapple from the side or back at stove height and push across the surface rather than stopping upon contact.
+  Summary: align trajectory; push pineapple off stove
+- **justification:** By around 3.5 seconds, the robot's end-effector reaches the pineapple on the stove plate but freezes/halts instead of sweeping forward to push the pineapple off, failing to make any further progress.
diff --git a/push_pineapple/ep030/labels.json b/push_pineapple/ep030/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..de3cc51e692a42c24314438cc20e8d09db10c181
--- /dev/null
+++ b/push_pineapple/ep030/labels.json
@@ -0,0 +1,87 @@
+{
+  "rollout_id": "t005__ep030__push_pineapple",
+  "scene_variant": "",
+  "task_id": 5,
+  "task_file": "",
+  "instruction": "Push the pineapple off the stove",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:44:23+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00990099009901,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "moving gripper toward pineapple",
+      "description": "moving gripper toward pineapple"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00990099009901,
+      "t_end_sec": 5.485148514851486,
+      "control_step_start": 152,
+      "control_step_end": 276,
+      "policy_step_start": 25,
+      "policy_step_end": 46,
+      "phrase": "holding position over pineapple",
+      "description": "holding position over pineapple"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "stuck_or_no_progress",
+    "vlm_failure_onset_type": "operator_intervention",
+    "vlm_failure_onset_seconds": 2.9702970297029703,
+    "vlm_failure_onset_timestamp": "00:03.0",
+    "vlm_failure_onset_frame": 150,
+    "vlm_failure_onset_step": 150,
+    "vlm_coarse_failure_onset_seconds": 3.5,
+    "vlm_failure_window_start_seconds": 3.0,
+    "vlm_failure_window_end_seconds": 4.0,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot arm lowered its gripper onto the pineapple on the stove but failed to execute a forward pushing motion, becoming stuck in place over the stove tray.",
+    "vlm_failure_reason_summary": "stuck at target; failed to push pineapple off stove",
+    "vlm_recovery_action": "Lift and reposition the gripper behind the pineapple, then execute a horizontal forward sweep to push the pineapple off the stove tray.",
+    "vlm_recovery_action_summary": "reposition gripper; push pineapple off stove",
+    "vlm_prevention_action": "Approach the pineapple from the side or back at stove height and push across the surface rather than stopping upon contact.",
+    "vlm_prevention_action_summary": "align trajectory; push pineapple off stove",
+    "vlm_justification": "By around 3.5 seconds, the robot's end-effector reaches the pineapple on the stove plate but freezes/halts instead of sweeping forward to push the pineapple off, failing to make any further progress.",
+    "vlm_temporal_justification": "At 02:05 into the refined clip, the robot gripper finishes descending onto the stove tray next to the pineapple. From this point onward, it remains completely stationary without initiating any forward pushing movement to move the pineapple off the stove.",
+    "vlm_video_fps": 50.5,
+    "vlm_video_num_frames": 277,
+    "vlm_video_duration_seconds": 5.465346534653466,
+    "vlm_usage": {
+      "input_tokens": 18049,
+      "output_tokens": 370,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"stuck_or_no_progress\",\n \"onset_type\": \"operator_intervention\",\n \"failure_onset_seconds\": 3.5,\n \"failure_window_start_seconds\": 3.0,\n \"failure_window_end_seconds\": 4.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot arm lowered its gripper onto the pineapple on the stove but failed to execute a forward pushing motion, becoming stuck in place over the stove tray.\",\n \"failure_reason_summary\": \"stuck at target; failed to push pineapple off stove\",\n \"recovery_action\": \"Lift and reposition the gripper behind the pineapple, then execute a horizontal forward sweep to push the pineapple off the stove tray.\",\n \"recovery_action_summary\": \"reposition gripper; push pineapple off stove\",\n \"prevention_action\": \"Approach the pineapple from the side or back at stove height and push across the surface rather than stopping upon contact.\",\n \"prevention_action_summary\": \"align trajectory; push pineapple off stove\",\n \"justification\": \"By around 3.5 seconds, the robot's end-effector reaches the pineapple on the stove plate but freezes/halts instead of sweeping forward to push the pineapple off, failing to make any further progress.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 150,
+    "failure_sim_step": 150,
+    "failure_policy_step": 25,
+    "failure_chunk_index": 0,
+    "first_post_failure_policy_step": 26,
+    "failure_type": "stuck_or_no_progress",
+    "correction_action": "Lift and reposition the gripper behind the pineapple, then execute a horizontal forward sweep to push the pineapple off the stove tray."
+  }
+}
\ No newline at end of file
diff --git a/push_pineapple/ep030/labels.npz b/push_pineapple/ep030/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..614bbdb2e99a1a7615c381979b09facad9952deb
--- /dev/null
+++ b/push_pineapple/ep030/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:448c8434cc2898f7a643b5fd05ff918578016cad52f43fa5e912d60a585a0a18
+size 3768
diff --git a/push_pineapple/ep031/example.md b/push_pineapple/ep031/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..83e86d012392951878f871bb386a312f6ec6b16e
--- /dev/null
+++ b/push_pineapple/ep031/example.md
@@ -0,0 +1,23 @@
+# t005__ep031__push_pineapple
+
+- **suite / variant:** `` / ``
+- **instruction:** Push the pineapple off the stove
+- **success:** True
+- **control steps:** 720  (14.3s at 50 Hz)
+- **policy inferences:** 124  (replan_steps=6)
+- **video:** `push_pineapple/ep031/rollout.mp4`
+- **features:** `124` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching toward the pineapple |
+| 3.0–6.0 | 152–303 | 25–50 | pushing pineapple across the stove |
+| 6.0–9.0 | 304–455 | 50–75 | adjusting push on pineapple |
+| 9.0–12.0 | 456–607 | 76–101 | pushing pineapple off the stove |
+| 12.0–14.3 | 608–719 | 101–119 | retracting robot arm |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/push_pineapple/ep031/labels.json b/push_pineapple/ep031/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..67d1a8f835d0f568b76b3ceb72226ebb6adced21
--- /dev/null
+++ b/push_pineapple/ep031/labels.json
@@ -0,0 +1,90 @@
+{
+  "rollout_id": "t005__ep031__push_pineapple",
+  "scene_variant": "",
+  "task_id": 5,
+  "task_file": "",
+  "instruction": "Push the pineapple off the stove",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:44:33+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00990099009901,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching toward the pineapple",
+      "description": "reaching toward the pineapple"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00990099009901,
+      "t_end_sec": 6.01980198019802,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "pushing pineapple across the stove",
+      "description": "pushing pineapple across the stove"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01980198019802,
+      "t_end_sec": 9.029702970297029,
+      "control_step_start": 304,
+      "control_step_end": 455,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "adjusting push on pineapple",
+      "description": "adjusting push on pineapple"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.029702970297029,
+      "t_end_sec": 12.03960396039604,
+      "control_step_start": 456,
+      "control_step_end": 607,
+      "policy_step_start": 76,
+      "policy_step_end": 101,
+      "phrase": "pushing pineapple off the stove",
+      "description": "pushing pineapple off the stove"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 12.03960396039604,
+      "t_end_sec": 14.257425742574258,
+      "control_step_start": 608,
+      "control_step_end": 719,
+      "policy_step_start": 101,
+      "policy_step_end": 119,
+      "phrase": "retracting robot arm",
+      "description": "retracting robot arm"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/push_pineapple/ep031/labels.npz b/push_pineapple/ep031/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..9b21ab7bfc235b58e5d611c996f5c49d0af6a92c
--- /dev/null
+++ b/push_pineapple/ep031/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6403f35b23c051a7d441c98897f42031fdfc74655f5cda4841a89dab024cbf35
+size 3161
diff --git a/push_pineapple/ep032/example.md b/push_pineapple/ep032/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..3feda6dc72a03c5c413cffc9583a2868505777fd
--- /dev/null
+++ b/push_pineapple/ep032/example.md
@@ -0,0 +1,22 @@
+# t005__ep032__push_pineapple
+
+- **suite / variant:** `` / ``
+- **instruction:** Push the pineapple off the stove
+- **success:** True
+- **control steps:** 463  (9.2s at 50 Hz)
+- **policy inferences:** 79  (replan_steps=6)
+- **video:** `push_pineapple/ep032/rollout.mp4`
+- **features:** `79` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching for pineapple |
+| 3.0–6.0 | 152–303 | 25–50 | pushing pineapple off stove |
+| 6.0–9.0 | 304–455 | 50–75 | retracting robot arm |
+| 9.0–9.2 | 456–462 | 76–77 | idle at home position |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/push_pineapple/ep032/labels.json b/push_pineapple/ep032/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..50e45cc5188cc4aa0c4356cdd1f7c60c40255a50
--- /dev/null
+++ b/push_pineapple/ep032/labels.json
@@ -0,0 +1,79 @@
+{
+  "rollout_id": "t005__ep032__push_pineapple",
+  "scene_variant": "",
+  "task_id": 5,
+  "task_file": "",
+  "instruction": "Push the pineapple off the stove",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:44:39+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00990099009901,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for pineapple",
+      "description": "reaching for pineapple"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00990099009901,
+      "t_end_sec": 6.01980198019802,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "pushing pineapple off stove",
+      "description": "pushing pineapple off stove"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01980198019802,
+      "t_end_sec": 9.029702970297029,
+      "control_step_start": 304,
+      "control_step_end": 455,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "retracting robot arm",
+      "description": "retracting robot arm"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.029702970297029,
+      "t_end_sec": 9.168316831683168,
+      "control_step_start": 456,
+      "control_step_end": 462,
+      "policy_step_start": 76,
+      "policy_step_end": 77,
+      "phrase": "idle at home position",
+      "description": "idle at home position"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/push_pineapple/ep032/labels.npz b/push_pineapple/ep032/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..99cf74432fed96d1a8b899ca0a31efd3a1eabb98
--- /dev/null
+++ b/push_pineapple/ep032/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:89460f1ebae8ddde07c2e931c02c2e566106557f769b193d9a6e8ef687bf8dd5
+size 3126
diff --git a/push_pineapple/ep033/example.md b/push_pineapple/ep033/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..942b308afb3c72ccee78765a0d92b8ebce027728
--- /dev/null
+++ b/push_pineapple/ep033/example.md
@@ -0,0 +1,33 @@
+# t005__ep033__push_pineapple
+
+- **suite / variant:** `` / ``
+- **instruction:** Push the pineapple off the stove
+- **success:** False
+- **control steps:** 354  (7.0s at 50 Hz)
+- **policy inferences:** 60  (replan_steps=6)
+- **video:** `push_pineapple/ep033/rollout.mp4`
+- **features:** `60` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | lowering gripper toward pineapple |
+| 3.0–6.0 | 151–301 | 25–50 | grasping pineapple and remaining stationary |
+| 6.0–7.0 | 302–353 | 50–58 | idle while holding pineapple |
+
+## Failure
+
+- **yes**
+- **onset type:** operator_intervention
+- **onset:** 3.6779324055666005 s (00:03.7)
+- **frame / control_step:** 185 (policy_step=30, chunk_index=5)
+- **coarse onset:** 4.0 s (refined=True)
+- **mode:** stuck_or_no_progress
+- **reason:** The robot gripped the pineapple on the stove but stopped advancing the motion, hovering in place without pushing or sliding the fruit off the cooktop.
+  Summary: hovering while holding object; no progress pushing off stove
+- **recovery:** Move the end-effector horizontally toward the side of the table to slide or transport the pineapple completely off the stove surface.
+  Summary: slide object laterally; place pineapple on table
+- **prevention:** Execute a continuous horizontal pushing or pulling motion across the cooktop to clear the pineapple off the stove without stopping.
+  Summary: push pineapple off cooktop; maintain horizontal movement
+- **justification:** By frame 00:04 (around 4.0s), the robot has grasped the pineapple on the stove but freezes/hovers in place without moving it off the stove, making no further progress for the remainder of the rollout.
diff --git a/push_pineapple/ep033/labels.json b/push_pineapple/ep033/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..4a02e95440f976d6cc4238052a3446fb6c23e7c5
--- /dev/null
+++ b/push_pineapple/ep033/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t005__ep033__push_pineapple",
+  "scene_variant": "",
+  "task_id": 5,
+  "task_file": "",
+  "instruction": "Push the pineapple off the stove",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:45:09+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.001988071570577,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "lowering gripper toward pineapple",
+      "description": "lowering gripper toward pineapple"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.001988071570577,
+      "t_end_sec": 6.003976143141154,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "grasping pineapple and remaining stationary",
+      "description": "grasping pineapple and remaining stationary"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.003976143141154,
+      "t_end_sec": 7.037773359840955,
+      "control_step_start": 302,
+      "control_step_end": 353,
+      "policy_step_start": 50,
+      "policy_step_end": 58,
+      "phrase": "idle while holding pineapple",
+      "description": "idle while holding pineapple"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "stuck_or_no_progress",
+    "vlm_failure_onset_type": "operator_intervention",
+    "vlm_failure_onset_seconds": 3.6779324055666005,
+    "vlm_failure_onset_timestamp": "00:03.7",
+    "vlm_failure_onset_frame": 185,
+    "vlm_failure_onset_step": 185,
+    "vlm_coarse_failure_onset_seconds": 4.0,
+    "vlm_failure_window_start_seconds": 3.5,
+    "vlm_failure_window_end_seconds": 4.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot gripped the pineapple on the stove but stopped advancing the motion, hovering in place without pushing or sliding the fruit off the cooktop.",
+    "vlm_failure_reason_summary": "hovering while holding object; no progress pushing off stove",
+    "vlm_recovery_action": "Move the end-effector horizontally toward the side of the table to slide or transport the pineapple completely off the stove surface.",
+    "vlm_recovery_action_summary": "slide object laterally; place pineapple on table",
+    "vlm_prevention_action": "Execute a continuous horizontal pushing or pulling motion across the cooktop to clear the pineapple off the stove without stopping.",
+    "vlm_prevention_action_summary": "push pineapple off cooktop; maintain horizontal movement",
+    "vlm_justification": "By frame 00:04 (around 4.0s), the robot has grasped the pineapple on the stove but freezes/hovers in place without moving it off the stove, making no further progress for the remainder of the rollout.",
+    "vlm_temporal_justification": "At 02:15 (second 135), the robot gripper finishes making contact with the fruit on the stove. From this point onward, the robot completely ceases movement and remains frozen in place, failing to advance any motion to push the fruit off the cooktop.",
+    "vlm_video_fps": 50.3,
+    "vlm_video_num_frames": 354,
+    "vlm_video_duration_seconds": 7.017892644135189,
+    "vlm_usage": {
+      "input_tokens": 21386,
+      "output_tokens": 385,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"stuck_or_no_progress\",\n \"onset_type\": \"operator_intervention\",\n \"failure_onset_seconds\": 4.0,\n \"failure_window_start_seconds\": 3.5,\n \"failure_window_end_seconds\": 4.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot gripped the pineapple on the stove but stopped advancing the motion, hovering in place without pushing or sliding the fruit off the cooktop.\",\n \"failure_reason_summary\": \"hovering while holding object; no progress pushing off stove\",\n \"recovery_action\": \"Move the end-effector horizontally toward the side of the table to slide or transport the pineapple completely off the stove surface.\",\n \"recovery_action_summary\": \"slide object laterally; place pineapple on table\",\n \"prevention_action\": \"Execute a continuous horizontal pushing or pulling motion across the cooktop to clear the pineapple off the stove without stopping.\",\n \"prevention_action_summary\": \"push pineapple off cooktop; maintain horizontal movement\",\n \"justification\": \"By frame 00:04 (around 4.0s), the robot has grasped the pineapple on the stove but freezes/hovers in place without moving it off the stove, making no further progress for the remainder of the rollout.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 185,
+    "failure_sim_step": 185,
+    "failure_policy_step": 30,
+    "failure_chunk_index": 5,
+    "first_post_failure_policy_step": 31,
+    "failure_type": "stuck_or_no_progress",
+    "correction_action": "Move the end-effector horizontally toward the side of the table to slide or transport the pineapple completely off the stove surface."
+  }
+}
\ No newline at end of file
diff --git a/push_pineapple/ep033/labels.npz b/push_pineapple/ep033/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..317f6bba7454035fb6c8a0fcacf2833ec66e8f9c
--- /dev/null
+++ b/push_pineapple/ep033/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:df27844409ab8a84f24fd0a3f4b09ae230b18c28d18042866d2b2649c9023848
+size 3824
diff --git a/push_pineapple/ep034/example.md b/push_pineapple/ep034/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..565983b4d6f1457511d00fa8d7ac604e06ae845c
--- /dev/null
+++ b/push_pineapple/ep034/example.md
@@ -0,0 +1,33 @@
+# t005__ep034__push_pineapple
+
+- **suite / variant:** `` / ``
+- **instruction:** Push the pineapple off the stove
+- **success:** False
+- **control steps:** 392  (7.7s at 51 Hz)
+- **policy inferences:** 68  (replan_steps=6)
+- **video:** `push_pineapple/ep034/rollout.mp4`
+- **features:** `68` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching toward pineapple on stove |
+| 3.0–6.0 | 152–303 | 25–50 | contacting pineapple on stove |
+| 6.0–7.7 | 304–391 | 50–65 | hovering near pineapple; no progress |
+
+## Failure
+
+- **yes**
+- **onset type:** operator_intervention
+- **onset:** 4.841897233201581 s (00:04.8)
+- **frame / control_step:** 245 (policy_step=40, chunk_index=5)
+- **coarse onset:** 5.0 s (refined=True)
+- **mode:** stuck_or_no_progress
+- **reason:** The robot reaches the pineapple on the stove but presses lightly down on it rather than executing a horizontal pushing motion, failing to move the pineapple off the cooktop before time expires.
+  Summary: ineffective pushing motion; pineapple remains on stove
+- **recovery:** Reposition the end-effector behind or to the side of the pineapple and push it laterally until it falls off the stove surface.
+  Summary: reposition behind pineapple; push off stove
+- **prevention:** Align the end-effector beside the pineapple before pushing horizontally across the stove top.
+  Summary: align gripper laterally; push pineapple off stove
+- **justification:** At around 5.0 seconds, the robot reaches the pineapple but makes ineffective contact from above rather than sweeping horizontally. It remains hovering/pressing on the object without sliding it off the stove until the episode ends.
diff --git a/push_pineapple/ep034/labels.json b/push_pineapple/ep034/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..b0b14bd0cdb2b45ee6ba1ba7a575e6d125e8d2ef
--- /dev/null
+++ b/push_pineapple/ep034/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t005__ep034__push_pineapple",
+  "scene_variant": "",
+  "task_id": 5,
+  "task_file": "",
+  "instruction": "Push the pineapple off the stove",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:45:44+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0039525691699605,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching toward pineapple on stove",
+      "description": "reaching toward pineapple on stove"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0039525691699605,
+      "t_end_sec": 6.007905138339921,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "contacting pineapple on stove",
+      "description": "contacting pineapple on stove"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.007905138339921,
+      "t_end_sec": 7.747035573122529,
+      "control_step_start": 304,
+      "control_step_end": 391,
+      "policy_step_start": 50,
+      "policy_step_end": 65,
+      "phrase": "hovering near pineapple; no progress",
+      "description": "hovering near pineapple; no progress"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "stuck_or_no_progress",
+    "vlm_failure_onset_type": "operator_intervention",
+    "vlm_failure_onset_seconds": 4.841897233201581,
+    "vlm_failure_onset_timestamp": "00:04.8",
+    "vlm_failure_onset_frame": 245,
+    "vlm_failure_onset_step": 245,
+    "vlm_coarse_failure_onset_seconds": 5.0,
+    "vlm_failure_window_start_seconds": 4.8,
+    "vlm_failure_window_end_seconds": 7.73,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot reaches the pineapple on the stove but presses lightly down on it rather than executing a horizontal pushing motion, failing to move the pineapple off the cooktop before time expires.",
+    "vlm_failure_reason_summary": "ineffective pushing motion; pineapple remains on stove",
+    "vlm_recovery_action": "Reposition the end-effector behind or to the side of the pineapple and push it laterally until it falls off the stove surface.",
+    "vlm_recovery_action_summary": "reposition behind pineapple; push off stove",
+    "vlm_prevention_action": "Align the end-effector beside the pineapple before pushing horizontally across the stove top.",
+    "vlm_prevention_action_summary": "align gripper laterally; push pineapple off stove",
+    "vlm_justification": "At around 5.0 seconds, the robot reaches the pineapple but makes ineffective contact from above rather than sweeping horizontally. It remains hovering/pressing on the object without sliding it off the stove until the episode ends.",
+    "vlm_temporal_justification": "At second 02:24 of the clip, the robot arm finishes descending and makes contact next to the pineapple. From this moment onward, it remains pressed down without making any horizontal pushing progress to clear the pineapple off the stove.",
+    "vlm_video_fps": 50.6,
+    "vlm_video_num_frames": 392,
+    "vlm_video_duration_seconds": 7.727272727272727,
+    "vlm_usage": {
+      "input_tokens": 20703,
+      "output_tokens": 371,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"stuck_or_no_progress\",\n \"onset_type\": \"operator_intervention\",\n \"failure_onset_seconds\": 5.0,\n \"failure_window_start_seconds\": 4.8,\n \"failure_window_end_seconds\": 7.73,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot reaches the pineapple on the stove but presses lightly down on it rather than executing a horizontal pushing motion, failing to move the pineapple off the cooktop before time expires.\",\n \"failure_reason_summary\": \"ineffective pushing motion; pineapple remains on stove\",\n \"recovery_action\": \"Reposition the end-effector behind or to the side of the pineapple and push it laterally until it falls off the stove surface.\",\n \"recovery_action_summary\": \"reposition behind pineapple; push off stove\",\n \"prevention_action\": \"Align the end-effector beside the pineapple before pushing horizontally across the stove top.\",\n \"prevention_action_summary\": \"align gripper laterally; push pineapple off stove\",\n \"justification\": \"At around 5.0 seconds, the robot reaches the pineapple but makes ineffective contact from above rather than sweeping horizontally. It remains hovering/pressing on the object without sliding it off the stove until the episode ends.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 245,
+    "failure_sim_step": 245,
+    "failure_policy_step": 40,
+    "failure_chunk_index": 5,
+    "first_post_failure_policy_step": 41,
+    "failure_type": "stuck_or_no_progress",
+    "correction_action": "Reposition the end-effector behind or to the side of the pineapple and push it laterally until it falls off the stove surface."
+  }
+}
\ No newline at end of file
diff --git a/push_pineapple/ep034/labels.npz b/push_pineapple/ep034/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..69f4367dadbd045be3537348c74cd63ae3f057cd
--- /dev/null
+++ b/push_pineapple/ep034/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:246cf9cfecf4ccb623dc780d15e26370c4c3b490178f439f58d3752ce65c1bc2
+size 3801
diff --git a/push_pineapple/ep035/example.md b/push_pineapple/ep035/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..cf5a4bdfdc9c263f457329b1fb890986bb6a5d55
--- /dev/null
+++ b/push_pineapple/ep035/example.md
@@ -0,0 +1,22 @@
+# t005__ep035__push_pineapple
+
+- **suite / variant:** `` / ``
+- **instruction:** Push the pineapple off the stove
+- **success:** True
+- **control steps:** 564  (11.1s at 51 Hz)
+- **policy inferences:** 96  (replan_steps=6)
+- **video:** `push_pineapple/ep035/rollout.mp4`
+- **features:** `96` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching toward pineapple |
+| 3.0–6.0 | 152–303 | 25–50 | positioning behind pineapple |
+| 6.0–9.0 | 304–455 | 50–75 | pushing pineapple off stove |
+| 9.0–11.1 | 456–563 | 76–93 | retracting robot arm |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/push_pineapple/ep035/labels.json b/push_pineapple/ep035/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..6e7d067e200b8a9bf4f2831ddabe9a19590339bd
--- /dev/null
+++ b/push_pineapple/ep035/labels.json
@@ -0,0 +1,79 @@
+{
+  "rollout_id": "t005__ep035__push_pineapple",
+  "scene_variant": "",
+  "task_id": 5,
+  "task_file": "",
+  "instruction": "Push the pineapple off the stove",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:45:53+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9980276134122286,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching toward pineapple",
+      "description": "reaching toward pineapple"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9980276134122286,
+      "t_end_sec": 5.996055226824457,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "positioning behind pineapple",
+      "description": "positioning behind pineapple"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.996055226824457,
+      "t_end_sec": 8.994082840236686,
+      "control_step_start": 304,
+      "control_step_end": 455,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "pushing pineapple off stove",
+      "description": "pushing pineapple off stove"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.994082840236686,
+      "t_end_sec": 11.124260355029586,
+      "control_step_start": 456,
+      "control_step_end": 563,
+      "policy_step_start": 76,
+      "policy_step_end": 93,
+      "phrase": "retracting robot arm",
+      "description": "retracting robot arm"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/push_pineapple/ep035/labels.npz b/push_pineapple/ep035/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..c2b619d2a02c29b68ef95d3f39dadc0f9aae996c
--- /dev/null
+++ b/push_pineapple/ep035/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:231a6be2e65f0343ae8c6a0e0839ecfd45a151472bc5eed2ce46e6eff50b8a4d
+size 3130
diff --git a/push_pineapple/ep036/example.md b/push_pineapple/ep036/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..a6b917810eac6e6a375a521e2dfdda0e868473f6
--- /dev/null
+++ b/push_pineapple/ep036/example.md
@@ -0,0 +1,33 @@
+# t005__ep036__push_pineapple
+
+- **suite / variant:** `` / ``
+- **instruction:** Push the pineapple off the stove
+- **success:** False
+- **control steps:** 355  (7.0s at 51 Hz)
+- **policy inferences:** 62  (replan_steps=6)
+- **video:** `push_pineapple/ep036/rollout.mp4`
+- **features:** `62` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | descending toward the stove |
+| 3.0–6.0 | 152–303 | 25–50 | lowering and pushing pineapple |
+| 6.0–7.0 | 304–354 | 50–59 | holding near the stove |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 2.066929133858268 s (00:02.1)
+- **frame / control_step:** 105 (policy_step=17, chunk_index=3)
+- **coarse onset:** 5.0 s (refined=True)
+- **mode:** object_displacement
+- **reason:** The robot made improper contact with the pineapple and stove unit, moving the stove tray itself rather than pushing the pineapple off of it.
+  Summary: improper contact; dragging stove tray instead of pineapple
+- **recovery:** Reposition the stove tray back onto the table, place the end effector on the side of the pineapple, and push the pineapple sideways off the stove surface.
+  Summary: reposition stove tray; push pineapple off stove
+- **prevention:** Approach the pineapple directly from the side with the gripper positioned to apply horizontal force strictly to the pineapple, avoiding catching or displacing the stove tray.
+  Summary: apply horizontal push to pineapple; avoid catching stove tray
+- **justification:** At approximately 5.0 seconds into the video, the robot end effector makes contact near the top of the pineapple and the stove tray, dragging the entire stove tray across the table instead of sliding the pineapple off the stove top.
diff --git a/push_pineapple/ep036/labels.json b/push_pineapple/ep036/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..d8937758fda048f6ee9c48d0f13d3823c0840592
--- /dev/null
+++ b/push_pineapple/ep036/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t005__ep036__push_pineapple",
+  "scene_variant": "",
+  "task_id": 5,
+  "task_file": "",
+  "instruction": "Push the pineapple off the stove",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:46:29+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9921259842519685,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "descending toward the stove",
+      "description": "descending toward the stove"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9921259842519685,
+      "t_end_sec": 5.984251968503937,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "lowering and pushing pineapple",
+      "description": "lowering and pushing pineapple"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.984251968503937,
+      "t_end_sec": 6.988188976377953,
+      "control_step_start": 304,
+      "control_step_end": 354,
+      "policy_step_start": 50,
+      "policy_step_end": 59,
+      "phrase": "holding near the stove",
+      "description": "holding near the stove"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "object_displacement",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 2.066929133858268,
+    "vlm_failure_onset_timestamp": "00:02.1",
+    "vlm_failure_onset_frame": 105,
+    "vlm_failure_onset_step": 105,
+    "vlm_coarse_failure_onset_seconds": 5.0,
+    "vlm_failure_window_start_seconds": 4.5,
+    "vlm_failure_window_end_seconds": 5.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot made improper contact with the pineapple and stove unit, moving the stove tray itself rather than pushing the pineapple off of it.",
+    "vlm_failure_reason_summary": "improper contact; dragging stove tray instead of pineapple",
+    "vlm_recovery_action": "Reposition the stove tray back onto the table, place the end effector on the side of the pineapple, and push the pineapple sideways off the stove surface.",
+    "vlm_recovery_action_summary": "reposition stove tray; push pineapple off stove",
+    "vlm_prevention_action": "Approach the pineapple directly from the side with the gripper positioned to apply horizontal force strictly to the pineapple, avoiding catching or displacing the stove tray.",
+    "vlm_prevention_action_summary": "apply horizontal push to pineapple; avoid catching stove tray",
+    "vlm_justification": "At approximately 5.0 seconds into the video, the robot end effector makes contact near the top of the pineapple and the stove tray, dragging the entire stove tray across the table instead of sliding the pineapple off the stove top.",
+    "vlm_temporal_justification": "At this timestamp, the robotic end-effector pushes against the pineapple and stove assembly with improper force/angle, causing the entire stove tray to start sliding across the table surface instead of separating the pineapple from the stove.",
+    "vlm_video_fps": 50.8,
+    "vlm_video_num_frames": 355,
+    "vlm_video_duration_seconds": 6.968503937007874,
+    "vlm_usage": {
+      "input_tokens": 18234,
+      "output_tokens": 377,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"object_displacement\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 5.0,\n \"failure_window_start_seconds\": 4.5,\n \"failure_window_end_seconds\": 5.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot made improper contact with the pineapple and stove unit, moving the stove tray itself rather than pushing the pineapple off of it.\",\n \"failure_reason_summary\": \"improper contact; dragging stove tray instead of pineapple\",\n \"recovery_action\": \"Reposition the stove tray back onto the table, place the end effector on the side of the pineapple, and push the pineapple sideways off the stove surface.\",\n \"recovery_action_summary\": \"reposition stove tray; push pineapple off stove\",\n \"prevention_action\": \"Approach the pineapple directly from the side with the gripper positioned to apply horizontal force strictly to the pineapple, avoiding catching or displacing the stove tray.\",\n \"prevention_action_summary\": \"apply horizontal push to pineapple; avoid catching stove tray\",\n \"justification\": \"At approximately 5.0 seconds into the video, the robot end effector makes contact near the top of the pineapple and the stove tray, dragging the entire stove tray across the table instead of sliding the pineapple off the stove top.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 105,
+    "failure_sim_step": 105,
+    "failure_policy_step": 17,
+    "failure_chunk_index": 3,
+    "first_post_failure_policy_step": 18,
+    "failure_type": "object_displacement",
+    "correction_action": "Reposition the stove tray back onto the table, place the end effector on the side of the pineapple, and push the pineapple sideways off the stove surface."
+  }
+}
\ No newline at end of file
diff --git a/push_pineapple/ep036/labels.npz b/push_pineapple/ep036/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..fab128d5b4060f5e9eb208c525496ea0b4455341
--- /dev/null
+++ b/push_pineapple/ep036/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6e21db5b3ed9e2fd6da301c9033109e1dbd4791abebb9302dd496b050e8d6d56
+size 3835
diff --git a/push_pineapple/ep037/example.md b/push_pineapple/ep037/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..74414296638fed86602d6f9ec77014db60b79b85
--- /dev/null
+++ b/push_pineapple/ep037/example.md
@@ -0,0 +1,32 @@
+# t005__ep037__push_pineapple
+
+- **suite / variant:** `` / ``
+- **instruction:** Push the pineapple off the stove
+- **success:** False
+- **control steps:** 291  (5.7s at 51 Hz)
+- **policy inferences:** 50  (replan_steps=6)
+- **video:** `push_pineapple/ep037/rollout.mp4`
+- **features:** `50` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching toward pineapple on stove |
+| 3.0–5.7 | 152–290 | 25–48 | contacting and pushing pineapple |
+
+## Failure
+
+- **yes**
+- **onset type:** timeout
+- **onset:** 5.708661417322835 s (00:05.7)
+- **frame / control_step:** 290 (policy_step=48, chunk_index=2)
+- **coarse onset:** 5.71 s (refined=False)
+- **mode:** timeout_or_insufficient_progress
+- **reason:** The robot moved purposefully toward the pineapple and prepared to push it, but ran out of time before completing the pushing motion to knock it off the stove.
+  Summary: insufficient time; incomplete push motion
+- **recovery:** Continue extending the end-effector forward against the pineapple until it drops off the stove surface.
+  Summary: push pineapple off stove; complete task
+- **prevention:** Execute the approach and pushing trajectory faster and more directly to push the pineapple off the stove before the episode time limit.
+  Summary: move faster; push pineapple off stove
+- **justification:** The robot executes a direct approach to the pineapple on the stove and makes contact to begin pushing, but the 5.71-second time limit expires before the pineapple is pushed off. No mistakes or collisions occurred.
diff --git a/push_pineapple/ep037/labels.json b/push_pineapple/ep037/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..f3d0f727d52cf99aaaa7525b11484aaf33dc6ea0
--- /dev/null
+++ b/push_pineapple/ep037/labels.json
@@ -0,0 +1,87 @@
+{
+  "rollout_id": "t005__ep037__push_pineapple",
+  "scene_variant": "",
+  "task_id": 5,
+  "task_file": "",
+  "instruction": "Push the pineapple off the stove",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:46:45+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9921259842519685,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching toward pineapple on stove",
+      "description": "reaching toward pineapple on stove"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9921259842519685,
+      "t_end_sec": 5.728346456692914,
+      "control_step_start": 152,
+      "control_step_end": 290,
+      "policy_step_start": 25,
+      "policy_step_end": 48,
+      "phrase": "contacting and pushing pineapple",
+      "description": "contacting and pushing pineapple"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "timeout_or_insufficient_progress",
+    "vlm_failure_onset_type": "timeout",
+    "vlm_failure_onset_seconds": 5.708661417322835,
+    "vlm_failure_onset_timestamp": "00:05.7",
+    "vlm_failure_onset_frame": 290,
+    "vlm_failure_onset_step": 290,
+    "vlm_coarse_failure_onset_seconds": 5.71,
+    "vlm_failure_window_start_seconds": 5.0,
+    "vlm_failure_window_end_seconds": 5.71,
+    "vlm_temporal_refined": false,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot moved purposefully toward the pineapple and prepared to push it, but ran out of time before completing the pushing motion to knock it off the stove.",
+    "vlm_failure_reason_summary": "insufficient time; incomplete push motion",
+    "vlm_recovery_action": "Continue extending the end-effector forward against the pineapple until it drops off the stove surface.",
+    "vlm_recovery_action_summary": "push pineapple off stove; complete task",
+    "vlm_prevention_action": "Execute the approach and pushing trajectory faster and more directly to push the pineapple off the stove before the episode time limit.",
+    "vlm_prevention_action_summary": "move faster; push pineapple off stove",
+    "vlm_justification": "The robot executes a direct approach to the pineapple on the stove and makes contact to begin pushing, but the 5.71-second time limit expires before the pineapple is pushed off. No mistakes or collisions occurred.",
+    "vlm_temporal_justification": "No earlier failure event identified; using the final frame.",
+    "vlm_video_fps": 50.8,
+    "vlm_video_num_frames": 291,
+    "vlm_video_duration_seconds": 5.708661417322835,
+    "vlm_usage": {
+      "input_tokens": 1866,
+      "output_tokens": 280,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"timeout_or_insufficient_progress\",\n \"onset_type\": \"timeout\",\n \"failure_onset_seconds\": 5.71,\n \"failure_window_start_seconds\": 5.0,\n \"failure_window_end_seconds\": 5.71,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot moved purposefully toward the pineapple and prepared to push it, but ran out of time before completing the pushing motion to knock it off the stove.\",\n \"failure_reason_summary\": \"insufficient time; incomplete push motion\",\n \"recovery_action\": \"Continue extending the end-effector forward against the pineapple until it drops off the stove surface.\",\n \"recovery_action_summary\": \"push pineapple off stove; complete task\",\n \"prevention_action\": \"Execute the approach and pushing trajectory faster and more directly to push the pineapple off the stove before the episode time limit.\",\n \"prevention_action_summary\": \"move faster; push pineapple off stove\",\n \"justification\": \"The robot executes a direct approach to the pineapple on the stove and makes contact to begin pushing, but the 5.71-second time limit expires before the pineapple is pushed off. No mistakes or collisions occurred.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 290,
+    "failure_sim_step": 290,
+    "failure_policy_step": 48,
+    "failure_chunk_index": 2,
+    "first_post_failure_policy_step": 49,
+    "failure_type": "timeout_or_insufficient_progress",
+    "correction_action": "Continue extending the end-effector forward against the pineapple until it drops off the stove surface."
+  }
+}
\ No newline at end of file
diff --git a/push_pineapple/ep037/labels.npz b/push_pineapple/ep037/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..c29d2b299ff6faab54ac523c10244b5d1394ce44
--- /dev/null
+++ b/push_pineapple/ep037/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:80b5b66f5587846c8d3396c8930d6a1b1edfe857580def6f67df74dc555252e5
+size 3730
diff --git a/push_pineapple/ep038/example.md b/push_pineapple/ep038/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..5085da2545b189db3b457ef7bbffc16e3697db51
--- /dev/null
+++ b/push_pineapple/ep038/example.md
@@ -0,0 +1,32 @@
+# t005__ep038__push_pineapple
+
+- **suite / variant:** `` / ``
+- **instruction:** Push the pineapple off the stove
+- **success:** False
+- **control steps:** 267  (5.3s at 51 Hz)
+- **policy inferences:** 47  (replan_steps=6)
+- **video:** `push_pineapple/ep038/rollout.mp4`
+- **features:** `47` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching toward pineapple on stove |
+| 3.0–5.3 | 152–266 | 25–44 | approaching pineapple on stove plate |
+
+## Failure
+
+- **yes**
+- **onset type:** timeout
+- **onset:** 5.2465483234714005 s (00:05.2)
+- **frame / control_step:** 266 (policy_step=44, chunk_index=2)
+- **coarse onset:** 5.25 s (refined=False)
+- **mode:** timeout_or_insufficient_progress
+- **reason:** The robot moved smoothly toward the pineapple on the stove, but did not execute the pushing action quickly enough to complete the task before the time limit expired.
+  Summary: slow trajectory execution; timeout before task completion
+- **recovery:** Sweep the end-effector sideways against the pineapple to push it off the stove burner plate onto the table surface.
+  Summary: push pineapple off stove
+- **prevention:** Trajectory execution should be sped up so that the end-effector reaches the pineapple and slides it off the stove before timing out.
+  Summary: execute faster approach; push pineapple off stove plate
+- **justification:** The robot exhibits reasonable purposeful motion throughout the rollout without dropping objects, colliding incorrectly, or going off target, but runs out of time before fulfilling the predicate condition.
diff --git a/push_pineapple/ep038/labels.json b/push_pineapple/ep038/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..3f63549af7c933cfac89a088b8e4fa19c0d9c245
--- /dev/null
+++ b/push_pineapple/ep038/labels.json
@@ -0,0 +1,87 @@
+{
+  "rollout_id": "t005__ep038__push_pineapple",
+  "scene_variant": "",
+  "task_id": 5,
+  "task_file": "",
+  "instruction": "Push the pineapple off the stove",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:47:00+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9980276134122286,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching toward pineapple on stove",
+      "description": "reaching toward pineapple on stove"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9980276134122286,
+      "t_end_sec": 5.266272189349112,
+      "control_step_start": 152,
+      "control_step_end": 266,
+      "policy_step_start": 25,
+      "policy_step_end": 44,
+      "phrase": "approaching pineapple on stove plate",
+      "description": "approaching pineapple on stove plate"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "timeout_or_insufficient_progress",
+    "vlm_failure_onset_type": "timeout",
+    "vlm_failure_onset_seconds": 5.2465483234714005,
+    "vlm_failure_onset_timestamp": "00:05.2",
+    "vlm_failure_onset_frame": 266,
+    "vlm_failure_onset_step": 266,
+    "vlm_coarse_failure_onset_seconds": 5.25,
+    "vlm_failure_window_start_seconds": 4.75,
+    "vlm_failure_window_end_seconds": 5.25,
+    "vlm_temporal_refined": false,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot moved smoothly toward the pineapple on the stove, but did not execute the pushing action quickly enough to complete the task before the time limit expired.",
+    "vlm_failure_reason_summary": "slow trajectory execution; timeout before task completion",
+    "vlm_recovery_action": "Sweep the end-effector sideways against the pineapple to push it off the stove burner plate onto the table surface.",
+    "vlm_recovery_action_summary": "push pineapple off stove",
+    "vlm_prevention_action": "Trajectory execution should be sped up so that the end-effector reaches the pineapple and slides it off the stove before timing out.",
+    "vlm_prevention_action_summary": "execute faster approach; push pineapple off stove plate",
+    "vlm_justification": "The robot exhibits reasonable purposeful motion throughout the rollout without dropping objects, colliding incorrectly, or going off target, but runs out of time before fulfilling the predicate condition.",
+    "vlm_temporal_justification": "No earlier failure event identified; using the final frame.",
+    "vlm_video_fps": 50.7,
+    "vlm_video_num_frames": 267,
+    "vlm_video_duration_seconds": 5.2465483234714005,
+    "vlm_usage": {
+      "input_tokens": 1803,
+      "output_tokens": 275,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"timeout_or_insufficient_progress\",\n \"onset_type\": \"timeout\",\n \"failure_onset_seconds\": 5.25,\n \"failure_window_start_seconds\": 4.75,\n \"failure_window_end_seconds\": 5.25,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot moved smoothly toward the pineapple on the stove, but did not execute the pushing action quickly enough to complete the task before the time limit expired.\",\n \"failure_reason_summary\": \"slow trajectory execution; timeout before task completion\",\n \"recovery_action\": \"Sweep the end-effector sideways against the pineapple to push it off the stove burner plate onto the table surface.\",\n \"recovery_action_summary\": \"push pineapple off stove\",\n \"prevention_action\": \"Trajectory execution should be sped up so that the end-effector reaches the pineapple and slides it off the stove before timing out.\",\n \"prevention_action_summary\": \"execute faster approach; push pineapple off stove plate\",\n \"justification\": \"The robot exhibits reasonable purposeful motion throughout the rollout without dropping objects, colliding incorrectly, or going off target, but runs out of time before fulfilling the predicate condition.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 266,
+    "failure_sim_step": 266,
+    "failure_policy_step": 44,
+    "failure_chunk_index": 2,
+    "first_post_failure_policy_step": 45,
+    "failure_type": "timeout_or_insufficient_progress",
+    "correction_action": "Sweep the end-effector sideways against the pineapple to push it off the stove burner plate onto the table surface."
+  }
+}
\ No newline at end of file
diff --git a/push_pineapple/ep038/labels.npz b/push_pineapple/ep038/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..f007425981adb8919b776757f42a41cddd6a74b1
--- /dev/null
+++ b/push_pineapple/ep038/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:96a2f81fd37b6013f7de2b37a04e26c238e7ee78259e1e93ba0f6c7be1f013d6
+size 3770
diff --git a/push_pineapple/ep039/example.md b/push_pineapple/ep039/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..db988fc3e1dfac46d721e37462778b4165ff8da3
--- /dev/null
+++ b/push_pineapple/ep039/example.md
@@ -0,0 +1,32 @@
+# t005__ep039__push_pineapple
+
+- **suite / variant:** `` / ``
+- **instruction:** Push the pineapple off the stove
+- **success:** False
+- **control steps:** 257  (5.1s at 50 Hz)
+- **policy inferences:** 45  (replan_steps=6)
+- **video:** `push_pineapple/ep039/rollout.mp4`
+- **features:** `45` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | lowering end-effector toward pineapple |
+| 3.0–5.1 | 152–256 | 25–42 | hovering near pineapple; remaining idle |
+
+## Failure
+
+- **yes**
+- **onset type:** operator_intervention
+- **onset:** 4.158415841584159 s (00:04.2)
+- **frame / control_step:** 210 (policy_step=35, chunk_index=0)
+- **coarse onset:** 4.0 s (refined=True)
+- **mode:** stuck_or_no_progress
+- **reason:** The robot arm descended towards the pineapple on the stove but froze and made no further pushing motion, leaving the pineapple resting on the stove.
+  Summary: stuck near pineapple; failing to push object
+- **recovery:** Reposition the end-effector beside the pineapple and apply a horizontal pushing force until the pineapple slides off the stove surface.
+  Summary: re-align end-effector; push pineapple off stove
+- **prevention:** Approach the pineapple from the side at stove level and execute a continuous lateral pushing motion to clear it off the stove.
+  Summary: approach pineapple from side; push off stove directly
+- **justification:** Around 3.8 to 4.0 seconds into the trajectory, the robot reaches the pineapple but stops making useful progress, failing to carry out the pushing motion required to satisfy the goal predicate.
diff --git a/push_pineapple/ep039/labels.json b/push_pineapple/ep039/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..48b486fa3ae0a397c150247a58bc9a0df84dce17
--- /dev/null
+++ b/push_pineapple/ep039/labels.json
@@ -0,0 +1,87 @@
+{
+  "rollout_id": "t005__ep039__push_pineapple",
+  "scene_variant": "",
+  "task_id": 5,
+  "task_file": "",
+  "instruction": "Push the pineapple off the stove",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:47:31+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00990099009901,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "lowering end-effector toward pineapple",
+      "description": "lowering end-effector toward pineapple"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00990099009901,
+      "t_end_sec": 5.089108910891089,
+      "control_step_start": 152,
+      "control_step_end": 256,
+      "policy_step_start": 25,
+      "policy_step_end": 42,
+      "phrase": "hovering near pineapple; remaining idle",
+      "description": "hovering near pineapple; remaining idle"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "stuck_or_no_progress",
+    "vlm_failure_onset_type": "operator_intervention",
+    "vlm_failure_onset_seconds": 4.158415841584159,
+    "vlm_failure_onset_timestamp": "00:04.2",
+    "vlm_failure_onset_frame": 210,
+    "vlm_failure_onset_step": 210,
+    "vlm_coarse_failure_onset_seconds": 4.0,
+    "vlm_failure_window_start_seconds": 3.5,
+    "vlm_failure_window_end_seconds": 4.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot arm descended towards the pineapple on the stove but froze and made no further pushing motion, leaving the pineapple resting on the stove.",
+    "vlm_failure_reason_summary": "stuck near pineapple; failing to push object",
+    "vlm_recovery_action": "Reposition the end-effector beside the pineapple and apply a horizontal pushing force until the pineapple slides off the stove surface.",
+    "vlm_recovery_action_summary": "re-align end-effector; push pineapple off stove",
+    "vlm_prevention_action": "Approach the pineapple from the side at stove level and execute a continuous lateral pushing motion to clear it off the stove.",
+    "vlm_prevention_action_summary": "approach pineapple from side; push off stove directly",
+    "vlm_justification": "Around 3.8 to 4.0 seconds into the trajectory, the robot reaches the pineapple but stops making useful progress, failing to carry out the pushing motion required to satisfy the goal predicate.",
+    "vlm_temporal_justification": "The robot arm completes its descent and settles onto the surface next to the pineapple around 02:40 (160 seconds). From this frame onward, it freezes completely and makes no attempt to push the pineapple, resulting in stuck/no progress failure.",
+    "vlm_video_fps": 50.5,
+    "vlm_video_num_frames": 257,
+    "vlm_video_duration_seconds": 5.069306930693069,
+    "vlm_usage": {
+      "input_tokens": 15211,
+      "output_tokens": 372,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"stuck_or_no_progress\",\n \"onset_type\": \"operator_intervention\",\n \"failure_onset_seconds\": 4.0,\n \"failure_window_start_seconds\": 3.5,\n \"failure_window_end_seconds\": 4.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot arm descended towards the pineapple on the stove but froze and made no further pushing motion, leaving the pineapple resting on the stove.\",\n \"failure_reason_summary\": \"stuck near pineapple; failing to push object\",\n \"recovery_action\": \"Reposition the end-effector beside the pineapple and apply a horizontal pushing force until the pineapple slides off the stove surface.\",\n \"recovery_action_summary\": \"re-align end-effector; push pineapple off stove\",\n \"prevention_action\": \"Approach the pineapple from the side at stove level and execute a continuous lateral pushing motion to clear it off the stove.\",\n \"prevention_action_summary\": \"approach pineapple from side; push off stove directly\",\n \"justification\": \"Around 3.8 to 4.0 seconds into the trajectory, the robot reaches the pineapple but stops making useful progress, failing to carry out the pushing motion required to satisfy the goal predicate.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 210,
+    "failure_sim_step": 210,
+    "failure_policy_step": 35,
+    "failure_chunk_index": 0,
+    "first_post_failure_policy_step": 36,
+    "failure_type": "stuck_or_no_progress",
+    "correction_action": "Reposition the end-effector beside the pineapple and apply a horizontal pushing force until the pineapple slides off the stove surface."
+  }
+}
\ No newline at end of file
diff --git a/push_pineapple/ep039/labels.npz b/push_pineapple/ep039/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..6a0ffee68ff127b5d4e08fe4f4c4df7c720bfeb3
--- /dev/null
+++ b/push_pineapple/ep039/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c906e2fd0a71b388f727c5a46cf6b56934a6d195a750be2a42e9f9ef299b85ee
+size 3760
diff --git a/push_pineapple/ep040/example.md b/push_pineapple/ep040/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..23a643f5054b4889f3c2bbbfd94540f870415e74
--- /dev/null
+++ b/push_pineapple/ep040/example.md
@@ -0,0 +1,33 @@
+# t005__ep040__push_pineapple
+
+- **suite / variant:** `` / ``
+- **instruction:** Push the pineapple off the stove
+- **success:** False
+- **control steps:** 316  (6.3s at 50 Hz)
+- **policy inferences:** 55  (replan_steps=6)
+- **video:** `push_pineapple/ep040/rollout.mp4`
+- **features:** `55` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | lowering gripper toward pineapple |
+| 3.0–6.0 | 151–301 | 25–50 | sliding pineapple along cooktop |
+| 6.0–6.3 | 302–315 | 50–52 | idle above stove cooktop |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 3.0952380952380953 s (00:03.1)
+- **frame / control_step:** 156 (policy_step=26, chunk_index=0)
+- **coarse onset:** 3.8 s (refined=True)
+- **mode:** stuck_or_no_progress
+- **reason:** The robot made contact with the pineapple but only slid it around on top of the stove cooktop rather than pushing it off the edge.
+  Summary: pineapple slid on cooktop; object not pushed off stove
+- **recovery:** Re-align the end-effector behind the pineapple and apply forward pushing force to slide it completely off the stove cooktop.
+  Summary: re-align end-effector; push pineapple off cooktop edge
+- **prevention:** Approach the pineapple from a trajectory aligned with the edge of the stove and push through until the pineapple falls off the stove surface.
+  Summary: align trajectory with cooktop edge; execute full pushing motion
+- **justification:** Around 3.8 seconds, the robot arm makes contact with the pineapple on the stove. Instead of pushing it forward off the stove, the arm slides it sideways across the cooktop and then retracts, leaving the pineapple entirely on the stove surface.
diff --git a/push_pineapple/ep040/labels.json b/push_pineapple/ep040/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..e77a54a86e11386a8dd2268f55ad15a2a35d2dde
--- /dev/null
+++ b/push_pineapple/ep040/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t005__ep040__push_pineapple",
+  "scene_variant": "",
+  "task_id": 5,
+  "task_file": "",
+  "instruction": "Push the pineapple off the stove",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:48:12+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.996031746031746,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "lowering gripper toward pineapple",
+      "description": "lowering gripper toward pineapple"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.996031746031746,
+      "t_end_sec": 5.992063492063492,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "sliding pineapple along cooktop",
+      "description": "sliding pineapple along cooktop"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.992063492063492,
+      "t_end_sec": 6.26984126984127,
+      "control_step_start": 302,
+      "control_step_end": 315,
+      "policy_step_start": 50,
+      "policy_step_end": 52,
+      "phrase": "idle above stove cooktop",
+      "description": "idle above stove cooktop"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "stuck_or_no_progress",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 3.0952380952380953,
+    "vlm_failure_onset_timestamp": "00:03.1",
+    "vlm_failure_onset_frame": 156,
+    "vlm_failure_onset_step": 156,
+    "vlm_coarse_failure_onset_seconds": 3.8,
+    "vlm_failure_window_start_seconds": 3.2,
+    "vlm_failure_window_end_seconds": 4.8,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot made contact with the pineapple but only slid it around on top of the stove cooktop rather than pushing it off the edge.",
+    "vlm_failure_reason_summary": "pineapple slid on cooktop; object not pushed off stove",
+    "vlm_recovery_action": "Re-align the end-effector behind the pineapple and apply forward pushing force to slide it completely off the stove cooktop.",
+    "vlm_recovery_action_summary": "re-align end-effector; push pineapple off cooktop edge",
+    "vlm_prevention_action": "Approach the pineapple from a trajectory aligned with the edge of the stove and push through until the pineapple falls off the stove surface.",
+    "vlm_prevention_action_summary": "align trajectory with cooktop edge; execute full pushing motion",
+    "vlm_justification": "Around 3.8 seconds, the robot arm makes contact with the pineapple on the stove. Instead of pushing it forward off the stove, the arm slides it sideways across the cooktop and then retracts, leaving the pineapple entirely on the stove surface.",
+    "vlm_temporal_justification": "At 01:55 (frame 115 of the original rollout), the robot gripper makes initial contact with the top of the pineapple and presses down vertically onto it rather than positioning behind it to push it off the stove, causing it to merely crush and slide around on the cooktop surface.",
+    "vlm_video_fps": 50.4,
+    "vlm_video_num_frames": 316,
+    "vlm_video_duration_seconds": 6.25,
+    "vlm_usage": {
+      "input_tokens": 19560,
+      "output_tokens": 403,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"stuck_or_no_progress\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 3.8,\n \"failure_window_start_seconds\": 3.2,\n \"failure_window_end_seconds\": 4.8,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot made contact with the pineapple but only slid it around on top of the stove cooktop rather than pushing it off the edge.\",\n \"failure_reason_summary\": \"pineapple slid on cooktop; object not pushed off stove\",\n \"recovery_action\": \"Re-align the end-effector behind the pineapple and apply forward pushing force to slide it completely off the stove cooktop.\",\n \"recovery_action_summary\": \"re-align end-effector; push pineapple off cooktop edge\",\n \"prevention_action\": \"Approach the pineapple from a trajectory aligned with the edge of the stove and push through until the pineapple falls off the stove surface.\",\n \"prevention_action_summary\": \"align trajectory with cooktop edge; execute full pushing motion\",\n \"justification\": \"Around 3.8 seconds, the robot arm makes contact with the pineapple on the stove. Instead of pushing it forward off the stove, the arm slides it sideways across the cooktop and then retracts, leaving the pineapple entirely on the stove surface.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 156,
+    "failure_sim_step": 156,
+    "failure_policy_step": 26,
+    "failure_chunk_index": 0,
+    "first_post_failure_policy_step": 27,
+    "failure_type": "stuck_or_no_progress",
+    "correction_action": "Re-align the end-effector behind the pineapple and apply forward pushing force to slide it completely off the stove cooktop."
+  }
+}
\ No newline at end of file
diff --git a/push_pineapple/ep040/labels.npz b/push_pineapple/ep040/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..c02a739fed200f9b0a971c3cf24da8948962fcf0
--- /dev/null
+++ b/push_pineapple/ep040/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f0b53469183965c17009d22ee732ae1eb7f424935a7cdb28478b3572367e7076
+size 3812
diff --git a/push_pineapple/ep041/example.md b/push_pineapple/ep041/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..82454a371cdc96e3653af317e58c23e657bae756
--- /dev/null
+++ b/push_pineapple/ep041/example.md
@@ -0,0 +1,33 @@
+# t005__ep041__push_pineapple
+
+- **suite / variant:** `` / ``
+- **instruction:** Push the pineapple off the stove
+- **success:** False
+- **control steps:** 352  (6.9s at 51 Hz)
+- **policy inferences:** 61  (replan_steps=6)
+- **video:** `push_pineapple/ep041/rollout.mp4`
+- **features:** `61` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | lowering gripper toward pineapple |
+| 3.0–6.0 | 152–303 | 25–50 | wiggling gripper over pineapple |
+| 6.0–6.9 | 304–351 | 50–58 | idle over pineapple |
+
+## Failure
+
+- **yes**
+- **onset type:** operator_intervention
+- **onset:** 3.1692913385826773 s (00:03.2)
+- **frame / control_step:** 161 (policy_step=26, chunk_index=5)
+- **coarse onset:** 3.5 s (refined=True)
+- **mode:** stuck_or_no_progress
+- **reason:** The robot reached the pineapple on the stove tray but became stuck hovering and wiggling over it without applying effective horizontal force to push it off.
+  Summary: stuck above object; failing to push horizontally
+- **recovery:** Reposition the gripper to the side of the pineapple and apply a direct horizontal push until the pineapple slides completely off the stove tray.
+  Summary: reposition gripper; push pineapple off stove
+- **prevention:** Approach the pineapple from a side trajectory rather than pressing straight down, ensuring clear horizontal contact to sweep it off the stove.
+  Summary: approach from side; sweep pineapple off stove
+- **justification:** Around 3.5 seconds into the rollout, the gripper reaches the pineapple on the stove tray but ceases forward progress, remaining stuck hovering and oscillating in place without pushing the item off the stove.
diff --git a/push_pineapple/ep041/labels.json b/push_pineapple/ep041/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..370fc0c9821f5fb816630691d32ec8acfcc0dec0
--- /dev/null
+++ b/push_pineapple/ep041/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t005__ep041__push_pineapple",
+  "scene_variant": "",
+  "task_id": 5,
+  "task_file": "",
+  "instruction": "Push the pineapple off the stove",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:48:43+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9921259842519685,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "lowering gripper toward pineapple",
+      "description": "lowering gripper toward pineapple"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9921259842519685,
+      "t_end_sec": 5.984251968503937,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "wiggling gripper over pineapple",
+      "description": "wiggling gripper over pineapple"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.984251968503937,
+      "t_end_sec": 6.929133858267717,
+      "control_step_start": 304,
+      "control_step_end": 351,
+      "policy_step_start": 50,
+      "policy_step_end": 58,
+      "phrase": "idle over pineapple",
+      "description": "idle over pineapple"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "stuck_or_no_progress",
+    "vlm_failure_onset_type": "operator_intervention",
+    "vlm_failure_onset_seconds": 3.1692913385826773,
+    "vlm_failure_onset_timestamp": "00:03.2",
+    "vlm_failure_onset_frame": 161,
+    "vlm_failure_onset_step": 161,
+    "vlm_coarse_failure_onset_seconds": 3.5,
+    "vlm_failure_window_start_seconds": 3.0,
+    "vlm_failure_window_end_seconds": 4.0,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot reached the pineapple on the stove tray but became stuck hovering and wiggling over it without applying effective horizontal force to push it off.",
+    "vlm_failure_reason_summary": "stuck above object; failing to push horizontally",
+    "vlm_recovery_action": "Reposition the gripper to the side of the pineapple and apply a direct horizontal push until the pineapple slides completely off the stove tray.",
+    "vlm_recovery_action_summary": "reposition gripper; push pineapple off stove",
+    "vlm_prevention_action": "Approach the pineapple from a side trajectory rather than pressing straight down, ensuring clear horizontal contact to sweep it off the stove.",
+    "vlm_prevention_action_summary": "approach from side; sweep pineapple off stove",
+    "vlm_justification": "Around 3.5 seconds into the rollout, the gripper reaches the pineapple on the stove tray but ceases forward progress, remaining stuck hovering and oscillating in place without pushing the item off the stove.",
+    "vlm_temporal_justification": " around second 02:15 (frame 161), the robot arm completes its descent onto the pineapple and becomes completely stuck hovering and wiggling in place, making no further progress toward pushing the pineapple off the stove tray.",
+    "vlm_video_fps": 50.8,
+    "vlm_video_num_frames": 352,
+    "vlm_video_duration_seconds": 6.909448818897638,
+    "vlm_usage": {
+      "input_tokens": 21513,
+      "output_tokens": 368,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"stuck_or_no_progress\",\n \"onset_type\": \"operator_intervention\",\n \"failure_onset_seconds\": 3.5,\n \"failure_window_start_seconds\": 3.0,\n \"failure_window_end_seconds\": 4.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot reached the pineapple on the stove tray but became stuck hovering and wiggling over it without applying effective horizontal force to push it off.\",\n \"failure_reason_summary\": \"stuck above object; failing to push horizontally\",\n \"recovery_action\": \"Reposition the gripper to the side of the pineapple and apply a direct horizontal push until the pineapple slides completely off the stove tray.\",\n \"recovery_action_summary\": \"reposition gripper; push pineapple off stove\",\n \"prevention_action\": \"Approach the pineapple from a side trajectory rather than pressing straight down, ensuring clear horizontal contact to sweep it off the stove.\",\n \"prevention_action_summary\": \"approach from side; sweep pineapple off stove\",\n \"justification\": \"Around 3.5 seconds into the rollout, the gripper reaches the pineapple on the stove tray but ceases forward progress, remaining stuck hovering and oscillating in place without pushing the item off the stove.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 161,
+    "failure_sim_step": 161,
+    "failure_policy_step": 26,
+    "failure_chunk_index": 5,
+    "first_post_failure_policy_step": 27,
+    "failure_type": "stuck_or_no_progress",
+    "correction_action": "Reposition the gripper to the side of the pineapple and apply a direct horizontal push until the pineapple slides completely off the stove tray."
+  }
+}
\ No newline at end of file
diff --git a/push_pineapple/ep041/labels.npz b/push_pineapple/ep041/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..00c42bb8df5699d15ea48e3ad6d4867aa5b3dd9f
--- /dev/null
+++ b/push_pineapple/ep041/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:55ede7ed68001fa23de562a23a11a3dabec29970ce42d7dd52981a83128bfea3
+size 3807
diff --git a/push_pineapple/ep042/example.md b/push_pineapple/ep042/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..97885edb13328be49fafaf5c02abd4ddd8129765
--- /dev/null
+++ b/push_pineapple/ep042/example.md
@@ -0,0 +1,33 @@
+# t005__ep042__push_pineapple
+
+- **suite / variant:** `` / ``
+- **instruction:** Push the pineapple off the stove
+- **success:** False
+- **control steps:** 446  (8.8s at 51 Hz)
+- **policy inferences:** 76  (replan_steps=6)
+- **video:** `push_pineapple/ep042/rollout.mp4`
+- **features:** `76` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | lowering gripper toward pineapple |
+| 3.0–6.0 | 152–303 | 25–50 | grasping top of pineapple |
+| 6.0–8.8 | 304–445 | 50–74 | holding position above stove |
+
+## Failure
+
+- **yes**
+- **onset type:** operator_intervention
+- **onset:** 3.4645669291338583 s (00:03.5)
+- **frame / control_step:** 176 (policy_step=29, chunk_index=2)
+- **coarse onset:** 3.8 s (refined=True)
+- **mode:** stuck_or_no_progress
+- **reason:** The robot descended onto the pineapple on the stove burner but failed to execute a horizontal pushing motion, remaining stationary over the object for the rest of the rollout.
+  Summary: robot stuck on pineapple; no horizontal pushing motion
+- **recovery:** Raise the gripper slightly and execute a forward/lateral push to slide the pineapple off the stove surface.
+  Summary: lift end effector; push pineapple off stove
+- **prevention:** Maintain an appropriate offset height and execute a smooth horizontal sweeping trajectory across the stove burner to push the pineapple off.
+  Summary: approach from side; sweep horizontally across burner
+- **justification:** At around 3.8 seconds, the robot arm contacts the top of the pineapple on the stove, but stops advancing and fails to perform the required pushing trajectory, remaining motionless over the object until time runs out.
diff --git a/push_pineapple/ep042/labels.json b/push_pineapple/ep042/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..2b918b58a7aaa237f2aa8c84596670776ba1a035
--- /dev/null
+++ b/push_pineapple/ep042/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t005__ep042__push_pineapple",
+  "scene_variant": "",
+  "task_id": 5,
+  "task_file": "",
+  "instruction": "Push the pineapple off the stove",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:49:19+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9921259842519685,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "lowering gripper toward pineapple",
+      "description": "lowering gripper toward pineapple"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9921259842519685,
+      "t_end_sec": 5.984251968503937,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "grasping top of pineapple",
+      "description": "grasping top of pineapple"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.984251968503937,
+      "t_end_sec": 8.779527559055119,
+      "control_step_start": 304,
+      "control_step_end": 445,
+      "policy_step_start": 50,
+      "policy_step_end": 74,
+      "phrase": "holding position above stove",
+      "description": "holding position above stove"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "stuck_or_no_progress",
+    "vlm_failure_onset_type": "operator_intervention",
+    "vlm_failure_onset_seconds": 3.4645669291338583,
+    "vlm_failure_onset_timestamp": "00:03.5",
+    "vlm_failure_onset_frame": 176,
+    "vlm_failure_onset_step": 176,
+    "vlm_coarse_failure_onset_seconds": 3.8,
+    "vlm_failure_window_start_seconds": 3.3,
+    "vlm_failure_window_end_seconds": 4.3,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot descended onto the pineapple on the stove burner but failed to execute a horizontal pushing motion, remaining stationary over the object for the rest of the rollout.",
+    "vlm_failure_reason_summary": "robot stuck on pineapple; no horizontal pushing motion",
+    "vlm_recovery_action": "Raise the gripper slightly and execute a forward/lateral push to slide the pineapple off the stove surface.",
+    "vlm_recovery_action_summary": "lift end effector; push pineapple off stove",
+    "vlm_prevention_action": "Maintain an appropriate offset height and execute a smooth horizontal sweeping trajectory across the stove burner to push the pineapple off.",
+    "vlm_prevention_action_summary": "approach from side; sweep horizontally across burner",
+    "vlm_justification": "At around 3.8 seconds, the robot arm contacts the top of the pineapple on the stove, but stops advancing and fails to perform the required pushing trajectory, remaining motionless over the object until time runs out.",
+    "vlm_temporal_justification": "At 02:15 (135 seconds into the clip), the robot gripper completes its downward motion onto the pineapple and comes to a complete halt. It fails to execute the horizontal push and remains stationary for the remainder of the rollout.",
+    "vlm_video_fps": 50.8,
+    "vlm_video_num_frames": 446,
+    "vlm_video_duration_seconds": 8.75984251968504,
+    "vlm_usage": {
+      "input_tokens": 21641,
+      "output_tokens": 367,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"stuck_or_no_progress\",\n \"onset_type\": \"operator_intervention\",\n \"failure_onset_seconds\": 3.8,\n \"failure_window_start_seconds\": 3.3,\n \"failure_window_end_seconds\": 4.3,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot descended onto the pineapple on the stove burner but failed to execute a horizontal pushing motion, remaining stationary over the object for the rest of the rollout.\",\n \"failure_reason_summary\": \"robot stuck on pineapple; no horizontal pushing motion\",\n \"recovery_action\": \"Raise the gripper slightly and execute a forward/lateral push to slide the pineapple off the stove surface.\",\n \"recovery_action_summary\": \"lift end effector; push pineapple off stove\",\n \"prevention_action\": \"Maintain an appropriate offset height and execute a smooth horizontal sweeping trajectory across the stove burner to push the pineapple off.\",\n \"prevention_action_summary\": \"approach from side; sweep horizontally across burner\",\n \"justification\": \"At around 3.8 seconds, the robot arm contacts the top of the pineapple on the stove, but stops advancing and fails to perform the required pushing trajectory, remaining motionless over the object until time runs out.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 176,
+    "failure_sim_step": 176,
+    "failure_policy_step": 29,
+    "failure_chunk_index": 2,
+    "first_post_failure_policy_step": 30,
+    "failure_type": "stuck_or_no_progress",
+    "correction_action": "Raise the gripper slightly and execute a forward/lateral push to slide the pineapple off the stove surface."
+  }
+}
\ No newline at end of file
diff --git a/push_pineapple/ep042/labels.npz b/push_pineapple/ep042/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..011b04113ded4338700ac46d26524ba0a9a68e07
--- /dev/null
+++ b/push_pineapple/ep042/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:877cc58ec3c10a093ef11ae0394397ba5f66e0cee3ad57df45dea21a30a5ae53
+size 3838
diff --git a/push_pineapple/ep043/example.md b/push_pineapple/ep043/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..923fca9638848b76994a8bfb444780f37ce52d12
--- /dev/null
+++ b/push_pineapple/ep043/example.md
@@ -0,0 +1,35 @@
+# t005__ep043__push_pineapple
+
+- **suite / variant:** `` / ``
+- **instruction:** Push the pineapple off the stove
+- **success:** False
+- **control steps:** 630  (12.5s at 50 Hz)
+- **policy inferences:** 108  (replan_steps=6)
+- **video:** `push_pineapple/ep043/rollout.mp4`
+- **features:** `108` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | lowering arm toward stove |
+| 3.0–6.0 | 151–301 | 25–50 | approaching pineapple on stove |
+| 6.0–9.0 | 302–452 | 50–75 | lowering gripper onto pineapple |
+| 9.0–12.0 | 453–603 | 75–100 | pressing against pineapple without moving |
+| 12.0–12.5 | 604–629 | 100–104 | idle at pineapple |
+
+## Failure
+
+- **yes**
+- **onset type:** operator_intervention
+- **onset:** 7.698412698412699 s (00:07.7)
+- **frame / control_step:** 388 (policy_step=64, chunk_index=4)
+- **coarse onset:** 9.2 s (refined=True)
+- **mode:** stuck_or_no_progress
+- **reason:** The robot positioned its gripper near the pineapple on the stove but failed to execute a effective pushing motion, remaining stuck in contact with the object without moving it off the stove.
+  Summary: stuck near object; failing to push pineapple
+- **recovery:** Reposition the gripper behind or to the side of the pineapple and push it horizontally until it is completely off the stove surface.
+  Summary: reposition gripper; push pineapple off stove
+- **prevention:** Align the gripper properly with the pineapple and drive horizontally across the stove surface to push the pineapple off.
+  Summary: align gripper behind object; push off stove surface
+- **justification:** Around 9.2 seconds, the robot arm reaches the pineapple on the stove but makes no further progress in pushing it off, staying in place against the stove and pineapple until the rollout ends.
diff --git a/push_pineapple/ep043/labels.json b/push_pineapple/ep043/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..c89d6068bfe40214a500431231406bfd6068ce00
--- /dev/null
+++ b/push_pineapple/ep043/labels.json
@@ -0,0 +1,120 @@
+{
+  "rollout_id": "t005__ep043__push_pineapple",
+  "scene_variant": "",
+  "task_id": 5,
+  "task_file": "",
+  "instruction": "Push the pineapple off the stove",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:49:50+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.996031746031746,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "lowering arm toward stove",
+      "description": "lowering arm toward stove"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.996031746031746,
+      "t_end_sec": 5.992063492063492,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "approaching pineapple on stove",
+      "description": "approaching pineapple on stove"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.992063492063492,
+      "t_end_sec": 8.988095238095239,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "lowering gripper onto pineapple",
+      "description": "lowering gripper onto pineapple"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.988095238095239,
+      "t_end_sec": 11.984126984126984,
+      "control_step_start": 453,
+      "control_step_end": 603,
+      "policy_step_start": 75,
+      "policy_step_end": 100,
+      "phrase": "pressing against pineapple without moving",
+      "description": "pressing against pineapple without moving"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 11.984126984126984,
+      "t_end_sec": 12.5,
+      "control_step_start": 604,
+      "control_step_end": 629,
+      "policy_step_start": 100,
+      "policy_step_end": 104,
+      "phrase": "idle at pineapple",
+      "description": "idle at pineapple"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "stuck_or_no_progress",
+    "vlm_failure_onset_type": "operator_intervention",
+    "vlm_failure_onset_seconds": 7.698412698412699,
+    "vlm_failure_onset_timestamp": "00:07.7",
+    "vlm_failure_onset_frame": 388,
+    "vlm_failure_onset_step": 388,
+    "vlm_coarse_failure_onset_seconds": 9.2,
+    "vlm_failure_window_start_seconds": 9.0,
+    "vlm_failure_window_end_seconds": 12.48,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot positioned its gripper near the pineapple on the stove but failed to execute a effective pushing motion, remaining stuck in contact with the object without moving it off the stove.",
+    "vlm_failure_reason_summary": "stuck near object; failing to push pineapple",
+    "vlm_recovery_action": "Reposition the gripper behind or to the side of the pineapple and push it horizontally until it is completely off the stove surface.",
+    "vlm_recovery_action_summary": "reposition gripper; push pineapple off stove",
+    "vlm_prevention_action": "Align the gripper properly with the pineapple and drive horizontally across the stove surface to push the pineapple off.",
+    "vlm_prevention_action_summary": "align gripper behind object; push off stove surface",
+    "vlm_justification": "Around 9.2 seconds, the robot arm reaches the pineapple on the stove but makes no further progress in pushing it off, staying in place against the stove and pineapple until the rollout ends.",
+    "vlm_temporal_justification": "At clip second 75 (01:15), the robot finishes positioning its gripper against the object on the stove and becomes completely stationary, remaining stuck in contact without executing a pushing motion for the remainder of the rollout.",
+    "vlm_video_fps": 50.4,
+    "vlm_video_num_frames": 630,
+    "vlm_video_duration_seconds": 12.480158730158731,
+    "vlm_usage": {
+      "input_tokens": 21775,
+      "output_tokens": 380,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n  \"failure_mode\": \"stuck_or_no_progress\",\n  \"onset_type\": \"operator_intervention\",\n  \"failure_onset_seconds\": 9.2,\n  \"failure_window_start_seconds\": 9.0,\n  \"failure_window_end_seconds\": 12.48,\n  \"confidence\": \"high\",\n  \"failure_reason\": \"The robot positioned its gripper near the pineapple on the stove but failed to execute a effective pushing motion, remaining stuck in contact with the object without moving it off the stove.\",\n  \"failure_reason_summary\": \"stuck near object; failing to push pineapple\",\n  \"recovery_action\": \"Reposition the gripper behind or to the side of the pineapple and push it horizontally until it is completely off the stove surface.\",\n  \"recovery_action_summary\": \"reposition gripper; push pineapple off stove\",\n  \"prevention_action\": \"Align the gripper properly with the pineapple and drive horizontally across the stove surface to push the pineapple off.\",\n  \"prevention_action_summary\": \"align gripper behind object; push off stove surface\",\n  \"justification\": \"Around 9.2 seconds, the robot arm reaches the pineapple on the stove but makes no further progress in pushing it off, staying in place against the stove and pineapple until the rollout ends.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 388,
+    "failure_sim_step": 388,
+    "failure_policy_step": 64,
+    "failure_chunk_index": 4,
+    "first_post_failure_policy_step": 65,
+    "failure_type": "stuck_or_no_progress",
+    "correction_action": "Reposition the gripper behind or to the side of the pineapple and push it horizontally until it is completely off the stove surface."
+  }
+}
\ No newline at end of file
diff --git a/push_pineapple/ep043/labels.npz b/push_pineapple/ep043/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..450a4ed5f725ca6562d2363b5603fba87c4fb61b
--- /dev/null
+++ b/push_pineapple/ep043/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5cd13d95d7d5fdfdd3383b671d07b3917a400353bc82d4b92825252e790fe606
+size 3905
diff --git a/push_pineapple/ep044/example.md b/push_pineapple/ep044/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..39122f3fb846be2ef25570da7968fa85feec916d
--- /dev/null
+++ b/push_pineapple/ep044/example.md
@@ -0,0 +1,35 @@
+# t005__ep044__push_pineapple
+
+- **suite / variant:** `` / ``
+- **instruction:** Push the pineapple off the stove
+- **success:** False
+- **control steps:** 639  (12.7s at 50 Hz)
+- **policy inferences:** 111  (replan_steps=6)
+- **video:** `push_pineapple/ep044/rollout.mp4`
+- **features:** `111` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching toward pineapple on stove |
+| 3.0–6.0 | 152–303 | 25–50 | lowering gripper near pineapple |
+| 6.0–9.0 | 304–455 | 50–75 | idle next to pineapple |
+| 9.0–12.0 | 456–607 | 76–101 | idle / no progress |
+| 12.0–12.7 | 608–638 | 101–106 | remaining stationary |
+
+## Failure
+
+- **yes**
+- **onset type:** operator_intervention
+- **onset:** 4.01980198019802 s (00:04.0)
+- **frame / control_step:** 203 (policy_step=33, chunk_index=5)
+- **coarse onset:** 4.2 s (refined=True)
+- **mode:** stuck_or_no_progress
+- **reason:** The robot lowered its gripper onto the stove next to the pineapple but failed to execute a forward or lateral pushing stroke, remaining stuck in place until the end of the episode.
+  Summary: gripper stuck near pineapple; failure to push off stove
+- **recovery:** Reposition the end-effector behind the pineapple and apply continuous lateral force to slide the pineapple completely off the stove.
+  Summary: reposition end-effector; push pineapple off stove
+- **prevention:** Maintain smooth lateral forward trajectory after making contact with the pineapple instead of stopping motion near the stove surface.
+  Summary: apply continuous lateral movement; avoid stopping near object
+- **justification:** Around 4.2 seconds into the video, the end-effector reaches the pineapple but stops making any lateral progress to push it off the stove, remaining idle/stuck in the same pose for the remainder of the rollout.
diff --git a/push_pineapple/ep044/labels.json b/push_pineapple/ep044/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..ffd2f0574abd28b3cc40459ae6afc072c2295f01
--- /dev/null
+++ b/push_pineapple/ep044/labels.json
@@ -0,0 +1,120 @@
+{
+  "rollout_id": "t005__ep044__push_pineapple",
+  "scene_variant": "",
+  "task_id": 5,
+  "task_file": "",
+  "instruction": "Push the pineapple off the stove",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:50:22+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00990099009901,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching toward pineapple on stove",
+      "description": "reaching toward pineapple on stove"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00990099009901,
+      "t_end_sec": 6.01980198019802,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "lowering gripper near pineapple",
+      "description": "lowering gripper near pineapple"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01980198019802,
+      "t_end_sec": 9.029702970297029,
+      "control_step_start": 304,
+      "control_step_end": 455,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "idle next to pineapple",
+      "description": "idle next to pineapple"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.029702970297029,
+      "t_end_sec": 12.03960396039604,
+      "control_step_start": 456,
+      "control_step_end": 607,
+      "policy_step_start": 76,
+      "policy_step_end": 101,
+      "phrase": "idle / no progress",
+      "description": "idle / no progress"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 12.03960396039604,
+      "t_end_sec": 12.653465346534654,
+      "control_step_start": 608,
+      "control_step_end": 638,
+      "policy_step_start": 101,
+      "policy_step_end": 106,
+      "phrase": "remaining stationary",
+      "description": "remaining stationary"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "stuck_or_no_progress",
+    "vlm_failure_onset_type": "operator_intervention",
+    "vlm_failure_onset_seconds": 4.01980198019802,
+    "vlm_failure_onset_timestamp": "00:04.0",
+    "vlm_failure_onset_frame": 203,
+    "vlm_failure_onset_step": 203,
+    "vlm_coarse_failure_onset_seconds": 4.2,
+    "vlm_failure_window_start_seconds": 4.2,
+    "vlm_failure_window_end_seconds": 12.63,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot lowered its gripper onto the stove next to the pineapple but failed to execute a forward or lateral pushing stroke, remaining stuck in place until the end of the episode.",
+    "vlm_failure_reason_summary": "gripper stuck near pineapple; failure to push off stove",
+    "vlm_recovery_action": "Reposition the end-effector behind the pineapple and apply continuous lateral force to slide the pineapple completely off the stove.",
+    "vlm_recovery_action_summary": "reposition end-effector; push pineapple off stove",
+    "vlm_prevention_action": "Maintain smooth lateral forward trajectory after making contact with the pineapple instead of stopping motion near the stove surface.",
+    "vlm_prevention_action_summary": "apply continuous lateral movement; avoid stopping near object",
+    "vlm_justification": "Around 4.2 seconds into the video, the end-effector reaches the pineapple but stops making any lateral progress to push it off the stove, remaining idle/stuck in the same pose for the remainder of the rollout.",
+    "vlm_temporal_justification": "The robot completes its downward approach and lowers the gripper beside the pineapple at 02:22 (142 seconds into the clip). From 02:23 (143 seconds into the clip) onwards, all arm movement ceases completely and the robot remains stuck in place without attempting to push the pineapple.",
+    "vlm_video_fps": 50.5,
+    "vlm_video_num_frames": 639,
+    "vlm_video_duration_seconds": 12.633663366336634,
+    "vlm_usage": {
+      "input_tokens": 21898,
+      "output_tokens": 396,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"stuck_or_no_progress\",\n \"onset_type\": \"operator_intervention\",\n \"failure_onset_seconds\": 4.2,\n \"failure_window_start_seconds\": 4.2,\n \"failure_window_end_seconds\": 12.63,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot lowered its gripper onto the stove next to the pineapple but failed to execute a forward or lateral pushing stroke, remaining stuck in place until the end of the episode.\",\n \"failure_reason_summary\": \"gripper stuck near pineapple; failure to push off stove\",\n \"recovery_action\": \"Reposition the end-effector behind the pineapple and apply continuous lateral force to slide the pineapple completely off the stove.\",\n \"recovery_action_summary\": \"reposition end-effector; push pineapple off stove\",\n \"prevention_action\": \"Maintain smooth lateral forward trajectory after making contact with the pineapple instead of stopping motion near the stove surface.\",\n \"prevention_action_summary\": \"apply continuous lateral movement; avoid stopping near object\",\n \"justification\": \"Around 4.2 seconds into the video, the end-effector reaches the pineapple but stops making any lateral progress to push it off the stove, remaining idle/stuck in the same pose for the remainder of the rollout.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 203,
+    "failure_sim_step": 203,
+    "failure_policy_step": 33,
+    "failure_chunk_index": 5,
+    "first_post_failure_policy_step": 34,
+    "failure_type": "stuck_or_no_progress",
+    "correction_action": "Reposition the end-effector behind the pineapple and apply continuous lateral force to slide the pineapple completely off the stove."
+  }
+}
\ No newline at end of file
diff --git a/push_pineapple/ep044/labels.npz b/push_pineapple/ep044/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..c2d89b7a311f58914a57a1a2d680cc5aa2b6cc1a
--- /dev/null
+++ b/push_pineapple/ep044/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ffee0933e477ad0096d5febe8c6073dea6d5b60a2a427e64c0267b863d449995
+size 3943
diff --git a/push_pineapple/ep045/example.md b/push_pineapple/ep045/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..8229de6e61d615fb3b2f1c84ffa9b0935d2a6316
--- /dev/null
+++ b/push_pineapple/ep045/example.md
@@ -0,0 +1,35 @@
+# t005__ep045__push_pineapple
+
+- **suite / variant:** `` / ``
+- **instruction:** Push the pineapple off the stove
+- **success:** False
+- **control steps:** 756  (14.9s at 51 Hz)
+- **policy inferences:** 129  (replan_steps=6)
+- **video:** `push_pineapple/ep045/rollout.mp4`
+- **features:** `129` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | moving end-effector toward stove |
+| 3.0–6.0 | 152–303 | 25–50 | positioning gripper beside pineapple |
+| 6.0–9.0 | 304–455 | 50–75 | pushing pineapple across stove |
+| 9.0–12.0 | 456–607 | 76–101 | pushing pineapple onto left burner |
+| 12.0–14.9 | 608–755 | 101–125 | hovering above left burner |
+
+## Failure
+
+- **yes**
+- **onset type:** operator_intervention
+- **onset:** 12.642998027613412 s (00:12.6)
+- **frame / control_step:** 641 (policy_step=106, chunk_index=5)
+- **coarse onset:** 11.2 s (refined=True)
+- **mode:** stuck_or_no_progress
+- **reason:** The robot pushes the pineapple across the stove top but stops prematurely on the left burner, failing to push it completely off the stove surface.
+  Summary: incomplete push; pineapple left on stove
+- **recovery:** Re-engage the pineapple with the end-effector and push it further left until it falls off the stove surface onto the table.
+  Summary: push pineapple further left; clear off stove
+- **prevention:** Maintain a continuous pushing motion across the entire stove top surface until the pineapple is fully pushed off the edge.
+  Summary: push pineapple completely off stove surface; maintain continuous contact
+- **justification:** The robot successfully contacts the pineapple and pushes it leftward across the stove, but halts its motion around 11.2 seconds before the pineapple is pushed off the stove edge. It then hovers in place without completing the task.
diff --git a/push_pineapple/ep045/labels.json b/push_pineapple/ep045/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..a92d0768d31c4ea58d902ab9695eaac262fb38e6
--- /dev/null
+++ b/push_pineapple/ep045/labels.json
@@ -0,0 +1,120 @@
+{
+  "rollout_id": "t005__ep045__push_pineapple",
+  "scene_variant": "",
+  "task_id": 5,
+  "task_file": "",
+  "instruction": "Push the pineapple off the stove",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:50:59+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9980276134122286,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "moving end-effector toward stove",
+      "description": "moving end-effector toward stove"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9980276134122286,
+      "t_end_sec": 5.996055226824457,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "positioning gripper beside pineapple",
+      "description": "positioning gripper beside pineapple"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.996055226824457,
+      "t_end_sec": 8.994082840236686,
+      "control_step_start": 304,
+      "control_step_end": 455,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "pushing pineapple across stove",
+      "description": "pushing pineapple across stove"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.994082840236686,
+      "t_end_sec": 11.992110453648914,
+      "control_step_start": 456,
+      "control_step_end": 607,
+      "policy_step_start": 76,
+      "policy_step_end": 101,
+      "phrase": "pushing pineapple onto left burner",
+      "description": "pushing pineapple onto left burner"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 11.992110453648914,
+      "t_end_sec": 14.911242603550296,
+      "control_step_start": 608,
+      "control_step_end": 755,
+      "policy_step_start": 101,
+      "policy_step_end": 125,
+      "phrase": "hovering above left burner",
+      "description": "hovering above left burner"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "stuck_or_no_progress",
+    "vlm_failure_onset_type": "operator_intervention",
+    "vlm_failure_onset_seconds": 12.642998027613412,
+    "vlm_failure_onset_timestamp": "00:12.6",
+    "vlm_failure_onset_frame": 641,
+    "vlm_failure_onset_step": 641,
+    "vlm_coarse_failure_onset_seconds": 11.2,
+    "vlm_failure_window_start_seconds": 11.0,
+    "vlm_failure_window_end_seconds": 14.89,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot pushes the pineapple across the stove top but stops prematurely on the left burner, failing to push it completely off the stove surface.",
+    "vlm_failure_reason_summary": "incomplete push; pineapple left on stove",
+    "vlm_recovery_action": "Re-engage the pineapple with the end-effector and push it further left until it falls off the stove surface onto the table.",
+    "vlm_recovery_action_summary": "push pineapple further left; clear off stove",
+    "vlm_prevention_action": "Maintain a continuous pushing motion across the entire stove top surface until the pineapple is fully pushed off the edge.",
+    "vlm_prevention_action_summary": "push pineapple completely off stove surface; maintain continuous contact",
+    "vlm_justification": "The robot successfully contacts the pineapple and pushes it leftward across the stove, but halts its motion around 11.2 seconds before the pineapple is pushed off the stove edge. It then hovers in place without completing the task.",
+    "vlm_temporal_justification": "At 03:45 (225 seconds in the clip), the robot gripper has fully opened and released the fruit, coming to a complete stop while leaving the fruit sitting on the stove top rather than pushing it completely off.",
+    "vlm_video_fps": 50.7,
+    "vlm_video_num_frames": 756,
+    "vlm_video_duration_seconds": 14.891518737672582,
+    "vlm_usage": {
+      "input_tokens": 22021,
+      "output_tokens": 374,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"stuck_or_no_progress\",\n \"onset_type\": \"operator_intervention\",\n \"failure_onset_seconds\": 11.2,\n \"failure_window_start_seconds\": 11.0,\n \"failure_window_end_seconds\": 14.89,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot pushes the pineapple across the stove top but stops prematurely on the left burner, failing to push it completely off the stove surface.\",\n \"failure_reason_summary\": \"incomplete push; pineapple left on stove\",\n \"recovery_action\": \"Re-engage the pineapple with the end-effector and push it further left until it falls off the stove surface onto the table.\",\n \"recovery_action_summary\": \"push pineapple further left; clear off stove\",\n \"prevention_action\": \"Maintain a continuous pushing motion across the entire stove top surface until the pineapple is fully pushed off the edge.\",\n \"prevention_action_summary\": \"push pineapple completely off stove surface; maintain continuous contact\",\n \"justification\": \"The robot successfully contacts the pineapple and pushes it leftward across the stove, but halts its motion around 11.2 seconds before the pineapple is pushed off the stove edge. It then hovers in place without completing the task.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 641,
+    "failure_sim_step": 641,
+    "failure_policy_step": 106,
+    "failure_chunk_index": 5,
+    "first_post_failure_policy_step": 107,
+    "failure_type": "stuck_or_no_progress",
+    "correction_action": "Re-engage the pineapple with the end-effector and push it further left until it falls off the stove surface onto the table."
+  }
+}
\ No newline at end of file
diff --git a/push_pineapple/ep045/labels.npz b/push_pineapple/ep045/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..5d8b3a98be32b21e508227373550661e30ed0d4a
--- /dev/null
+++ b/push_pineapple/ep045/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2fa5ccafb51b3a1fd00d6060f409bcb314488338aff3e6cd9dab0b810b37106b
+size 3872
diff --git a/push_pineapple/ep046/example.md b/push_pineapple/ep046/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..b0b2d19f0ddce819cce58a836e27f097b9898602
--- /dev/null
+++ b/push_pineapple/ep046/example.md
@@ -0,0 +1,33 @@
+# t005__ep046__push_pineapple
+
+- **suite / variant:** `` / ``
+- **instruction:** Push the pineapple off the stove
+- **success:** False
+- **control steps:** 424  (8.4s at 50 Hz)
+- **policy inferences:** 73  (replan_steps=6)
+- **video:** `push_pineapple/ep046/rollout.mp4`
+- **features:** `73` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | lowering arm toward stove |
+| 3.0–6.0 | 151–301 | 25–50 | approaching and contacting pineapple |
+| 6.0–8.4 | 302–423 | 50–70 | holding position at pineapple |
+
+## Failure
+
+- **yes**
+- **onset type:** operator_intervention
+- **onset:** 3.8645418326693224 s (00:03.9)
+- **frame / control_step:** 194 (policy_step=32, chunk_index=2)
+- **coarse onset:** 5.8 s (refined=True)
+- **mode:** stuck_or_no_progress
+- **reason:** The robot reaches the pineapple on the stove top around 5.8 seconds but freezes upon making contact, failing to apply the lateral force necessary to push it off the stove.
+  Summary: arm freezes upon contact; pineapple remains on stove
+- **recovery:** Re-engage end-effector control to push the pineapple laterally until it moves completely off the stove plate onto the table.
+  Summary: apply lateral force; push pineapple off stove
+- **prevention:** Maintain a continuous lateral trajectory when approaching the pineapple to smoothly sweep it off the stove surface without stopping.
+  Summary: maintain sweeping trajectory; push object off stove
+- **justification:** The robot correctly identifies and approaches the pineapple on the stove around 5.8 seconds. However, once touching the object, it stops advancing and remains stationary on top of the stove until the video ends, making no progress toward pushing the object off the stove.
diff --git a/push_pineapple/ep046/labels.json b/push_pineapple/ep046/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..3ccf38a82f7c23c2591f22abf879aecf453e79a9
--- /dev/null
+++ b/push_pineapple/ep046/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t005__ep046__push_pineapple",
+  "scene_variant": "",
+  "task_id": 5,
+  "task_file": "",
+  "instruction": "Push the pineapple off the stove",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:51:34+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00796812749004,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "lowering arm toward stove",
+      "description": "lowering arm toward stove"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00796812749004,
+      "t_end_sec": 6.01593625498008,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "approaching and contacting pineapple",
+      "description": "approaching and contacting pineapple"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01593625498008,
+      "t_end_sec": 8.44621513944223,
+      "control_step_start": 302,
+      "control_step_end": 423,
+      "policy_step_start": 50,
+      "policy_step_end": 70,
+      "phrase": "holding position at pineapple",
+      "description": "holding position at pineapple"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "stuck_or_no_progress",
+    "vlm_failure_onset_type": "operator_intervention",
+    "vlm_failure_onset_seconds": 3.8645418326693224,
+    "vlm_failure_onset_timestamp": "00:03.9",
+    "vlm_failure_onset_frame": 194,
+    "vlm_failure_onset_step": 194,
+    "vlm_coarse_failure_onset_seconds": 5.8,
+    "vlm_failure_window_start_seconds": 5.3,
+    "vlm_failure_window_end_seconds": 6.8,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot reaches the pineapple on the stove top around 5.8 seconds but freezes upon making contact, failing to apply the lateral force necessary to push it off the stove.",
+    "vlm_failure_reason_summary": "arm freezes upon contact; pineapple remains on stove",
+    "vlm_recovery_action": "Re-engage end-effector control to push the pineapple laterally until it moves completely off the stove plate onto the table.",
+    "vlm_recovery_action_summary": "apply lateral force; push pineapple off stove",
+    "vlm_prevention_action": "Maintain a continuous lateral trajectory when approaching the pineapple to smoothly sweep it off the stove surface without stopping.",
+    "vlm_prevention_action_summary": "maintain sweeping trajectory; push object off stove",
+    "vlm_justification": "The robot correctly identifies and approaches the pineapple on the stove around 5.8 seconds. However, once touching the object, it stops advancing and remains stationary on top of the stove until the video ends, making no progress toward pushing the object off the stove.",
+    "vlm_temporal_justification": "At 00:54, the robot arm completes its downward movement and makes contact with the stovetop next to the pineapple. It then freezes completely and fails to perform any lateral pushing movement.",
+    "vlm_video_fps": 50.2,
+    "vlm_video_num_frames": 424,
+    "vlm_video_duration_seconds": 8.42629482071713,
+    "vlm_usage": {
+      "input_tokens": 20262,
+      "output_tokens": 371,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"stuck_or_no_progress\",\n \"onset_type\": \"operator_intervention\",\n \"failure_onset_seconds\": 5.8,\n \"failure_window_start_seconds\": 5.3,\n \"failure_window_end_seconds\": 6.8,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot reaches the pineapple on the stove top around 5.8 seconds but freezes upon making contact, failing to apply the lateral force necessary to push it off the stove.\",\n \"failure_reason_summary\": \"arm freezes upon contact; pineapple remains on stove\",\n \"recovery_action\": \"Re-engage end-effector control to push the pineapple laterally until it moves completely off the stove plate onto the table.\",\n \"recovery_action_summary\": \"apply lateral force; push pineapple off stove\",\n \"prevention_action\": \"Maintain a continuous lateral trajectory when approaching the pineapple to smoothly sweep it off the stove surface without stopping.\",\n \"prevention_action_summary\": \"maintain sweeping trajectory; push object off stove\",\n \"justification\": \"The robot correctly identifies and approaches the pineapple on the stove around 5.8 seconds. However, once touching the object, it stops advancing and remains stationary on top of the stove until the video ends, making no progress toward pushing the object off the stove.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 194,
+    "failure_sim_step": 194,
+    "failure_policy_step": 32,
+    "failure_chunk_index": 2,
+    "first_post_failure_policy_step": 33,
+    "failure_type": "stuck_or_no_progress",
+    "correction_action": "Re-engage end-effector control to push the pineapple laterally until it moves completely off the stove plate onto the table."
+  }
+}
\ No newline at end of file
diff --git a/push_pineapple/ep046/labels.npz b/push_pineapple/ep046/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..a48bf7050004f1fd0852e81a456228942509f81a
--- /dev/null
+++ b/push_pineapple/ep046/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9bf51529924407227dfdab30b3912d2dab5c2ad8713a962d15811f3856462a2f
+size 3818
diff --git a/push_pineapple/ep047/example.md b/push_pineapple/ep047/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..1629f9ba6461fe438da60bab095c9cfb4ea43472
--- /dev/null
+++ b/push_pineapple/ep047/example.md
@@ -0,0 +1,35 @@
+# t005__ep047__push_pineapple
+
+- **suite / variant:** `` / ``
+- **instruction:** Push the pineapple off the stove
+- **success:** False
+- **control steps:** 657  (13.0s at 51 Hz)
+- **policy inferences:** 112  (replan_steps=6)
+- **video:** `push_pineapple/ep047/rollout.mp4`
+- **features:** `112` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | moving arm toward stove |
+| 3.0–6.0 | 152–303 | 25–50 | lowering gripper near pineapple |
+| 6.0–9.0 | 304–455 | 50–75 | pushing pineapple across stove |
+| 9.0–12.0 | 456–607 | 76–101 | idle near pineapple |
+| 12.0–13.0 | 608–656 | 101–109 | remaining stationary |
+
+## Failure
+
+- **yes**
+- **onset type:** operator_intervention
+- **onset:** 7.8063241106719365 s (00:07.8)
+- **frame / control_step:** 395 (policy_step=65, chunk_index=5)
+- **coarse onset:** 7.8 s (refined=False)
+- **mode:** stuck_or_no_progress
+- **reason:** The robot pushed the pineapple across the stove surface but stopped before pushing it completely off the black stove plate, remaining frozen in place for the remainder of the episode.
+  Summary: stopped pushing prematurely; pineapple left on stove
+- **recovery:** Resume pushing the pineapple sideways or forward until it slides completely off the black stove plate onto the table surface.
+  Summary: continue pushing pineapple; push completely off stove
+- **prevention:** Maintain continuous pushing force against the pineapple until it is fully dislodged from the stove top plate before concluding the movement.
+  Summary: push pineapple continuously; ensure object leaves stove
+- **justification:** Around 00:07.8, the robot pushes the pineapple onto the left burner of the stove top plate and then stops moving entirely. It remains static until the rollout ends at 12.96s without completing the task requirement to push the object off the stove.
diff --git a/push_pineapple/ep047/labels.json b/push_pineapple/ep047/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..55639347896b54fcd732346ae68070cde617614b
--- /dev/null
+++ b/push_pineapple/ep047/labels.json
@@ -0,0 +1,120 @@
+{
+  "rollout_id": "t005__ep047__push_pineapple",
+  "scene_variant": "",
+  "task_id": 5,
+  "task_file": "",
+  "instruction": "Push the pineapple off the stove",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:52:07+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0039525691699605,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "moving arm toward stove",
+      "description": "moving arm toward stove"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0039525691699605,
+      "t_end_sec": 6.007905138339921,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "lowering gripper near pineapple",
+      "description": "lowering gripper near pineapple"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.007905138339921,
+      "t_end_sec": 9.011857707509881,
+      "control_step_start": 304,
+      "control_step_end": 455,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "pushing pineapple across stove",
+      "description": "pushing pineapple across stove"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.011857707509881,
+      "t_end_sec": 12.015810276679842,
+      "control_step_start": 456,
+      "control_step_end": 607,
+      "policy_step_start": 76,
+      "policy_step_end": 101,
+      "phrase": "idle near pineapple",
+      "description": "idle near pineapple"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 12.015810276679842,
+      "t_end_sec": 12.984189723320158,
+      "control_step_start": 608,
+      "control_step_end": 656,
+      "policy_step_start": 101,
+      "policy_step_end": 109,
+      "phrase": "remaining stationary",
+      "description": "remaining stationary"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "stuck_or_no_progress",
+    "vlm_failure_onset_type": "operator_intervention",
+    "vlm_failure_onset_seconds": 7.8063241106719365,
+    "vlm_failure_onset_timestamp": "00:07.8",
+    "vlm_failure_onset_frame": 395,
+    "vlm_failure_onset_step": 395,
+    "vlm_coarse_failure_onset_seconds": 7.8,
+    "vlm_failure_window_start_seconds": 7.8,
+    "vlm_failure_window_end_seconds": 12.96,
+    "vlm_temporal_refined": false,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot pushed the pineapple across the stove surface but stopped before pushing it completely off the black stove plate, remaining frozen in place for the remainder of the episode.",
+    "vlm_failure_reason_summary": "stopped pushing prematurely; pineapple left on stove",
+    "vlm_recovery_action": "Resume pushing the pineapple sideways or forward until it slides completely off the black stove plate onto the table surface.",
+    "vlm_recovery_action_summary": "continue pushing pineapple; push completely off stove",
+    "vlm_prevention_action": "Maintain continuous pushing force against the pineapple until it is fully dislodged from the stove top plate before concluding the movement.",
+    "vlm_prevention_action_summary": "push pineapple continuously; ensure object leaves stove",
+    "vlm_justification": "Around 00:07.8, the robot pushes the pineapple onto the left burner of the stove top plate and then stops moving entirely. It remains static until the rollout ends at 12.96s without completing the task requirement to push the object off the stove.",
+    "vlm_temporal_justification": "At 01:57, the robot arm completes its motion after pushing the pineapple partially across the stove top and comes to a complete halt, remaining frozen in place without making further progress for the remainder of the clip.",
+    "vlm_video_fps": 50.6,
+    "vlm_video_num_frames": 657,
+    "vlm_video_duration_seconds": 12.964426877470355,
+    "vlm_usage": {
+      "input_tokens": 21899,
+      "output_tokens": 384,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"stuck_or_no_progress\",\n \"onset_type\": \"operator_intervention\",\n \"failure_onset_seconds\": 7.8,\n \"failure_window_start_seconds\": 7.8,\n \"failure_window_end_seconds\": 12.96,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot pushed the pineapple across the stove surface but stopped before pushing it completely off the black stove plate, remaining frozen in place for the remainder of the episode.\",\n \"failure_reason_summary\": \"stopped pushing prematurely; pineapple left on stove\",\n \"recovery_action\": \"Resume pushing the pineapple sideways or forward until it slides completely off the black stove plate onto the table surface.\",\n \"recovery_action_summary\": \"continue pushing pineapple; push completely off stove\",\n \"prevention_action\": \"Maintain continuous pushing force against the pineapple until it is fully dislodged from the stove top plate before concluding the movement.\",\n \"prevention_action_summary\": \"push pineapple continuously; ensure object leaves stove\",\n \"justification\": \"Around 00:07.8, the robot pushes the pineapple onto the left burner of the stove top plate and then stops moving entirely. It remains static until the rollout ends at 12.96s without completing the task requirement to push the object off the stove.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 395,
+    "failure_sim_step": 395,
+    "failure_policy_step": 65,
+    "failure_chunk_index": 5,
+    "first_post_failure_policy_step": 66,
+    "failure_type": "stuck_or_no_progress",
+    "correction_action": "Resume pushing the pineapple sideways or forward until it slides completely off the black stove plate onto the table surface."
+  }
+}
\ No newline at end of file
diff --git a/push_pineapple/ep047/labels.npz b/push_pineapple/ep047/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..f0a4959071d3d5a220111225cb830d0c6800377a
--- /dev/null
+++ b/push_pineapple/ep047/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b8b7555a5e81207ea0d3e353d970edebedc665ad96db2e1e69be5fce8ca229c5
+size 3916
diff --git a/push_pineapple/ep048/example.md b/push_pineapple/ep048/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..71c6ff199cf4fbbdae6c201978308a710ead824b
--- /dev/null
+++ b/push_pineapple/ep048/example.md
@@ -0,0 +1,35 @@
+# t005__ep048__push_pineapple
+
+- **suite / variant:** `` / ``
+- **instruction:** Push the pineapple off the stove
+- **success:** False
+- **control steps:** 749  (15.0s at 50 Hz)
+- **policy inferences:** 129  (replan_steps=6)
+- **video:** `push_pineapple/ep048/rollout.mp4`
+- **features:** `129` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–149 | 0–24 | hovering near base |
+| 3.0–6.0 | 150–299 | 25–49 | idle near center table |
+| 6.0–9.0 | 300–449 | 50–74 | remaining stationary |
+| 9.0–12.0 | 450–599 | 75–99 | idle without moving |
+| 12.0–15.0 | 600–748 | 100–124 | hovering near initial position |
+
+## Failure
+
+- **yes**
+- **onset type:** operator_intervention
+- **onset:** 1.12 s (00:01.1)
+- **frame / control_step:** 56 (policy_step=9, chunk_index=2)
+- **coarse onset:** 2.5 s (refined=True)
+- **mode:** stuck_or_no_progress
+- **reason:** The robot arm remains stuck near its initial position near the center of the workspace and fails to move toward the pineapple on the stove.
+  Summary: robot arm stuck; no progress toward target
+- **recovery:** Command the arm to move to the right side toward the stove, position the gripper behind the pineapple, and push it off the stove.
+  Summary: reach toward stove; push pineapple off stove
+- **prevention:** Generate trajectory towards the stove and pineapple from the beginning of execution.
+  Summary: move directly to stove; push pineapple off
+- **justification:** The robot arm hovers near its base from the start and makes no meaningful motion toward the stove or pineapple throughout the entire 15-second duration, indicating a clear trajectory execution failure.
diff --git a/push_pineapple/ep048/labels.json b/push_pineapple/ep048/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..e8308c53bbaab4ed2af13922f223dcf7312f5757
--- /dev/null
+++ b/push_pineapple/ep048/labels.json
@@ -0,0 +1,120 @@
+{
+  "rollout_id": "t005__ep048__push_pineapple",
+  "scene_variant": "",
+  "task_id": 5,
+  "task_file": "",
+  "instruction": "Push the pineapple off the stove",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:52:39+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0,
+      "control_step_start": 0,
+      "control_step_end": 149,
+      "policy_step_start": 0,
+      "policy_step_end": 24,
+      "phrase": "hovering near base",
+      "description": "hovering near base"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0,
+      "t_end_sec": 6.0,
+      "control_step_start": 150,
+      "control_step_end": 299,
+      "policy_step_start": 25,
+      "policy_step_end": 49,
+      "phrase": "idle near center table",
+      "description": "idle near center table"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.0,
+      "t_end_sec": 9.0,
+      "control_step_start": 300,
+      "control_step_end": 449,
+      "policy_step_start": 50,
+      "policy_step_end": 74,
+      "phrase": "remaining stationary",
+      "description": "remaining stationary"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.0,
+      "t_end_sec": 12.0,
+      "control_step_start": 450,
+      "control_step_end": 599,
+      "policy_step_start": 75,
+      "policy_step_end": 99,
+      "phrase": "idle without moving",
+      "description": "idle without moving"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 12.0,
+      "t_end_sec": 14.98,
+      "control_step_start": 600,
+      "control_step_end": 748,
+      "policy_step_start": 100,
+      "policy_step_end": 124,
+      "phrase": "hovering near initial position",
+      "description": "hovering near initial position"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "stuck_or_no_progress",
+    "vlm_failure_onset_type": "operator_intervention",
+    "vlm_failure_onset_seconds": 1.12,
+    "vlm_failure_onset_timestamp": "00:01.1",
+    "vlm_failure_onset_frame": 56,
+    "vlm_failure_onset_step": 56,
+    "vlm_coarse_failure_onset_seconds": 2.5,
+    "vlm_failure_window_start_seconds": 2.0,
+    "vlm_failure_window_end_seconds": 3.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot arm remains stuck near its initial position near the center of the workspace and fails to move toward the pineapple on the stove.",
+    "vlm_failure_reason_summary": "robot arm stuck; no progress toward target",
+    "vlm_recovery_action": "Command the arm to move to the right side toward the stove, position the gripper behind the pineapple, and push it off the stove.",
+    "vlm_recovery_action_summary": "reach toward stove; push pineapple off stove",
+    "vlm_prevention_action": "Generate trajectory towards the stove and pineapple from the beginning of execution.",
+    "vlm_prevention_action_summary": "move directly to stove; push pineapple off",
+    "vlm_justification": "The robot arm hovers near its base from the start and makes no meaningful motion toward the stove or pineapple throughout the entire 15-second duration, indicating a clear trajectory execution failure.",
+    "vlm_temporal_justification": "At frame 56 (clip second 56), the robot arm initiates its first movement by making a small wrist rotation in place rather than moving toward the pineapple on the stove, marking the onset of the stuck/no-progress failure.",
+    "vlm_video_fps": 50.0,
+    "vlm_video_num_frames": 749,
+    "vlm_video_duration_seconds": 14.96,
+    "vlm_usage": {
+      "input_tokens": 20186,
+      "output_tokens": 352,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"stuck_or_no_progress\",\n \"onset_type\": \"operator_intervention\",\n \"failure_onset_seconds\": 2.5,\n \"failure_window_start_seconds\": 2.0,\n \"failure_window_end_seconds\": 3.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot arm remains stuck near its initial position near the center of the workspace and fails to move toward the pineapple on the stove.\",\n \"failure_reason_summary\": \"robot arm stuck; no progress toward target\",\n \"recovery_action\": \"Command the arm to move to the right side toward the stove, position the gripper behind the pineapple, and push it off the stove.\",\n \"recovery_action_summary\": \"reach toward stove; push pineapple off stove\",\n \"prevention_action\": \"Generate trajectory towards the stove and pineapple from the beginning of execution.\",\n \"prevention_action_summary\": \"move directly to stove; push pineapple off\",\n \"justification\": \"The robot arm hovers near its base from the start and makes no meaningful motion toward the stove or pineapple throughout the entire 15-second duration, indicating a clear trajectory execution failure.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 56,
+    "failure_sim_step": 56,
+    "failure_policy_step": 9,
+    "failure_chunk_index": 2,
+    "first_post_failure_policy_step": 10,
+    "failure_type": "stuck_or_no_progress",
+    "correction_action": "Command the arm to move to the right side toward the stove, position the gripper behind the pineapple, and push it off the stove."
+  }
+}
\ No newline at end of file
diff --git a/push_pineapple/ep048/labels.npz b/push_pineapple/ep048/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..952c6b5d83a9ea781f3887462d098bd47d9dc613
--- /dev/null
+++ b/push_pineapple/ep048/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:32f8d5e05cabd9b1dfc0b7baefa9396bfbca7471d05956defbde068ada841103
+size 3785
diff --git a/push_pineapple/ep049/example.md b/push_pineapple/ep049/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..319a4af34a2a2cb96433043dd6688c411eb64425
--- /dev/null
+++ b/push_pineapple/ep049/example.md
@@ -0,0 +1,35 @@
+# t005__ep049__push_pineapple
+
+- **suite / variant:** `` / ``
+- **instruction:** Push the pineapple off the stove
+- **success:** False
+- **control steps:** 670  (13.0s at 51 Hz)
+- **policy inferences:** 113  (replan_steps=6)
+- **video:** `push_pineapple/ep049/rollout.mp4`
+- **features:** `113` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–153 | 0–25 | hovering above center table |
+| 3.0–6.0 | 154–307 | 25–51 | idle near home position |
+| 6.0–9.0 | 308–461 | 51–76 | oscillating slightly in place |
+| 9.0–12.0 | 462–615 | 77–102 | remaining idle above table |
+| 12.0–13.0 | 616–669 | 102–111 | static near home position |
+
+## Failure
+
+- **yes**
+- **onset type:** operator_intervention
+- **onset:** 0.058365758754863814 s (00:00.1)
+- **frame / control_step:** 3 (policy_step=0, chunk_index=3)
+- **coarse onset:** 3.0 s (refined=True)
+- **mode:** stuck_or_no_progress
+- **reason:** The robot arm becomes stuck hovering near its home position above the table, making tiny ineffective motions and failing to approach the stove or pineapple.
+  Summary: robot stuck near home position; no progress toward pineapple
+- **recovery:** Reset arm position controller and re-plan a trajectory targeting the pineapple on the stove.
+  Summary: reset controller; target pineapple on stove
+- **prevention:** Improve policy motion execution to ensure end-effector navigates toward the stove target instead of remaining motionless near home position.
+  Summary: generate trajectory toward stove; avoid hovering in place
+- **justification:** By timestamp 00:03, the robot arm stops making forward progress and hovers in mid-air above the table for the remainder of the rollout, never reaching the stove.
diff --git a/push_pineapple/ep049/labels.json b/push_pineapple/ep049/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..5fc531c5c8e0316a43ba51217c9c39c077756860
--- /dev/null
+++ b/push_pineapple/ep049/labels.json
@@ -0,0 +1,120 @@
+{
+  "rollout_id": "t005__ep049__push_pineapple",
+  "scene_variant": "",
+  "task_id": 5,
+  "task_file": "",
+  "instruction": "Push the pineapple off the stove",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:53:13+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9961089494163424,
+      "control_step_start": 0,
+      "control_step_end": 153,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "hovering above center table",
+      "description": "hovering above center table"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9961089494163424,
+      "t_end_sec": 5.992217898832685,
+      "control_step_start": 154,
+      "control_step_end": 307,
+      "policy_step_start": 25,
+      "policy_step_end": 51,
+      "phrase": "idle near home position",
+      "description": "idle near home position"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.992217898832685,
+      "t_end_sec": 8.988326848249027,
+      "control_step_start": 308,
+      "control_step_end": 461,
+      "policy_step_start": 51,
+      "policy_step_end": 76,
+      "phrase": "oscillating slightly in place",
+      "description": "oscillating slightly in place"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.988326848249027,
+      "t_end_sec": 11.98443579766537,
+      "control_step_start": 462,
+      "control_step_end": 615,
+      "policy_step_start": 77,
+      "policy_step_end": 102,
+      "phrase": "remaining idle above table",
+      "description": "remaining idle above table"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 11.98443579766537,
+      "t_end_sec": 13.035019455252918,
+      "control_step_start": 616,
+      "control_step_end": 669,
+      "policy_step_start": 102,
+      "policy_step_end": 111,
+      "phrase": "static near home position",
+      "description": "static near home position"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "stuck_or_no_progress",
+    "vlm_failure_onset_type": "operator_intervention",
+    "vlm_failure_onset_seconds": 0.058365758754863814,
+    "vlm_failure_onset_timestamp": "00:00.1",
+    "vlm_failure_onset_frame": 3,
+    "vlm_failure_onset_step": 3,
+    "vlm_coarse_failure_onset_seconds": 3.0,
+    "vlm_failure_window_start_seconds": 3.0,
+    "vlm_failure_window_end_seconds": 13.02,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot arm becomes stuck hovering near its home position above the table, making tiny ineffective motions and failing to approach the stove or pineapple.",
+    "vlm_failure_reason_summary": "robot stuck near home position; no progress toward pineapple",
+    "vlm_recovery_action": "Reset arm position controller and re-plan a trajectory targeting the pineapple on the stove.",
+    "vlm_recovery_action_summary": "reset controller; target pineapple on stove",
+    "vlm_prevention_action": "Improve policy motion execution to ensure end-effector navigates toward the stove target instead of remaining motionless near home position.",
+    "vlm_prevention_action_summary": "generate trajectory toward stove; avoid hovering in place",
+    "vlm_justification": "By timestamp 00:03, the robot arm stops making forward progress and hovers in mid-air above the table for the remainder of the rollout, never reaching the stove.",
+    "vlm_temporal_justification": "At second 3 (corresponding to original frame 3), the robot fails to begin moving toward the target and instead remains stuck hovering near its initial home position without making progress on the task.",
+    "vlm_video_fps": 51.4,
+    "vlm_video_num_frames": 670,
+    "vlm_video_duration_seconds": 13.01556420233463,
+    "vlm_usage": {
+      "input_tokens": 22140,
+      "output_tokens": 345,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"stuck_or_no_progress\",\n \"onset_type\": \"operator_intervention\",\n \"failure_onset_seconds\": 3.0,\n \"failure_window_start_seconds\": 3.0,\n \"failure_window_end_seconds\": 13.02,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot arm becomes stuck hovering near its home position above the table, making tiny ineffective motions and failing to approach the stove or pineapple.\",\n \"failure_reason_summary\": \"robot stuck near home position; no progress toward pineapple\",\n \"recovery_action\": \"Reset arm position controller and re-plan a trajectory targeting the pineapple on the stove.\",\n \"recovery_action_summary\": \"reset controller; target pineapple on stove\",\n \"prevention_action\": \"Improve policy motion execution to ensure end-effector navigates toward the stove target instead of remaining motionless near home position.\",\n \"prevention_action_summary\": \"generate trajectory toward stove; avoid hovering in place\",\n \"justification\": \"By timestamp 00:03, the robot arm stops making forward progress and hovers in mid-air above the table for the remainder of the rollout, never reaching the stove.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 3,
+    "failure_sim_step": 3,
+    "failure_policy_step": 0,
+    "failure_chunk_index": 3,
+    "first_post_failure_policy_step": 1,
+    "failure_type": "stuck_or_no_progress",
+    "correction_action": "Reset arm position controller and re-plan a trajectory targeting the pineapple on the stove."
+  }
+}
\ No newline at end of file
diff --git a/push_pineapple/ep049/labels.npz b/push_pineapple/ep049/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..a72176e0231c765c1e07471dd027f601c90cefa9
--- /dev/null
+++ b/push_pineapple/ep049/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5db8a16b3e9bbd2837d4327a985ec9233d408dc8f60081f961ce09f2fe0eff4f
+size 3892
diff --git a/tomato/ep000/example.md b/tomato/ep000/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..b4534b9757258754ba37ddfccf6854722a5323f0
--- /dev/null
+++ b/tomato/ep000/example.md
@@ -0,0 +1,22 @@
+# t003__ep000__tomato
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the tomato in the bin
+- **success:** True
+- **control steps:** 503  (10.0s at 50 Hz)
+- **policy inferences:** 86  (replan_steps=6)
+- **video:** `tomato/ep000/rollout.mp4`
+- **features:** `86` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching for the tomato |
+| 3.0–6.0 | 152–303 | 25–50 | grasping and lifting tomato |
+| 6.0–9.0 | 304–455 | 50–75 | moving and dropping tomato in bin |
+| 9.0–10.0 | 456–502 | 76–83 | retracting arm |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/tomato/ep000/labels.json b/tomato/ep000/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..f95a8d059d7483ed04fba743d59c0f0602982114
--- /dev/null
+++ b/tomato/ep000/labels.json
@@ -0,0 +1,79 @@
+{
+  "rollout_id": "t003__ep000__tomato",
+  "scene_variant": "",
+  "task_id": 3,
+  "task_file": "",
+  "instruction": "Put the tomato in the bin",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:53:22+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00990099009901,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for the tomato",
+      "description": "reaching for the tomato"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00990099009901,
+      "t_end_sec": 6.01980198019802,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "grasping and lifting tomato",
+      "description": "grasping and lifting tomato"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01980198019802,
+      "t_end_sec": 9.029702970297029,
+      "control_step_start": 304,
+      "control_step_end": 455,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "moving and dropping tomato in bin",
+      "description": "moving and dropping tomato in bin"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.029702970297029,
+      "t_end_sec": 9.96039603960396,
+      "control_step_start": 456,
+      "control_step_end": 502,
+      "policy_step_start": 76,
+      "policy_step_end": 83,
+      "phrase": "retracting arm",
+      "description": "retracting arm"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/tomato/ep000/labels.npz b/tomato/ep000/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..c1bfc363b97af08395ec4103aab47e92b506a916
--- /dev/null
+++ b/tomato/ep000/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7d87cba4d82717fd0e20c15b756a8b4d6517b183d69ff5ae8e01893e9aad85ca
+size 3116
diff --git a/tomato/ep001/example.md b/tomato/ep001/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..d92cc9026c90ffe3388560d79a83bb5802aa23b0
--- /dev/null
+++ b/tomato/ep001/example.md
@@ -0,0 +1,21 @@
+# t003__ep001__tomato
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the tomato in the bin
+- **success:** True
+- **control steps:** 434  (8.6s at 50 Hz)
+- **policy inferences:** 76  (replan_steps=6)
+- **video:** `tomato/ep001/rollout.mp4`
+- **features:** `76` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | reaching for tomato |
+| 3.0–6.0 | 151–301 | 25–50 | grasping and lifting tomato |
+| 6.0–8.6 | 302–433 | 50–72 | placing tomato into bin |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/tomato/ep001/labels.json b/tomato/ep001/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..f3c0f371f0b0fcc8c80219faa2b755f57649938f
--- /dev/null
+++ b/tomato/ep001/labels.json
@@ -0,0 +1,68 @@
+{
+  "rollout_id": "t003__ep001__tomato",
+  "scene_variant": "",
+  "task_id": 3,
+  "task_file": "",
+  "instruction": "Put the tomato in the bin",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:53:29+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00796812749004,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for tomato",
+      "description": "reaching for tomato"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00796812749004,
+      "t_end_sec": 6.01593625498008,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "grasping and lifting tomato",
+      "description": "grasping and lifting tomato"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01593625498008,
+      "t_end_sec": 8.645418326693227,
+      "control_step_start": 302,
+      "control_step_end": 433,
+      "policy_step_start": 50,
+      "policy_step_end": 72,
+      "phrase": "placing tomato into bin",
+      "description": "placing tomato into bin"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/tomato/ep001/labels.npz b/tomato/ep001/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..fc1b921c937ecd44f01d6be04f9467b3ebf963c7
--- /dev/null
+++ b/tomato/ep001/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5d9d53d9dd6978dbafe0a7be0f2629f8b9bea602fe12b81a3d2c6f3d9957fa24
+size 3064
diff --git a/tomato/ep002/example.md b/tomato/ep002/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..b7f2e7c0a5e411ecafeaf29134db17ca546307d1
--- /dev/null
+++ b/tomato/ep002/example.md
@@ -0,0 +1,21 @@
+# t003__ep002__tomato
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the tomato in the bin
+- **success:** True
+- **control steps:** 386  (7.7s at 50 Hz)
+- **policy inferences:** 67  (replan_steps=5)
+- **video:** `tomato/ep002/rollout.mp4`
+- **features:** `67` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–149 | 0–29 | reaching for tomato |
+| 3.0–6.0 | 150–299 | 30–59 | placing tomato in bin |
+| 6.0–7.7 | 300–385 | 60–77 | retracting arm |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/tomato/ep002/labels.json b/tomato/ep002/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..fd52262eb43d0d6eeaeb160018553d65126ca312
--- /dev/null
+++ b/tomato/ep002/labels.json
@@ -0,0 +1,68 @@
+{
+  "rollout_id": "t003__ep002__tomato",
+  "scene_variant": "",
+  "task_id": 3,
+  "task_file": "",
+  "instruction": "Put the tomato in the bin",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:53:37+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9940119760479043,
+      "control_step_start": 0,
+      "control_step_end": 149,
+      "policy_step_start": 0,
+      "policy_step_end": 29,
+      "phrase": "reaching for tomato",
+      "description": "reaching for tomato"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9940119760479043,
+      "t_end_sec": 5.9880239520958085,
+      "control_step_start": 150,
+      "control_step_end": 299,
+      "policy_step_start": 30,
+      "policy_step_end": 59,
+      "phrase": "placing tomato in bin",
+      "description": "placing tomato in bin"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.9880239520958085,
+      "t_end_sec": 7.704590818363274,
+      "control_step_start": 300,
+      "control_step_end": 385,
+      "policy_step_start": 60,
+      "policy_step_end": 77,
+      "phrase": "retracting arm",
+      "description": "retracting arm"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/tomato/ep002/labels.npz b/tomato/ep002/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..9046d53243f920a8c4ea1df6bdfcc73683889c24
--- /dev/null
+++ b/tomato/ep002/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:93b88b3b952dd8cf757208e6f84fecd70250e15ea8b0539a5fa9c773e3192a10
+size 3058
diff --git a/tomato/ep003/example.md b/tomato/ep003/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..33e4dda580611d2c2feaed1b81cb7706298eaa78
--- /dev/null
+++ b/tomato/ep003/example.md
@@ -0,0 +1,22 @@
+# t003__ep003__tomato
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the tomato in the bin
+- **success:** True
+- **control steps:** 572  (11.4s at 50 Hz)
+- **policy inferences:** 99  (replan_steps=6)
+- **video:** `tomato/ep003/rollout.mp4`
+- **features:** `99` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | approaching the tomato |
+| 3.0–6.0 | 151–301 | 25–50 | grasping the tomato |
+| 6.0–9.0 | 302–452 | 50–75 | moving tomato to bin |
+| 9.0–11.4 | 453–571 | 75–95 | retracting robot arm |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/tomato/ep003/labels.json b/tomato/ep003/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..cc5b89262d16aeb777e507fe6ce43178be60e7d7
--- /dev/null
+++ b/tomato/ep003/labels.json
@@ -0,0 +1,79 @@
+{
+  "rollout_id": "t003__ep003__tomato",
+  "scene_variant": "",
+  "task_id": 3,
+  "task_file": "",
+  "instruction": "Put the tomato in the bin",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:53:45+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00796812749004,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "approaching the tomato",
+      "description": "approaching the tomato"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00796812749004,
+      "t_end_sec": 6.01593625498008,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "grasping the tomato",
+      "description": "grasping the tomato"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01593625498008,
+      "t_end_sec": 9.02390438247012,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "moving tomato to bin",
+      "description": "moving tomato to bin"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.02390438247012,
+      "t_end_sec": 11.39442231075697,
+      "control_step_start": 453,
+      "control_step_end": 571,
+      "policy_step_start": 75,
+      "policy_step_end": 95,
+      "phrase": "retracting robot arm",
+      "description": "retracting robot arm"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/tomato/ep003/labels.npz b/tomato/ep003/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..662a316f1491183678678fa7fe009c5dade9b315
--- /dev/null
+++ b/tomato/ep003/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:aea2a8b360308ea4aa98d62963b7a8a13c79a7ca10da5e8548dd7371756d20e5
+size 3101
diff --git a/tomato/ep004/example.md b/tomato/ep004/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..6d4bf244f7fbd6286f85d77064e88935294ef3c1
--- /dev/null
+++ b/tomato/ep004/example.md
@@ -0,0 +1,22 @@
+# t003__ep004__tomato
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the tomato in the bin
+- **success:** True
+- **control steps:** 513  (10.2s at 50 Hz)
+- **policy inferences:** 88  (replan_steps=6)
+- **video:** `tomato/ep004/rollout.mp4`
+- **features:** `88` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | reaching for the tomato |
+| 3.0–6.0 | 151–301 | 25–50 | grasping and lifting the tomato |
+| 6.0–9.0 | 302–452 | 50–75 | placing tomato into the bin |
+| 9.0–10.2 | 453–512 | 75–85 | retracting arm to initial position |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/tomato/ep004/labels.json b/tomato/ep004/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..a2b0d0b44f0ce8e91115a9e4b91d46ae5e6ec966
--- /dev/null
+++ b/tomato/ep004/labels.json
@@ -0,0 +1,79 @@
+{
+  "rollout_id": "t003__ep004__tomato",
+  "scene_variant": "",
+  "task_id": 3,
+  "task_file": "",
+  "instruction": "Put the tomato in the bin",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:53:53+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.996031746031746,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for the tomato",
+      "description": "reaching for the tomato"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.996031746031746,
+      "t_end_sec": 5.992063492063492,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "grasping and lifting the tomato",
+      "description": "grasping and lifting the tomato"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.992063492063492,
+      "t_end_sec": 8.988095238095239,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "placing tomato into the bin",
+      "description": "placing tomato into the bin"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.988095238095239,
+      "t_end_sec": 10.178571428571429,
+      "control_step_start": 453,
+      "control_step_end": 512,
+      "policy_step_start": 75,
+      "policy_step_end": 85,
+      "phrase": "retracting arm to initial position",
+      "description": "retracting arm to initial position"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/tomato/ep004/labels.npz b/tomato/ep004/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..cb8ddc1a680773aee9e703370ab5d95d0589a7a6
--- /dev/null
+++ b/tomato/ep004/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6c43b4b1bb4c92aab5269de29f811500a87ee72b1200fc2b64b11de6806428cf
+size 3127
diff --git a/tomato/ep005/example.md b/tomato/ep005/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..099a28d4b1220f2f1a53a187586de827329e7784
--- /dev/null
+++ b/tomato/ep005/example.md
@@ -0,0 +1,33 @@
+# t003__ep005__tomato
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the tomato in the bin
+- **success:** False
+- **control steps:** 421  (8.2s at 51 Hz)
+- **policy inferences:** 72  (replan_steps=6)
+- **video:** `tomato/ep005/rollout.mp4`
+- **features:** `72` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–152 | 0–25 | approaching tomato |
+| 3.0–6.0 | 153–305 | 25–50 | attempting to grasp tomato |
+| 6.0–8.2 | 306–420 | 51–70 | retracting arm empty |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 5.499021526418787 s (00:05.5)
+- **frame / control_step:** 281 (policy_step=46, chunk_index=5)
+- **coarse onset:** 5.0 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the tomato, but failed to secure a firm grip, causing the tomato to slip out as the arm lifted empty.
+  Summary: failed grasp on tomato; tomato slipped out
+- **recovery:** Re-align the end-effector over the tomato, descend to table level, establish a secure grasp, and place it into the gray bin.
+  Summary: re-align over tomato; grasp tomato securely; place in bin
+- **prevention:** Ensure the gripper is properly centered and aligned around the tomato before closing fingers and lifting.
+  Summary: center gripper over tomato; close fingers firmly before lifting
+- **justification:** At around 5.0 seconds, the robot attempts to close its gripper around the tomato, but fails to hold it securely. By 6.0 seconds, the robot arm lifts away leaving the tomato sitting on the table surface.
diff --git a/tomato/ep005/labels.json b/tomato/ep005/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..0d596d848d322043147c6b43078e35c0689e123e
--- /dev/null
+++ b/tomato/ep005/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t003__ep005__tomato",
+  "scene_variant": "",
+  "task_id": 3,
+  "task_file": "",
+  "instruction": "Put the tomato in the bin",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:54:25+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.99412915851272,
+      "control_step_start": 0,
+      "control_step_end": 152,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "approaching tomato",
+      "description": "approaching tomato"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.99412915851272,
+      "t_end_sec": 5.98825831702544,
+      "control_step_start": 153,
+      "control_step_end": 305,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "attempting to grasp tomato",
+      "description": "attempting to grasp tomato"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.98825831702544,
+      "t_end_sec": 8.238747553816047,
+      "control_step_start": 306,
+      "control_step_end": 420,
+      "policy_step_start": 51,
+      "policy_step_end": 70,
+      "phrase": "retracting arm empty",
+      "description": "retracting arm empty"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 5.499021526418787,
+    "vlm_failure_onset_timestamp": "00:05.5",
+    "vlm_failure_onset_frame": 281,
+    "vlm_failure_onset_step": 281,
+    "vlm_coarse_failure_onset_seconds": 5.0,
+    "vlm_failure_window_start_seconds": 4.5,
+    "vlm_failure_window_end_seconds": 6.0,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the tomato, but failed to secure a firm grip, causing the tomato to slip out as the arm lifted empty.",
+    "vlm_failure_reason_summary": "failed grasp on tomato; tomato slipped out",
+    "vlm_recovery_action": "Re-align the end-effector over the tomato, descend to table level, establish a secure grasp, and place it into the gray bin.",
+    "vlm_recovery_action_summary": "re-align over tomato; grasp tomato securely; place in bin",
+    "vlm_prevention_action": "Ensure the gripper is properly centered and aligned around the tomato before closing fingers and lifting.",
+    "vlm_prevention_action_summary": "center gripper over tomato; close fingers firmly before lifting",
+    "vlm_justification": "At around 5.0 seconds, the robot attempts to close its gripper around the tomato, but fails to hold it securely. By 6.0 seconds, the robot arm lifts away leaving the tomato sitting on the table surface.",
+    "vlm_temporal_justification": "The robot lowers its gripper over the tomato and closes its fingers, but fails to encompass the tomato. At second 178 (02:58 in the clip), the gripper fingers close completely empty while the tomato remains sitting on the table, marking the exact moment the grasp attempt fails.",
+    "vlm_video_fps": 51.1,
+    "vlm_video_num_frames": 421,
+    "vlm_video_duration_seconds": 8.21917808219178,
+    "vlm_usage": {
+      "input_tokens": 21701,
+      "output_tokens": 385,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 5.0,\n \"failure_window_start_seconds\": 4.5,\n \"failure_window_end_seconds\": 6.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the tomato, but failed to secure a firm grip, causing the tomato to slip out as the arm lifted empty.\",\n \"failure_reason_summary\": \"failed grasp on tomato; tomato slipped out\",\n \"recovery_action\": \"Re-align the end-effector over the tomato, descend to table level, establish a secure grasp, and place it into the gray bin.\",\n \"recovery_action_summary\": \"re-align over tomato; grasp tomato securely; place in bin\",\n \"prevention_action\": \"Ensure the gripper is properly centered and aligned around the tomato before closing fingers and lifting.\",\n \"prevention_action_summary\": \"center gripper over tomato; close fingers firmly before lifting\",\n \"justification\": \"At around 5.0 seconds, the robot attempts to close its gripper around the tomato, but fails to hold it securely. By 6.0 seconds, the robot arm lifts away leaving the tomato sitting on the table surface.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 281,
+    "failure_sim_step": 281,
+    "failure_policy_step": 46,
+    "failure_chunk_index": 5,
+    "first_post_failure_policy_step": 47,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-align the end-effector over the tomato, descend to table level, establish a secure grasp, and place it into the gray bin."
+  }
+}
\ No newline at end of file
diff --git a/tomato/ep005/labels.npz b/tomato/ep005/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..b00a4d02edd8f17e25a353fe7729062bc3b86549
--- /dev/null
+++ b/tomato/ep005/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a5bbe39d0ebf83ff83883718e853fa451bd459b8d3906dd0c7addf070f6eb53f
+size 3730
diff --git a/tomato/ep006/example.md b/tomato/ep006/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..61dc31a9e5ea57dbb3bb0c84bec60c6c921b3f42
--- /dev/null
+++ b/tomato/ep006/example.md
@@ -0,0 +1,22 @@
+# t003__ep006__tomato
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the tomato in the bin
+- **success:** True
+- **control steps:** 531  (10.7s at 50 Hz)
+- **policy inferences:** 92  (replan_steps=6)
+- **video:** `tomato/ep006/rollout.mp4`
+- **features:** `92` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–148 | 0–24 | reaching for tomato |
+| 3.0–6.0 | 149–297 | 24–49 | grasping and lifting tomato |
+| 6.0–9.0 | 298–446 | 49–74 | placing tomato into bin |
+| 9.0–10.7 | 447–530 | 74–88 | retracting arm |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/tomato/ep006/labels.json b/tomato/ep006/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..5b2e2867b3000ebc747cc6f03a73bb076ee2238f
--- /dev/null
+++ b/tomato/ep006/labels.json
@@ -0,0 +1,79 @@
+{
+  "rollout_id": "t003__ep006__tomato",
+  "scene_variant": "",
+  "task_id": 3,
+  "task_file": "",
+  "instruction": "Put the tomato in the bin",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:54:32+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.991967871485944,
+      "control_step_start": 0,
+      "control_step_end": 148,
+      "policy_step_start": 0,
+      "policy_step_end": 24,
+      "phrase": "reaching for tomato",
+      "description": "reaching for tomato"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.991967871485944,
+      "t_end_sec": 5.983935742971888,
+      "control_step_start": 149,
+      "control_step_end": 297,
+      "policy_step_start": 24,
+      "policy_step_end": 49,
+      "phrase": "grasping and lifting tomato",
+      "description": "grasping and lifting tomato"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.983935742971888,
+      "t_end_sec": 8.975903614457831,
+      "control_step_start": 298,
+      "control_step_end": 446,
+      "policy_step_start": 49,
+      "policy_step_end": 74,
+      "phrase": "placing tomato into bin",
+      "description": "placing tomato into bin"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.975903614457831,
+      "t_end_sec": 10.66265060240964,
+      "control_step_start": 447,
+      "control_step_end": 530,
+      "policy_step_start": 74,
+      "policy_step_end": 88,
+      "phrase": "retracting arm",
+      "description": "retracting arm"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/tomato/ep006/labels.npz b/tomato/ep006/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..81beb4f7663eb27e4e4d0e43c8c48d5312b0c7cb
--- /dev/null
+++ b/tomato/ep006/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ea4d3e182aabb45315cbbcfdfe22bab89b5dad466ac7f9f0807f7858eb0a27c8
+size 3104
diff --git a/tomato/ep007/example.md b/tomato/ep007/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..ece3f7765f15de57f7f0c94f5fadbce1d9e6ae26
--- /dev/null
+++ b/tomato/ep007/example.md
@@ -0,0 +1,33 @@
+# t003__ep007__tomato
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the tomato in the bin
+- **success:** False
+- **control steps:** 402  (8.0s at 50 Hz)
+- **policy inferences:** 69  (replan_steps=6)
+- **video:** `tomato/ep007/rollout.mp4`
+- **features:** `69` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | reaching toward tomato |
+| 3.0–6.0 | 151–301 | 25–50 | closing gripper on tomato |
+| 6.0–8.0 | 302–401 | 50–66 | holding gripper over ungrasped tomato |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 7.211155378486056 s (00:07.2)
+- **frame / control_step:** 362 (policy_step=60, chunk_index=2)
+- **coarse onset:** 4.5 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot lowered its end-effector over the tomato and closed its gripper, but failed to establish a secure grasp, leaving the tomato behind on the table.
+  Summary: failed grasp on tomato; object ungrasped
+- **recovery:** Re-open the gripper, center the end-effector over the tomato, descend to the correct gripping height, and close the gripper securely before lifting.
+  Summary: re-align gripper; regrasp tomato; place in bin
+- **prevention:** Center the end-effector accurately over the tomato and adjust grasping depth prior to closing the fingers so that the object is held firmly.
+  Summary: align centered with tomato; close gripper firmly
+- **justification:** Around 4.5 seconds into the rollout, the gripper closes around the tomato but fails to secure it. When the arm attempts to complete the maneuver, the tomato remains untouched on the tabletop.
diff --git a/tomato/ep007/labels.json b/tomato/ep007/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..96cc6a0107bf67d3449028346eee0d248c440781
--- /dev/null
+++ b/tomato/ep007/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t003__ep007__tomato",
+  "scene_variant": "",
+  "task_id": 3,
+  "task_file": "",
+  "instruction": "Put the tomato in the bin",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:55:05+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00796812749004,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching toward tomato",
+      "description": "reaching toward tomato"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00796812749004,
+      "t_end_sec": 6.01593625498008,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "closing gripper on tomato",
+      "description": "closing gripper on tomato"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01593625498008,
+      "t_end_sec": 8.00796812749004,
+      "control_step_start": 302,
+      "control_step_end": 401,
+      "policy_step_start": 50,
+      "policy_step_end": 66,
+      "phrase": "holding gripper over ungrasped tomato",
+      "description": "holding gripper over ungrasped tomato"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 7.211155378486056,
+    "vlm_failure_onset_timestamp": "00:07.2",
+    "vlm_failure_onset_frame": 362,
+    "vlm_failure_onset_step": 362,
+    "vlm_coarse_failure_onset_seconds": 4.5,
+    "vlm_failure_window_start_seconds": 3.5,
+    "vlm_failure_window_end_seconds": 5.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot lowered its end-effector over the tomato and closed its gripper, but failed to establish a secure grasp, leaving the tomato behind on the table.",
+    "vlm_failure_reason_summary": "failed grasp on tomato; object ungrasped",
+    "vlm_recovery_action": "Re-open the gripper, center the end-effector over the tomato, descend to the correct gripping height, and close the gripper securely before lifting.",
+    "vlm_recovery_action_summary": "re-align gripper; regrasp tomato; place in bin",
+    "vlm_prevention_action": "Center the end-effector accurately over the tomato and adjust grasping depth prior to closing the fingers so that the object is held firmly.",
+    "vlm_prevention_action_summary": "align centered with tomato; close gripper firmly",
+    "vlm_justification": "Around 4.5 seconds into the rollout, the gripper closes around the tomato but fails to secure it. When the arm attempts to complete the maneuver, the tomato remains untouched on the tabletop.",
+    "vlm_temporal_justification": "At clip second 04:47 (corresponding to frame 287), the robot gripper closes completely without grasping the tomato, leaving it on the table as the arm begins to lift away.",
+    "vlm_video_fps": 50.2,
+    "vlm_video_num_frames": 402,
+    "vlm_video_duration_seconds": 7.988047808764939,
+    "vlm_usage": {
+      "input_tokens": 21449,
+      "output_tokens": 369,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 4.5,\n \"failure_window_start_seconds\": 3.5,\n \"failure_window_end_seconds\": 5.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot lowered its end-effector over the tomato and closed its gripper, but failed to establish a secure grasp, leaving the tomato behind on the table.\",\n \"failure_reason_summary\": \"failed grasp on tomato; object ungrasped\",\n \"recovery_action\": \"Re-open the gripper, center the end-effector over the tomato, descend to the correct gripping height, and close the gripper securely before lifting.\",\n \"recovery_action_summary\": \"re-align gripper; regrasp tomato; place in bin\",\n \"prevention_action\": \"Center the end-effector accurately over the tomato and adjust grasping depth prior to closing the fingers so that the object is held firmly.\",\n \"prevention_action_summary\": \"align centered with tomato; close gripper firmly\",\n \"justification\": \"Around 4.5 seconds into the rollout, the gripper closes around the tomato but fails to secure it. When the arm attempts to complete the maneuver, the tomato remains untouched on the tabletop.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 362,
+    "failure_sim_step": 362,
+    "failure_policy_step": 60,
+    "failure_chunk_index": 2,
+    "first_post_failure_policy_step": 61,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-open the gripper, center the end-effector over the tomato, descend to the correct gripping height, and close the gripper securely before lifting."
+  }
+}
\ No newline at end of file
diff --git a/tomato/ep007/labels.npz b/tomato/ep007/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..810b3cffb9b2ec2ca90a838563437e50e80a78a1
--- /dev/null
+++ b/tomato/ep007/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a33a860b58378254717d9613d1a9b9908009208425f4e802f64557e04076ffff
+size 3788
diff --git a/tomato/ep008/example.md b/tomato/ep008/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..b2271eb3bea2c139d36b3eae198313189ba3aad6
--- /dev/null
+++ b/tomato/ep008/example.md
@@ -0,0 +1,23 @@
+# t003__ep008__tomato
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the tomato in the bin
+- **success:** True
+- **control steps:** 659  (13.0s at 50 Hz)
+- **policy inferences:** 113  (replan_steps=6)
+- **video:** `tomato/ep008/rollout.mp4`
+- **features:** `113` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching for tomato |
+| 3.0–6.0 | 152–303 | 25–50 | grasping tomato |
+| 6.0–9.0 | 304–455 | 50–75 | lifting tomato |
+| 9.0–12.0 | 456–607 | 76–101 | placing tomato in bin |
+| 12.0–13.0 | 608–658 | 101–109 | idle / task complete |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/tomato/ep008/labels.json b/tomato/ep008/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..f25bd9d79f469afdafd9d3ef038f7300c8cb36cd
--- /dev/null
+++ b/tomato/ep008/labels.json
@@ -0,0 +1,90 @@
+{
+  "rollout_id": "t003__ep008__tomato",
+  "scene_variant": "",
+  "task_id": 3,
+  "task_file": "",
+  "instruction": "Put the tomato in the bin",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:55:13+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00990099009901,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for tomato",
+      "description": "reaching for tomato"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00990099009901,
+      "t_end_sec": 6.01980198019802,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "grasping tomato",
+      "description": "grasping tomato"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01980198019802,
+      "t_end_sec": 9.029702970297029,
+      "control_step_start": 304,
+      "control_step_end": 455,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "lifting tomato",
+      "description": "lifting tomato"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.029702970297029,
+      "t_end_sec": 12.03960396039604,
+      "control_step_start": 456,
+      "control_step_end": 607,
+      "policy_step_start": 76,
+      "policy_step_end": 101,
+      "phrase": "placing tomato in bin",
+      "description": "placing tomato in bin"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 12.03960396039604,
+      "t_end_sec": 13.049504950495049,
+      "control_step_start": 608,
+      "control_step_end": 658,
+      "policy_step_start": 101,
+      "policy_step_end": 109,
+      "phrase": "idle / task complete",
+      "description": "idle / task complete"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/tomato/ep008/labels.npz b/tomato/ep008/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..ec7d39e8c4b4540bae340f8b3e8eb42a8897d9a2
--- /dev/null
+++ b/tomato/ep008/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4ef96794673155660668c677f1c4c9c2e81d6caadf17adf44ac1eb7bc98687cf
+size 3128
diff --git a/tomato/ep009/example.md b/tomato/ep009/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..2f777776d1ed34e4f31959e474e7c08fed3d4ca7
--- /dev/null
+++ b/tomato/ep009/example.md
@@ -0,0 +1,22 @@
+# t003__ep009__tomato
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the tomato in the bin
+- **success:** True
+- **control steps:** 516  (10.2s at 50 Hz)
+- **policy inferences:** 89  (replan_steps=6)
+- **video:** `tomato/ep009/rollout.mp4`
+- **features:** `89` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | reaching for tomato |
+| 3.0–6.0 | 151–301 | 25–50 | grasping and lifting tomato |
+| 6.0–9.0 | 302–452 | 50–75 | moving tomato into bin |
+| 9.0–10.2 | 453–515 | 75–85 | retracting robot arm |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/tomato/ep009/labels.json b/tomato/ep009/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..8fcbada270e427b2ce626c7ecfc75da1c140c5c6
--- /dev/null
+++ b/tomato/ep009/labels.json
@@ -0,0 +1,79 @@
+{
+  "rollout_id": "t003__ep009__tomato",
+  "scene_variant": "",
+  "task_id": 3,
+  "task_file": "",
+  "instruction": "Put the tomato in the bin",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:55:21+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.996031746031746,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for tomato",
+      "description": "reaching for tomato"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.996031746031746,
+      "t_end_sec": 5.992063492063492,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "grasping and lifting tomato",
+      "description": "grasping and lifting tomato"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.992063492063492,
+      "t_end_sec": 8.988095238095239,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "moving tomato into bin",
+      "description": "moving tomato into bin"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.988095238095239,
+      "t_end_sec": 10.238095238095239,
+      "control_step_start": 453,
+      "control_step_end": 515,
+      "policy_step_start": 75,
+      "policy_step_end": 85,
+      "phrase": "retracting robot arm",
+      "description": "retracting robot arm"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/tomato/ep009/labels.npz b/tomato/ep009/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..26784f4d9e8b69c8c9a27d9ee10724c57c6bf492
--- /dev/null
+++ b/tomato/ep009/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a3a271c467dca29ffa72e2f49fffc450741a3604a903a98541147b5f5c5be444
+size 3112
diff --git a/tomato/ep010/example.md b/tomato/ep010/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..0a53bf5f9695816b62450f69a59a406205d54532
--- /dev/null
+++ b/tomato/ep010/example.md
@@ -0,0 +1,35 @@
+# t003__ep010__tomato
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the tomato in the bin
+- **success:** False
+- **control steps:** 713  (14.2s at 50 Hz)
+- **policy inferences:** 124  (replan_steps=6)
+- **video:** `tomato/ep010/rollout.mp4`
+- **features:** `124` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | approaching and grasping tomato |
+| 3.0–6.0 | 151–301 | 25–50 | dragging tomato across table |
+| 6.0–9.0 | 302–452 | 50–75 | continuing to drag tomato |
+| 9.0–12.0 | 453–603 | 75–100 | pushing tomato toward bin |
+| 12.0–14.2 | 604–712 | 100–118 | holding tomato against table |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 1.0337972166998013 s (00:01.0)
+- **frame / control_step:** 52 (policy_step=8, chunk_index=4)
+- **coarse onset:** 4.0 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot closed its gripper on the tomato but failed to elevate it off the table surface, dragging the object across the table instead of securely picking it up and placing it into the bin.
+  Summary: failed to lift tomato; dragging object across table
+- **recovery:** Re-align the gripper, close securely on the tomato, lift it vertically off the tabletop, and move it into the bin before releasing.
+  Summary: re-grasp tomato; lift vertically; place in bin
+- **prevention:** Raise the arm vertically immediately after closing the gripper so the tomato is fully lifted above the table before attempting horizontal motion towards the bin.
+  Summary: lift tomato vertically; move over bin; place inside
+- **justification:** At around 4.0 seconds, after closing its gripper around the tomato, the robot moves laterally across the table without lifting the object. It drags the tomato along the surface for the remainder of the video, failing to pick it up or deposit it into the bin.
diff --git a/tomato/ep010/labels.json b/tomato/ep010/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..1d0fec0a33607a3d8d3c76454f62439c25a57b42
--- /dev/null
+++ b/tomato/ep010/labels.json
@@ -0,0 +1,120 @@
+{
+  "rollout_id": "t003__ep010__tomato",
+  "scene_variant": "",
+  "task_id": 3,
+  "task_file": "",
+  "instruction": "Put the tomato in the bin",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:55:55+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.001988071570577,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "approaching and grasping tomato",
+      "description": "approaching and grasping tomato"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.001988071570577,
+      "t_end_sec": 6.003976143141154,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "dragging tomato across table",
+      "description": "dragging tomato across table"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.003976143141154,
+      "t_end_sec": 9.00596421471173,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "continuing to drag tomato",
+      "description": "continuing to drag tomato"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.00596421471173,
+      "t_end_sec": 12.007952286282308,
+      "control_step_start": 453,
+      "control_step_end": 603,
+      "policy_step_start": 75,
+      "policy_step_end": 100,
+      "phrase": "pushing tomato toward bin",
+      "description": "pushing tomato toward bin"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 12.007952286282308,
+      "t_end_sec": 14.174950298210737,
+      "control_step_start": 604,
+      "control_step_end": 712,
+      "policy_step_start": 100,
+      "policy_step_end": 118,
+      "phrase": "holding tomato against table",
+      "description": "holding tomato against table"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 1.0337972166998013,
+    "vlm_failure_onset_timestamp": "00:01.0",
+    "vlm_failure_onset_frame": 52,
+    "vlm_failure_onset_step": 52,
+    "vlm_coarse_failure_onset_seconds": 4.0,
+    "vlm_failure_window_start_seconds": 3.5,
+    "vlm_failure_window_end_seconds": 4.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot closed its gripper on the tomato but failed to elevate it off the table surface, dragging the object across the table instead of securely picking it up and placing it into the bin.",
+    "vlm_failure_reason_summary": "failed to lift tomato; dragging object across table",
+    "vlm_recovery_action": "Re-align the gripper, close securely on the tomato, lift it vertically off the tabletop, and move it into the bin before releasing.",
+    "vlm_recovery_action_summary": "re-grasp tomato; lift vertically; place in bin",
+    "vlm_prevention_action": "Raise the arm vertically immediately after closing the gripper so the tomato is fully lifted above the table before attempting horizontal motion towards the bin.",
+    "vlm_prevention_action_summary": "lift tomato vertically; move over bin; place inside",
+    "vlm_justification": "At around 4.0 seconds, after closing its gripper around the tomato, the robot moves laterally across the table without lifting the object. It drags the tomato along the surface for the remainder of the video, failing to pick it up or deposit it into the bin.",
+    "vlm_temporal_justification": "The robot gripper finishes closing on the tomato at around second 02:01. At second 02:04, instead of elevating the object off the table surface, the robotic arm begins moving horizontally to the left, causing the tomato to drag across the table.",
+    "vlm_video_fps": 50.3,
+    "vlm_video_num_frames": 713,
+    "vlm_video_duration_seconds": 14.155069582504971,
+    "vlm_usage": {
+      "input_tokens": 21835,
+      "output_tokens": 405,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 4.0,\n \"failure_window_start_seconds\": 3.5,\n \"failure_window_end_seconds\": 4.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot closed its gripper on the tomato but failed to elevate it off the table surface, dragging the object across the table instead of securely picking it up and placing it into the bin.\",\n \"failure_reason_summary\": \"failed to lift tomato; dragging object across table\",\n \"recovery_action\": \"Re-align the gripper, close securely on the tomato, lift it vertically off the tabletop, and move it into the bin before releasing.\",\n \"recovery_action_summary\": \"re-grasp tomato; lift vertically; place in bin\",\n \"prevention_action\": \"Raise the arm vertically immediately after closing the gripper so the tomato is fully lifted above the table before attempting horizontal motion towards the bin.\",\n \"prevention_action_summary\": \"lift tomato vertically; move over bin; place inside\",\n \"justification\": \"At around 4.0 seconds, after closing its gripper around the tomato, the robot moves laterally across the table without lifting the object. It drags the tomato along the surface for the remainder of the video, failing to pick it up or deposit it into the bin.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 52,
+    "failure_sim_step": 52,
+    "failure_policy_step": 8,
+    "failure_chunk_index": 4,
+    "first_post_failure_policy_step": 9,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-align the gripper, close securely on the tomato, lift it vertically off the tabletop, and move it into the bin before releasing."
+  }
+}
\ No newline at end of file
diff --git a/tomato/ep010/labels.npz b/tomato/ep010/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..0677468a4dce40823cdb99f28df9a9ce8a55755c
--- /dev/null
+++ b/tomato/ep010/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2d31a49c74cf13ed5c2dcb11137d018053367dcab66c13d00a970e23d99aeb44
+size 3945
diff --git a/tomato/ep011/example.md b/tomato/ep011/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..80ffddaeed1d42136327d246ad39543bd9834084
--- /dev/null
+++ b/tomato/ep011/example.md
@@ -0,0 +1,22 @@
+# t003__ep011__tomato
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the tomato in the bin
+- **success:** True
+- **control steps:** 477  (9.5s at 50 Hz)
+- **policy inferences:** 83  (replan_steps=6)
+- **video:** `tomato/ep011/rollout.mp4`
+- **features:** `83` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | reaching for the tomato |
+| 3.0–6.0 | 151–301 | 25–50 | grasping and lifting the tomato |
+| 6.0–9.0 | 302–452 | 50–75 | dropping tomato into the bin |
+| 9.0–9.5 | 453–476 | 75–79 | retracting arm |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/tomato/ep011/labels.json b/tomato/ep011/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..10cc8d5f372dac5d6f132977cb0c774f5907cb25
--- /dev/null
+++ b/tomato/ep011/labels.json
@@ -0,0 +1,79 @@
+{
+  "rollout_id": "t003__ep011__tomato",
+  "scene_variant": "",
+  "task_id": 3,
+  "task_file": "",
+  "instruction": "Put the tomato in the bin",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:56:03+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.996031746031746,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for the tomato",
+      "description": "reaching for the tomato"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.996031746031746,
+      "t_end_sec": 5.992063492063492,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "grasping and lifting the tomato",
+      "description": "grasping and lifting the tomato"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.992063492063492,
+      "t_end_sec": 8.988095238095239,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "dropping tomato into the bin",
+      "description": "dropping tomato into the bin"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.988095238095239,
+      "t_end_sec": 9.464285714285715,
+      "control_step_start": 453,
+      "control_step_end": 476,
+      "policy_step_start": 75,
+      "policy_step_end": 79,
+      "phrase": "retracting arm",
+      "description": "retracting arm"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/tomato/ep011/labels.npz b/tomato/ep011/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..8cf4d1befe306af56ff83a0113b4bd1aafb1ca2e
--- /dev/null
+++ b/tomato/ep011/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e1df27223688fc2cf825c3743a75bded32b51dd1b2176b9128226698219c25a2
+size 3112
diff --git a/tomato/ep012/example.md b/tomato/ep012/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..1779fb88b16a1747d388f572e5337bf9c3302f27
--- /dev/null
+++ b/tomato/ep012/example.md
@@ -0,0 +1,22 @@
+# t003__ep012__tomato
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the tomato in the bin
+- **success:** True
+- **control steps:** 500  (9.8s at 51 Hz)
+- **policy inferences:** 86  (replan_steps=6)
+- **video:** `tomato/ep012/rollout.mp4`
+- **features:** `86` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching for tomato |
+| 3.0–6.0 | 152–303 | 25–50 | grasping and lifting tomato |
+| 6.0–9.0 | 304–455 | 50–75 | placing tomato into bin |
+| 9.0–9.8 | 456–499 | 76–83 | retracting arm |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/tomato/ep012/labels.json b/tomato/ep012/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..19cb5670c41caf5f26208471c9b065fffa2833d7
--- /dev/null
+++ b/tomato/ep012/labels.json
@@ -0,0 +1,79 @@
+{
+  "rollout_id": "t003__ep012__tomato",
+  "scene_variant": "",
+  "task_id": 3,
+  "task_file": "",
+  "instruction": "Put the tomato in the bin",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:56:11+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9921259842519685,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for tomato",
+      "description": "reaching for tomato"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9921259842519685,
+      "t_end_sec": 5.984251968503937,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "grasping and lifting tomato",
+      "description": "grasping and lifting tomato"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.984251968503937,
+      "t_end_sec": 8.976377952755906,
+      "control_step_start": 304,
+      "control_step_end": 455,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "placing tomato into bin",
+      "description": "placing tomato into bin"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.976377952755906,
+      "t_end_sec": 9.84251968503937,
+      "control_step_start": 456,
+      "control_step_end": 499,
+      "policy_step_start": 76,
+      "policy_step_end": 83,
+      "phrase": "retracting arm",
+      "description": "retracting arm"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/tomato/ep012/labels.npz b/tomato/ep012/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..f3308cb906bc71a1da960bdaee6409489d290ef4
--- /dev/null
+++ b/tomato/ep012/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e6fb981fc60f69a9305f5cc82330a7a68cb3d6a08af4489d88672779de6f04be
+size 3103
diff --git a/tomato/ep013/example.md b/tomato/ep013/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..7fc7cc6f0a7bc8877efb8a0684cd788ad4688e98
--- /dev/null
+++ b/tomato/ep013/example.md
@@ -0,0 +1,33 @@
+# t003__ep013__tomato
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the tomato in the bin
+- **success:** False
+- **control steps:** 409  (8.1s at 51 Hz)
+- **policy inferences:** 71  (replan_steps=6)
+- **video:** `tomato/ep013/rollout.mp4`
+- **features:** `71` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | lowering arm toward tomato |
+| 3.0–6.0 | 152–303 | 25–50 | closing gripper around tomato |
+| 6.0–8.1 | 304–408 | 50–68 | remaining stationary over tomato |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 5.374015748031496 s (00:05.4)
+- **frame / control_step:** 273 (policy_step=45, chunk_index=3)
+- **coarse onset:** 4.0 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot reaches down to the tomato and closes its gripper around it, but fails to establish a secure grasp to lift the tomato.
+  Summary: failed grasp on tomato; no lift achieved
+- **recovery:** Re-open the end-effector, re-center the gripper over the tomato, close firmly around it, and lift it toward the bin.
+  Summary: re-align gripper; grasp tomato securely; lift into bin
+- **prevention:** Ensure precise alignment and proper gripper positioning around the center of the tomato before closing fingers to secure the object.
+  Summary: center gripper on tomato; close fingers firmly; lift object
+- **justification:** At around 4.0 seconds into the video, the robot closes its gripper over the tomato on the table but fails to securely grasp or lift it, leaving the tomato on the table for the remainder of the rollout.
diff --git a/tomato/ep013/labels.json b/tomato/ep013/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..3bd40479e41a3f0e6fb04ccead0fb9292d4618b8
--- /dev/null
+++ b/tomato/ep013/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t003__ep013__tomato",
+  "scene_variant": "",
+  "task_id": 3,
+  "task_file": "",
+  "instruction": "Put the tomato in the bin",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:56:44+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9921259842519685,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "lowering arm toward tomato",
+      "description": "lowering arm toward tomato"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9921259842519685,
+      "t_end_sec": 5.984251968503937,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "closing gripper around tomato",
+      "description": "closing gripper around tomato"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.984251968503937,
+      "t_end_sec": 8.051181102362206,
+      "control_step_start": 304,
+      "control_step_end": 408,
+      "policy_step_start": 50,
+      "policy_step_end": 68,
+      "phrase": "remaining stationary over tomato",
+      "description": "remaining stationary over tomato"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 5.374015748031496,
+    "vlm_failure_onset_timestamp": "00:05.4",
+    "vlm_failure_onset_frame": 273,
+    "vlm_failure_onset_step": 273,
+    "vlm_coarse_failure_onset_seconds": 4.0,
+    "vlm_failure_window_start_seconds": 3.5,
+    "vlm_failure_window_end_seconds": 4.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot reaches down to the tomato and closes its gripper around it, but fails to establish a secure grasp to lift the tomato.",
+    "vlm_failure_reason_summary": "failed grasp on tomato; no lift achieved",
+    "vlm_recovery_action": "Re-open the end-effector, re-center the gripper over the tomato, close firmly around it, and lift it toward the bin.",
+    "vlm_recovery_action_summary": "re-align gripper; grasp tomato securely; lift into bin",
+    "vlm_prevention_action": "Ensure precise alignment and proper gripper positioning around the center of the tomato before closing fingers to secure the object.",
+    "vlm_prevention_action_summary": "center gripper on tomato; close fingers firmly; lift object",
+    "vlm_justification": "At around 4.0 seconds into the video, the robot closes its gripper over the tomato on the table but fails to securely grasp or lift it, leaving the tomato on the table for the remainder of the rollout.",
+    "vlm_temporal_justification": "At clip second 222 (03:42), the robot closes its gripper around the tomato, but fails to position its fingers correctly to establish a secure hold, making the grasp failure visibly apparent.",
+    "vlm_video_fps": 50.8,
+    "vlm_video_num_frames": 409,
+    "vlm_video_duration_seconds": 8.031496062992126,
+    "vlm_usage": {
+      "input_tokens": 21570,
+      "output_tokens": 367,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 4.0,\n \"failure_window_start_seconds\": 3.5,\n \"failure_window_end_seconds\": 4.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot reaches down to the tomato and closes its gripper around it, but fails to establish a secure grasp to lift the tomato.\",\n \"failure_reason_summary\": \"failed grasp on tomato; no lift achieved\",\n \"recovery_action\": \"Re-open the end-effector, re-center the gripper over the tomato, close firmly around it, and lift it toward the bin.\",\n \"recovery_action_summary\": \"re-align gripper; grasp tomato securely; lift into bin\",\n \"prevention_action\": \"Ensure precise alignment and proper gripper positioning around the center of the tomato before closing fingers to secure the object.\",\n \"prevention_action_summary\": \"center gripper on tomato; close fingers firmly; lift object\",\n \"justification\": \"At around 4.0 seconds into the video, the robot closes its gripper over the tomato on the table but fails to securely grasp or lift it, leaving the tomato on the table for the remainder of the rollout.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 273,
+    "failure_sim_step": 273,
+    "failure_policy_step": 45,
+    "failure_chunk_index": 3,
+    "first_post_failure_policy_step": 46,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-open the end-effector, re-center the gripper over the tomato, close firmly around it, and lift it toward the bin."
+  }
+}
\ No newline at end of file
diff --git a/tomato/ep013/labels.npz b/tomato/ep013/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..4c907b8c6e8bf476031f783be1951e5deadc2310
--- /dev/null
+++ b/tomato/ep013/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ec8a18e3415142e812d1c5d73837bd4c7cf4edd9bf9fab1ffce04a2d42bfc108
+size 3771
diff --git a/tomato/ep014/example.md b/tomato/ep014/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..2734b5436827346739c4d1a46a715b3397371ebc
--- /dev/null
+++ b/tomato/ep014/example.md
@@ -0,0 +1,23 @@
+# t003__ep014__tomato
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the tomato in the bin
+- **success:** True
+- **control steps:** 693  (13.7s at 51 Hz)
+- **policy inferences:** 119  (replan_steps=6)
+- **video:** `tomato/ep014/rollout.mp4`
+- **features:** `119` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching for tomato |
+| 3.0–6.0 | 152–303 | 25–50 | grasping tomato |
+| 6.0–9.0 | 304–455 | 50–75 | lifting tomato |
+| 9.0–12.0 | 456–607 | 76–101 | placing tomato into bin |
+| 12.0–13.7 | 608–692 | 101–115 | retracting arm |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/tomato/ep014/labels.json b/tomato/ep014/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..2eaba474a1e9df6f869e6d605816125d7a036e05
--- /dev/null
+++ b/tomato/ep014/labels.json
@@ -0,0 +1,90 @@
+{
+  "rollout_id": "t003__ep014__tomato",
+  "scene_variant": "",
+  "task_id": 3,
+  "task_file": "",
+  "instruction": "Put the tomato in the bin",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:56:56+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9980276134122286,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for tomato",
+      "description": "reaching for tomato"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9980276134122286,
+      "t_end_sec": 5.996055226824457,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "grasping tomato",
+      "description": "grasping tomato"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.996055226824457,
+      "t_end_sec": 8.994082840236686,
+      "control_step_start": 304,
+      "control_step_end": 455,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "lifting tomato",
+      "description": "lifting tomato"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.994082840236686,
+      "t_end_sec": 11.992110453648914,
+      "control_step_start": 456,
+      "control_step_end": 607,
+      "policy_step_start": 76,
+      "policy_step_end": 101,
+      "phrase": "placing tomato into bin",
+      "description": "placing tomato into bin"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 11.992110453648914,
+      "t_end_sec": 13.668639053254438,
+      "control_step_start": 608,
+      "control_step_end": 692,
+      "policy_step_start": 101,
+      "policy_step_end": 115,
+      "phrase": "retracting arm",
+      "description": "retracting arm"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/tomato/ep014/labels.npz b/tomato/ep014/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..c16c83436e24397779f1ee299f39587a31560192
--- /dev/null
+++ b/tomato/ep014/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ed7056261a2b126bfa761d3d79d6a86506a342f855612a4eeab2cf703612d159
+size 3118
diff --git a/tomato/ep015/example.md b/tomato/ep015/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..cc995fdc349157cd555e339ee307e5daa7edf9ed
--- /dev/null
+++ b/tomato/ep015/example.md
@@ -0,0 +1,24 @@
+# t003__ep015__tomato
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the tomato in the bin
+- **success:** True
+- **control steps:** 875  (17.3s at 51 Hz)
+- **policy inferences:** 150  (replan_steps=6)
+- **video:** `tomato/ep015/rollout.mp4`
+- **features:** `150` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching for tomato |
+| 3.0–6.0 | 152–303 | 25–50 | grasping tomato |
+| 6.0–9.0 | 304–455 | 50–75 | securing grip on tomato |
+| 9.0–12.0 | 456–607 | 76–101 | lifting tomato |
+| 12.0–15.0 | 608–759 | 101–126 | moving tomato to bin |
+| 15.0–17.3 | 760–874 | 126–145 | releasing tomato and retracting |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/tomato/ep015/labels.json b/tomato/ep015/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..9a7f8c08c724076de089ca6a307d261d3dfea3f0
--- /dev/null
+++ b/tomato/ep015/labels.json
@@ -0,0 +1,101 @@
+{
+  "rollout_id": "t003__ep015__tomato",
+  "scene_variant": "",
+  "task_id": 3,
+  "task_file": "",
+  "instruction": "Put the tomato in the bin",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:57:08+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0039525691699605,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for tomato",
+      "description": "reaching for tomato"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0039525691699605,
+      "t_end_sec": 6.007905138339921,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "grasping tomato",
+      "description": "grasping tomato"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.007905138339921,
+      "t_end_sec": 9.011857707509881,
+      "control_step_start": 304,
+      "control_step_end": 455,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "securing grip on tomato",
+      "description": "securing grip on tomato"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.011857707509881,
+      "t_end_sec": 12.015810276679842,
+      "control_step_start": 456,
+      "control_step_end": 607,
+      "policy_step_start": 76,
+      "policy_step_end": 101,
+      "phrase": "lifting tomato",
+      "description": "lifting tomato"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 12.015810276679842,
+      "t_end_sec": 15.019762845849803,
+      "control_step_start": 608,
+      "control_step_end": 759,
+      "policy_step_start": 101,
+      "policy_step_end": 126,
+      "phrase": "moving tomato to bin",
+      "description": "moving tomato to bin"
+    },
+    {
+      "segment_index": 5,
+      "t_start_sec": 15.019762845849803,
+      "t_end_sec": 17.292490118577074,
+      "control_step_start": 760,
+      "control_step_end": 874,
+      "policy_step_start": 126,
+      "policy_step_end": 145,
+      "phrase": "releasing tomato and retracting",
+      "description": "releasing tomato and retracting"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/tomato/ep015/labels.npz b/tomato/ep015/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..742069fadd7ccff5edb36405eb1509c62f75ead7
--- /dev/null
+++ b/tomato/ep015/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8249ed2519e754d5e4ad98cbb981213d1a8e23c27cc2775534a2e59a5d4047ba
+size 3168
diff --git a/tomato/ep016/example.md b/tomato/ep016/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..95690b85a2ce9d2f5eaa820613e2e881470e1494
--- /dev/null
+++ b/tomato/ep016/example.md
@@ -0,0 +1,33 @@
+# t003__ep016__tomato
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the tomato in the bin
+- **success:** False
+- **control steps:** 354  (7.0s at 51 Hz)
+- **policy inferences:** 61  (replan_steps=6)
+- **video:** `tomato/ep016/rollout.mp4`
+- **features:** `61` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | idle; no motion |
+| 3.0–6.0 | 152–303 | 25–50 | idle; arm stationary |
+| 6.0–7.0 | 304–353 | 50–58 | idle; remaining motionless |
+
+## Failure
+
+- **yes**
+- **onset type:** operator_intervention
+- **onset:** 0.0 s (00:00.0)
+- **frame / control_step:** 0 (policy_step=0, chunk_index=0)
+- **coarse onset:** 0.0 s (refined=True)
+- **mode:** stuck_or_no_progress
+- **reason:** The robot arm remains completely motionless throughout the entire episode, failing to initiate any movement toward the tomato.
+  Summary: robot frozen; no movement towards object
+- **recovery:** Re-initialize the robot controller and command the end-effector to lower, grasp the tomato, and place it into the bin.
+  Summary: re-initialize policy; grasp tomato; place in bin
+- **prevention:** Ensure action commands are correctly generated and passed to the controller at the start of the rollout so the robot begins approaching the tomato immediately.
+  Summary: send action commands; approach tomato; navigate to object
+- **justification:** The robot end-effector remains completely static from t=0.0s to the end of the video, making no attempt to reach or pick up the tomato.
diff --git a/tomato/ep016/labels.json b/tomato/ep016/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..ab910248abeca06f5110ec97a6f933540c8ff207
--- /dev/null
+++ b/tomato/ep016/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t003__ep016__tomato",
+  "scene_variant": "",
+  "task_id": 3,
+  "task_file": "",
+  "instruction": "Put the tomato in the bin",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:57:31+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9980276134122286,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "idle; no motion",
+      "description": "idle; no motion"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9980276134122286,
+      "t_end_sec": 5.996055226824457,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "idle; arm stationary",
+      "description": "idle; arm stationary"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.996055226824457,
+      "t_end_sec": 6.982248520710058,
+      "control_step_start": 304,
+      "control_step_end": 353,
+      "policy_step_start": 50,
+      "policy_step_end": 58,
+      "phrase": "idle; remaining motionless",
+      "description": "idle; remaining motionless"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "stuck_or_no_progress",
+    "vlm_failure_onset_type": "operator_intervention",
+    "vlm_failure_onset_seconds": 0.0,
+    "vlm_failure_onset_timestamp": "00:00.0",
+    "vlm_failure_onset_frame": 0,
+    "vlm_failure_onset_step": 0,
+    "vlm_coarse_failure_onset_seconds": 0.0,
+    "vlm_failure_window_start_seconds": 0.0,
+    "vlm_failure_window_end_seconds": 6.96,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot arm remains completely motionless throughout the entire episode, failing to initiate any movement toward the tomato.",
+    "vlm_failure_reason_summary": "robot frozen; no movement towards object",
+    "vlm_recovery_action": "Re-initialize the robot controller and command the end-effector to lower, grasp the tomato, and place it into the bin.",
+    "vlm_recovery_action_summary": "re-initialize policy; grasp tomato; place in bin",
+    "vlm_prevention_action": "Ensure action commands are correctly generated and passed to the controller at the start of the rollout so the robot begins approaching the tomato immediately.",
+    "vlm_prevention_action_summary": "send action commands; approach tomato; navigate to object",
+    "vlm_justification": "The robot end-effector remains completely static from t=0.0s to the end of the video, making no attempt to reach or pick up the tomato.",
+    "vlm_temporal_justification": "The robot arm remains completely stationary starting from the very first frame (0) and never attempts to move towards the tomato throughout the clip.",
+    "vlm_video_fps": 50.7,
+    "vlm_video_num_frames": 354,
+    "vlm_video_duration_seconds": 6.962524654832347,
+    "vlm_usage": {
+      "input_tokens": 11925,
+      "output_tokens": 335,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"stuck_or_no_progress\",\n \"onset_type\": \"operator_intervention\",\n \"failure_onset_seconds\": 0.0,\n \"failure_window_start_seconds\": 0.0,\n \"failure_window_end_seconds\": 6.96,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot arm remains completely motionless throughout the entire episode, failing to initiate any movement toward the tomato.\",\n \"failure_reason_summary\": \"robot frozen; no movement towards object\",\n \"recovery_action\": \"Re-initialize the robot controller and command the end-effector to lower, grasp the tomato, and place it into the bin.\",\n \"recovery_action_summary\": \"re-initialize policy; grasp tomato; place in bin\",\n \"prevention_action\": \"Ensure action commands are correctly generated and passed to the controller at the start of the rollout so the robot begins approaching the tomato immediately.\",\n \"prevention_action_summary\": \"send action commands; approach tomato; navigate to object\",\n \"justification\": \"The robot end-effector remains completely static from t=0.0s to the end of the video, making no attempt to reach or pick up the tomato.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 0,
+    "failure_sim_step": 0,
+    "failure_policy_step": 0,
+    "failure_chunk_index": 0,
+    "first_post_failure_policy_step": 1,
+    "failure_type": "stuck_or_no_progress",
+    "correction_action": "Re-initialize the robot controller and command the end-effector to lower, grasp the tomato, and place it into the bin."
+  }
+}
\ No newline at end of file
diff --git a/tomato/ep016/labels.npz b/tomato/ep016/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..64363dfcdd72fce309a89798b85d51a47dc01aff
--- /dev/null
+++ b/tomato/ep016/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:48f463d58249e89711122d9876d8234c855c9e785652ff9112a2c5549be9e904
+size 3767
diff --git a/tomato/ep017/example.md b/tomato/ep017/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..dfcabf373b7168a0af0573f0b117bb70c405518f
--- /dev/null
+++ b/tomato/ep017/example.md
@@ -0,0 +1,35 @@
+# t003__ep017__tomato
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the tomato in the bin
+- **success:** False
+- **control steps:** 639  (12.7s at 50 Hz)
+- **policy inferences:** 110  (replan_steps=6)
+- **video:** `tomato/ep017/rollout.mp4`
+- **features:** `110` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | approaching and targeting tomato |
+| 3.0–6.0 | 151–301 | 25–50 | closing gripper on tomato |
+| 6.0–9.0 | 302–452 | 50–75 | holding position near tomato |
+| 9.0–12.0 | 453–603 | 75–100 | hovering over tomato with closed gripper |
+| 12.0–12.7 | 604–638 | 100–106 | opening gripper and raising arm |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 2.191235059760956 s (00:02.2)
+- **frame / control_step:** 110 (policy_step=18, chunk_index=2)
+- **coarse onset:** 3.2 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to pick up the tomato, but closed its gripper without securing the object, leaving the tomato on the table.
+  Summary: failed grasp on tomato; improper gripper position
+- **recovery:** Reopen the gripper, realign the end-effector directly over the tomato, descend around it, and close the fingers firmly.
+  Summary: reopen gripper; realign over tomato; grasp firmly
+- **prevention:** Ensure precise positioning around the center of the tomato before closing the gripper to guarantee a secure grasp.
+  Summary: center gripper over tomato; grasp securely
+- **justification:** At approximately 3.2 seconds, the robot attempts to close its gripper around the tomato but fails to acquire a hold. It remains stuck at the tomato's location until lifting away empty-handed.
diff --git a/tomato/ep017/labels.json b/tomato/ep017/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..108ba34d527c8b4b75c54222a9e13ba264ad4aba
--- /dev/null
+++ b/tomato/ep017/labels.json
@@ -0,0 +1,120 @@
+{
+  "rollout_id": "t003__ep017__tomato",
+  "scene_variant": "",
+  "task_id": 3,
+  "task_file": "",
+  "instruction": "Put the tomato in the bin",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:58:06+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00796812749004,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "approaching and targeting tomato",
+      "description": "approaching and targeting tomato"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00796812749004,
+      "t_end_sec": 6.01593625498008,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "closing gripper on tomato",
+      "description": "closing gripper on tomato"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01593625498008,
+      "t_end_sec": 9.02390438247012,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "holding position near tomato",
+      "description": "holding position near tomato"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.02390438247012,
+      "t_end_sec": 12.03187250996016,
+      "control_step_start": 453,
+      "control_step_end": 603,
+      "policy_step_start": 75,
+      "policy_step_end": 100,
+      "phrase": "hovering over tomato with closed gripper",
+      "description": "hovering over tomato with closed gripper"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 12.03187250996016,
+      "t_end_sec": 12.729083665338644,
+      "control_step_start": 604,
+      "control_step_end": 638,
+      "policy_step_start": 100,
+      "policy_step_end": 106,
+      "phrase": "opening gripper and raising arm",
+      "description": "opening gripper and raising arm"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 2.191235059760956,
+    "vlm_failure_onset_timestamp": "00:02.2",
+    "vlm_failure_onset_frame": 110,
+    "vlm_failure_onset_step": 110,
+    "vlm_coarse_failure_onset_seconds": 3.2,
+    "vlm_failure_window_start_seconds": 2.7,
+    "vlm_failure_window_end_seconds": 4.2,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to pick up the tomato, but closed its gripper without securing the object, leaving the tomato on the table.",
+    "vlm_failure_reason_summary": "failed grasp on tomato; improper gripper position",
+    "vlm_recovery_action": "Reopen the gripper, realign the end-effector directly over the tomato, descend around it, and close the fingers firmly.",
+    "vlm_recovery_action_summary": "reopen gripper; realign over tomato; grasp firmly",
+    "vlm_prevention_action": "Ensure precise positioning around the center of the tomato before closing the gripper to guarantee a secure grasp.",
+    "vlm_prevention_action_summary": "center gripper over tomato; grasp securely",
+    "vlm_justification": "At approximately 3.2 seconds, the robot attempts to close its gripper around the tomato but fails to acquire a hold. It remains stuck at the tomato's location until lifting away empty-handed.",
+    "vlm_temporal_justification": "At 01:40 into the clip, the robot gripper fully closes shut without securing or holding the tomato, leaving it resting untouched on the table.",
+    "vlm_video_fps": 50.2,
+    "vlm_video_num_frames": 639,
+    "vlm_video_duration_seconds": 12.709163346613545,
+    "vlm_usage": {
+      "input_tokens": 21760,
+      "output_tokens": 342,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 3.2,\n \"failure_window_start_seconds\": 2.7,\n \"failure_window_end_seconds\": 4.2,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to pick up the tomato, but closed its gripper without securing the object, leaving the tomato on the table.\",\n \"failure_reason_summary\": \"failed grasp on tomato; improper gripper position\",\n \"recovery_action\": \"Reopen the gripper, realign the end-effector directly over the tomato, descend around it, and close the fingers firmly.\",\n \"recovery_action_summary\": \"reopen gripper; realign over tomato; grasp firmly\",\n \"prevention_action\": \"Ensure precise positioning around the center of the tomato before closing the gripper to guarantee a secure grasp.\",\n \"prevention_action_summary\": \"center gripper over tomato; grasp securely\",\n \"justification\": \"At approximately 3.2 seconds, the robot attempts to close its gripper around the tomato but fails to acquire a hold. It remains stuck at the tomato's location until lifting away empty-handed.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 110,
+    "failure_sim_step": 110,
+    "failure_policy_step": 18,
+    "failure_chunk_index": 2,
+    "first_post_failure_policy_step": 19,
+    "failure_type": "grasp_failure",
+    "correction_action": "Reopen the gripper, realign the end-effector directly over the tomato, descend around it, and close the fingers firmly."
+  }
+}
\ No newline at end of file
diff --git a/tomato/ep017/labels.npz b/tomato/ep017/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..375bb6ca42b4b3832be17bf8f53de3fecaaf9bd5
--- /dev/null
+++ b/tomato/ep017/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:022157fbcf8d3979b888fff7bf258ea34095ceeb7942b47cadc49208cc5571fe
+size 3838
diff --git a/tomato/ep018/example.md b/tomato/ep018/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..cb7d945291bff1888d378ce96407c91dbbfa3aab
--- /dev/null
+++ b/tomato/ep018/example.md
@@ -0,0 +1,35 @@
+# t003__ep018__tomato
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the tomato in the bin
+- **success:** False
+- **control steps:** 708  (14.0s at 50 Hz)
+- **policy inferences:** 123  (replan_steps=6)
+- **video:** `tomato/ep018/rollout.mp4`
+- **features:** `123` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | lowering arm toward table |
+| 3.0–6.0 | 151–301 | 25–50 | approaching tomato with gripper |
+| 6.0–9.0 | 302–452 | 50–75 | closing gripper on tomato |
+| 9.0–12.0 | 453–603 | 75–100 | lifting arm and dropping tomato |
+| 12.0–14.0 | 604–707 | 100–117 | lifting empty gripper |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 12.996031746031747 s (00:13.0)
+- **frame / control_step:** 655 (policy_step=109, chunk_index=1)
+- **coarse onset:** 11.5 s (refined=True)
+- **mode:** object_displacement
+- **reason:** The robot grasped the tomato and attempted to lift it, but lost its hold, causing the tomato to drop back onto the table surface.
+  Summary: dropping gripped tomato; lost grasp while lifting
+- **recovery:** Re-center the gripper over the dropped tomato, establish a secure grasp, and lift it into the bin.
+  Summary: re-grasp tomato; place in bin
+- **prevention:** Ensure the gripper securely aligns with and closes tightly around the tomato before initiating the upward motion.
+  Summary: secure firm grasp; lift tomato carefully
+- **justification:** The robot successfully reaches the tomato and closes its fingers on it, but as it lifts the arm at around 11.5 seconds, the tomato slips out of the gripper and falls back onto the desk, resulting in task failure.
diff --git a/tomato/ep018/labels.json b/tomato/ep018/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..ee171cb85e1063d68e9f75f6ac86039cd0032c01
--- /dev/null
+++ b/tomato/ep018/labels.json
@@ -0,0 +1,120 @@
+{
+  "rollout_id": "t003__ep018__tomato",
+  "scene_variant": "",
+  "task_id": 3,
+  "task_file": "",
+  "instruction": "Put the tomato in the bin",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:58:36+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.996031746031746,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "lowering arm toward table",
+      "description": "lowering arm toward table"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.996031746031746,
+      "t_end_sec": 5.992063492063492,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "approaching tomato with gripper",
+      "description": "approaching tomato with gripper"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.992063492063492,
+      "t_end_sec": 8.988095238095239,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "closing gripper on tomato",
+      "description": "closing gripper on tomato"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.988095238095239,
+      "t_end_sec": 11.984126984126984,
+      "control_step_start": 453,
+      "control_step_end": 603,
+      "policy_step_start": 75,
+      "policy_step_end": 100,
+      "phrase": "lifting arm and dropping tomato",
+      "description": "lifting arm and dropping tomato"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 11.984126984126984,
+      "t_end_sec": 14.047619047619047,
+      "control_step_start": 604,
+      "control_step_end": 707,
+      "policy_step_start": 100,
+      "policy_step_end": 117,
+      "phrase": "lifting empty gripper",
+      "description": "lifting empty gripper"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "object_displacement",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 12.996031746031747,
+    "vlm_failure_onset_timestamp": "00:13.0",
+    "vlm_failure_onset_frame": 655,
+    "vlm_failure_onset_step": 655,
+    "vlm_coarse_failure_onset_seconds": 11.5,
+    "vlm_failure_window_start_seconds": 10.5,
+    "vlm_failure_window_end_seconds": 12.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot grasped the tomato and attempted to lift it, but lost its hold, causing the tomato to drop back onto the table surface.",
+    "vlm_failure_reason_summary": "dropping gripped tomato; lost grasp while lifting",
+    "vlm_recovery_action": "Re-center the gripper over the dropped tomato, establish a secure grasp, and lift it into the bin.",
+    "vlm_recovery_action_summary": "re-grasp tomato; place in bin",
+    "vlm_prevention_action": "Ensure the gripper securely aligns with and closes tightly around the tomato before initiating the upward motion.",
+    "vlm_prevention_action_summary": "secure firm grasp; lift tomato carefully",
+    "vlm_justification": "The robot successfully reaches the tomato and closes its fingers on it, but as it lifts the arm at around 11.5 seconds, the tomato slips out of the gripper and falls back onto the desk, resulting in task failure.",
+    "vlm_temporal_justification": "At 03:46 into the clip (226 seconds), the tomato slips out of the robot gripper's grasp while being lifted and falls back onto the table.",
+    "vlm_video_fps": 50.4,
+    "vlm_video_num_frames": 708,
+    "vlm_video_duration_seconds": 14.027777777777779,
+    "vlm_usage": {
+      "input_tokens": 20317,
+      "output_tokens": 350,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"object_displacement\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 11.5,\n \"failure_window_start_seconds\": 10.5,\n \"failure_window_end_seconds\": 12.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot grasped the tomato and attempted to lift it, but lost its hold, causing the tomato to drop back onto the table surface.\",\n \"failure_reason_summary\": \"dropping gripped tomato; lost grasp while lifting\",\n \"recovery_action\": \"Re-center the gripper over the dropped tomato, establish a secure grasp, and lift it into the bin.\",\n \"recovery_action_summary\": \"re-grasp tomato; place in bin\",\n \"prevention_action\": \"Ensure the gripper securely aligns with and closes tightly around the tomato before initiating the upward motion.\",\n \"prevention_action_summary\": \"secure firm grasp; lift tomato carefully\",\n \"justification\": \"The robot successfully reaches the tomato and closes its fingers on it, but as it lifts the arm at around 11.5 seconds, the tomato slips out of the gripper and falls back onto the desk, resulting in task failure.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 655,
+    "failure_sim_step": 655,
+    "failure_policy_step": 109,
+    "failure_chunk_index": 1,
+    "first_post_failure_policy_step": 110,
+    "failure_type": "object_displacement",
+    "correction_action": "Re-center the gripper over the dropped tomato, establish a secure grasp, and lift it into the bin."
+  }
+}
\ No newline at end of file
diff --git a/tomato/ep018/labels.npz b/tomato/ep018/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..3bce96362604a081990e692202146708b1f07181
--- /dev/null
+++ b/tomato/ep018/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c068c2d67445b7bb8d4b3b2da185ff4559d756b2f0b6eb74d644cf58e63ed50e
+size 3779
diff --git a/tomato/ep019/example.md b/tomato/ep019/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..acf1c0268031c218527a4e351800b1a40714ad03
--- /dev/null
+++ b/tomato/ep019/example.md
@@ -0,0 +1,22 @@
+# t003__ep019__tomato
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the tomato in the bin
+- **success:** True
+- **control steps:** 564  (11.2s at 50 Hz)
+- **policy inferences:** 97  (replan_steps=6)
+- **video:** `tomato/ep019/rollout.mp4`
+- **features:** `97` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | reaching for tomato |
+| 3.0–6.0 | 151–301 | 25–50 | grasping the tomato |
+| 6.0–9.0 | 302–452 | 50–75 | moving tomato to bin |
+| 9.0–11.2 | 453–563 | 75–93 | releasing tomato and retracting |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/tomato/ep019/labels.json b/tomato/ep019/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..af2b9d3c08b4ea52a17ad3a8615f46417a452f91
--- /dev/null
+++ b/tomato/ep019/labels.json
@@ -0,0 +1,79 @@
+{
+  "rollout_id": "t003__ep019__tomato",
+  "scene_variant": "",
+  "task_id": 3,
+  "task_file": "",
+  "instruction": "Put the tomato in the bin",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:58:44+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00796812749004,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for tomato",
+      "description": "reaching for tomato"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00796812749004,
+      "t_end_sec": 6.01593625498008,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "grasping the tomato",
+      "description": "grasping the tomato"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01593625498008,
+      "t_end_sec": 9.02390438247012,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "moving tomato to bin",
+      "description": "moving tomato to bin"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.02390438247012,
+      "t_end_sec": 11.235059760956174,
+      "control_step_start": 453,
+      "control_step_end": 563,
+      "policy_step_start": 75,
+      "policy_step_end": 93,
+      "phrase": "releasing tomato and retracting",
+      "description": "releasing tomato and retracting"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/tomato/ep019/labels.npz b/tomato/ep019/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..689d7f28cda7e591be6981bdb99ef8cbed024f15
--- /dev/null
+++ b/tomato/ep019/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:aec37d66a873f3a44827bb5a43198ca0f2f8c8a805b1b30ed0ade67056b19820
+size 3106
diff --git a/tomato/ep020/example.md b/tomato/ep020/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..0490e982d4292e438b56a5623455309b000a5f83
--- /dev/null
+++ b/tomato/ep020/example.md
@@ -0,0 +1,23 @@
+# t003__ep020__tomato
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the tomato in the bin
+- **success:** True
+- **control steps:** 639  (12.7s at 50 Hz)
+- **policy inferences:** 110  (replan_steps=6)
+- **video:** `tomato/ep020/rollout.mp4`
+- **features:** `110` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching for the tomato |
+| 3.0–6.0 | 152–303 | 25–50 | grasping the tomato |
+| 6.0–9.0 | 304–455 | 50–75 | moving tomato over bin |
+| 9.0–12.0 | 456–607 | 76–101 | releasing tomato into bin |
+| 12.0–12.7 | 608–638 | 101–106 | retracting robot arm |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/tomato/ep020/labels.json b/tomato/ep020/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..d9265cfdbe6c8be615ec67ebfa7c9c399e475dbc
--- /dev/null
+++ b/tomato/ep020/labels.json
@@ -0,0 +1,90 @@
+{
+  "rollout_id": "t003__ep020__tomato",
+  "scene_variant": "",
+  "task_id": 3,
+  "task_file": "",
+  "instruction": "Put the tomato in the bin",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:58:53+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00990099009901,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for the tomato",
+      "description": "reaching for the tomato"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00990099009901,
+      "t_end_sec": 6.01980198019802,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "grasping the tomato",
+      "description": "grasping the tomato"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01980198019802,
+      "t_end_sec": 9.029702970297029,
+      "control_step_start": 304,
+      "control_step_end": 455,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "moving tomato over bin",
+      "description": "moving tomato over bin"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.029702970297029,
+      "t_end_sec": 12.03960396039604,
+      "control_step_start": 456,
+      "control_step_end": 607,
+      "policy_step_start": 76,
+      "policy_step_end": 101,
+      "phrase": "releasing tomato into bin",
+      "description": "releasing tomato into bin"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 12.03960396039604,
+      "t_end_sec": 12.653465346534654,
+      "control_step_start": 608,
+      "control_step_end": 638,
+      "policy_step_start": 101,
+      "policy_step_end": 106,
+      "phrase": "retracting robot arm",
+      "description": "retracting robot arm"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/tomato/ep020/labels.npz b/tomato/ep020/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..357ccc3c73dfc7f76629ff3c8e03a9717d2d105e
--- /dev/null
+++ b/tomato/ep020/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4137d7b1b7f70e6bd51ab2da97ff4b23e72291a4208f6333e80ec0292988472a
+size 3142
diff --git a/tomato/ep021/example.md b/tomato/ep021/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..cfc10ef63b218a733196d49b238aecbbec7f4775
--- /dev/null
+++ b/tomato/ep021/example.md
@@ -0,0 +1,34 @@
+# t003__ep021__tomato
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the tomato in the bin
+- **success:** False
+- **control steps:** 555  (11.1s at 50 Hz)
+- **policy inferences:** 97  (replan_steps=6)
+- **video:** `tomato/ep021/rollout.mp4`
+- **features:** `97` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | reaching toward tomato |
+| 3.0–6.0 | 151–301 | 25–50 | lowering gripper over tomato |
+| 6.0–9.0 | 302–452 | 50–75 | lifting empty gripper from table |
+| 9.0–11.1 | 453–554 | 75–92 | moving empty gripper over bin |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 6.49402390438247 s (00:06.5)
+- **frame / control_step:** 326 (policy_step=54, chunk_index=2)
+- **coarse onset:** 6.0 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the tomato, but the gripper failed to establish a secure hold, leaving the tomato behind on the table as the arm lifted away.
+  Summary: unsuccessful grasp attempt; leaving tomato on table
+- **recovery:** Lower the arm back over the tomato, align the gripper properly, firmly grasp the tomato, and transport it to the bin.
+  Summary: re-align gripper; regrasp tomato; place in bin
+- **prevention:** Center the gripper accurately around the tomato and ensure a firm grip before lifting the end-effector.
+  Summary: center gripper over tomato; secure firm grasp
+- **justification:** Around 6.0 seconds, the gripper closes on the tomato but slips off without securing it. The arm then rises and moves toward the bin empty while the tomato remains on the table.
diff --git a/tomato/ep021/labels.json b/tomato/ep021/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..57e404c7d349f5a831dd54c939ec6f9a7f063e7b
--- /dev/null
+++ b/tomato/ep021/labels.json
@@ -0,0 +1,109 @@
+{
+  "rollout_id": "t003__ep021__tomato",
+  "scene_variant": "",
+  "task_id": 3,
+  "task_file": "",
+  "instruction": "Put the tomato in the bin",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:59:32+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00796812749004,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching toward tomato",
+      "description": "reaching toward tomato"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00796812749004,
+      "t_end_sec": 6.01593625498008,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "lowering gripper over tomato",
+      "description": "lowering gripper over tomato"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01593625498008,
+      "t_end_sec": 9.02390438247012,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "lifting empty gripper from table",
+      "description": "lifting empty gripper from table"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.02390438247012,
+      "t_end_sec": 11.055776892430279,
+      "control_step_start": 453,
+      "control_step_end": 554,
+      "policy_step_start": 75,
+      "policy_step_end": 92,
+      "phrase": "moving empty gripper over bin",
+      "description": "moving empty gripper over bin"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 6.49402390438247,
+    "vlm_failure_onset_timestamp": "00:06.5",
+    "vlm_failure_onset_frame": 326,
+    "vlm_failure_onset_step": 326,
+    "vlm_coarse_failure_onset_seconds": 6.0,
+    "vlm_failure_window_start_seconds": 5.5,
+    "vlm_failure_window_end_seconds": 7.0,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the tomato, but the gripper failed to establish a secure hold, leaving the tomato behind on the table as the arm lifted away.",
+    "vlm_failure_reason_summary": "unsuccessful grasp attempt; leaving tomato on table",
+    "vlm_recovery_action": "Lower the arm back over the tomato, align the gripper properly, firmly grasp the tomato, and transport it to the bin.",
+    "vlm_recovery_action_summary": "re-align gripper; regrasp tomato; place in bin",
+    "vlm_prevention_action": "Center the gripper accurately around the tomato and ensure a firm grip before lifting the end-effector.",
+    "vlm_prevention_action_summary": "center gripper over tomato; secure firm grasp",
+    "vlm_justification": "Around 6.0 seconds, the gripper closes on the tomato but slips off without securing it. The arm then rises and moves toward the bin empty while the tomato remains on the table.",
+    "vlm_temporal_justification": "At second 176 (02:56 in the clip), the robot's gripper fingers close completely around the tomato without establishing a secure hold, leaving the tomato resting ungrasped on the table.",
+    "vlm_video_fps": 50.2,
+    "vlm_video_num_frames": 555,
+    "vlm_video_duration_seconds": 11.035856573705178,
+    "vlm_usage": {
+      "input_tokens": 21643,
+      "output_tokens": 359,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 6.0,\n \"failure_window_start_seconds\": 5.5,\n \"failure_window_end_seconds\": 7.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the tomato, but the gripper failed to establish a secure hold, leaving the tomato behind on the table as the arm lifted away.\",\n \"failure_reason_summary\": \"unsuccessful grasp attempt; leaving tomato on table\",\n \"recovery_action\": \"Lower the arm back over the tomato, align the gripper properly, firmly grasp the tomato, and transport it to the bin.\",\n \"recovery_action_summary\": \"re-align gripper; regrasp tomato; place in bin\",\n \"prevention_action\": \"Center the gripper accurately around the tomato and ensure a firm grip before lifting the end-effector.\",\n \"prevention_action_summary\": \"center gripper over tomato; secure firm grasp\",\n \"justification\": \"Around 6.0 seconds, the gripper closes on the tomato but slips off without securing it. The arm then rises and moves toward the bin empty while the tomato remains on the table.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 326,
+    "failure_sim_step": 326,
+    "failure_policy_step": 54,
+    "failure_chunk_index": 2,
+    "first_post_failure_policy_step": 55,
+    "failure_type": "grasp_failure",
+    "correction_action": "Lower the arm back over the tomato, align the gripper properly, firmly grasp the tomato, and transport it to the bin."
+  }
+}
\ No newline at end of file
diff --git a/tomato/ep021/labels.npz b/tomato/ep021/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..f2e43649c05cb396f75f634a1158c39d7a0908a9
--- /dev/null
+++ b/tomato/ep021/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e9444980dcf08f49bd990f390569e49b7726b262bc207a5f0ad2f18ca54165a5
+size 3788
diff --git a/tomato/ep022/example.md b/tomato/ep022/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..097003e10f91921228edd8de95557599f045c8f8
--- /dev/null
+++ b/tomato/ep022/example.md
@@ -0,0 +1,22 @@
+# t003__ep022__tomato
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the tomato in the bin
+- **success:** True
+- **control steps:** 603  (11.9s at 51 Hz)
+- **policy inferences:** 103  (replan_steps=6)
+- **video:** `tomato/ep022/rollout.mp4`
+- **features:** `103` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching for the tomato |
+| 3.0–6.0 | 152–303 | 25–50 | grasping the tomato |
+| 6.0–9.0 | 304–455 | 50–75 | moving tomato toward bin |
+| 9.0–11.9 | 456–602 | 76–100 | dropping tomato and retracting |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/tomato/ep022/labels.json b/tomato/ep022/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..9af590176258dc9028bd635a0cf254551e3cf6d3
--- /dev/null
+++ b/tomato/ep022/labels.json
@@ -0,0 +1,79 @@
+{
+  "rollout_id": "t003__ep022__tomato",
+  "scene_variant": "",
+  "task_id": 3,
+  "task_file": "",
+  "instruction": "Put the tomato in the bin",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:59:40+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0039525691699605,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for the tomato",
+      "description": "reaching for the tomato"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0039525691699605,
+      "t_end_sec": 6.007905138339921,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "grasping the tomato",
+      "description": "grasping the tomato"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.007905138339921,
+      "t_end_sec": 9.011857707509881,
+      "control_step_start": 304,
+      "control_step_end": 455,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "moving tomato toward bin",
+      "description": "moving tomato toward bin"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.011857707509881,
+      "t_end_sec": 11.91699604743083,
+      "control_step_start": 456,
+      "control_step_end": 602,
+      "policy_step_start": 76,
+      "policy_step_end": 100,
+      "phrase": "dropping tomato and retracting",
+      "description": "dropping tomato and retracting"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/tomato/ep022/labels.npz b/tomato/ep022/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..b80cdce6bff371c119d9aaae8f3b724f6831eeb9
--- /dev/null
+++ b/tomato/ep022/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:011ad8504eb92fdc8604e9dfc0723b7aa6ba870dbc6fa81818df43023f55d7a5
+size 3112
diff --git a/tomato/ep023/example.md b/tomato/ep023/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..9efbca0bc57282d22863d5aa262ba65aab66e390
--- /dev/null
+++ b/tomato/ep023/example.md
@@ -0,0 +1,22 @@
+# t003__ep023__tomato
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the tomato in the bin
+- **success:** True
+- **control steps:** 519  (10.3s at 50 Hz)
+- **policy inferences:** 89  (replan_steps=6)
+- **video:** `tomato/ep023/rollout.mp4`
+- **features:** `89` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | reaching for tomato |
+| 3.0–6.0 | 151–301 | 25–50 | grasping and lifting tomato |
+| 6.0–9.0 | 302–452 | 50–75 | dropping tomato into bin |
+| 9.0–10.3 | 453–518 | 75–86 | retracting to home position |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/tomato/ep023/labels.json b/tomato/ep023/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..d322e64d3eb49d8247662494c9c564e88d06c3b3
--- /dev/null
+++ b/tomato/ep023/labels.json
@@ -0,0 +1,79 @@
+{
+  "rollout_id": "t003__ep023__tomato",
+  "scene_variant": "",
+  "task_id": 3,
+  "task_file": "",
+  "instruction": "Put the tomato in the bin",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:59:49+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00796812749004,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for tomato",
+      "description": "reaching for tomato"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00796812749004,
+      "t_end_sec": 6.01593625498008,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "grasping and lifting tomato",
+      "description": "grasping and lifting tomato"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01593625498008,
+      "t_end_sec": 9.02390438247012,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "dropping tomato into bin",
+      "description": "dropping tomato into bin"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.02390438247012,
+      "t_end_sec": 10.338645418326692,
+      "control_step_start": 453,
+      "control_step_end": 518,
+      "policy_step_start": 75,
+      "policy_step_end": 86,
+      "phrase": "retracting to home position",
+      "description": "retracting to home position"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/tomato/ep023/labels.npz b/tomato/ep023/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..5352e4b786f223ad40b0949439145cbc9a05a01e
--- /dev/null
+++ b/tomato/ep023/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:80212ce1243f30f156b4c1522a10c85a61a80d1c851e769ef4a5c3907b7e4ac7
+size 3116
diff --git a/tomato/ep024/example.md b/tomato/ep024/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..aa84d5632bcbde83e56830a08d5dfccf3553977d
--- /dev/null
+++ b/tomato/ep024/example.md
@@ -0,0 +1,23 @@
+# t003__ep024__tomato
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the tomato in the bin
+- **success:** True
+- **control steps:** 679  (13.5s at 50 Hz)
+- **policy inferences:** 117  (replan_steps=6)
+- **video:** `tomato/ep024/rollout.mp4`
+- **features:** `117` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | reaching for tomato |
+| 3.0–6.0 | 151–301 | 25–50 | grasping the tomato |
+| 6.0–9.0 | 302–452 | 50–75 | moving tomato toward bin |
+| 9.0–12.0 | 453–603 | 75–100 | placing tomato in bin |
+| 12.0–13.5 | 604–678 | 100–113 | retracting arm |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/tomato/ep024/labels.json b/tomato/ep024/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..7800baee95e21d1b1cd546c74d70ef3c1cb80122
--- /dev/null
+++ b/tomato/ep024/labels.json
@@ -0,0 +1,90 @@
+{
+  "rollout_id": "t003__ep024__tomato",
+  "scene_variant": "",
+  "task_id": 3,
+  "task_file": "",
+  "instruction": "Put the tomato in the bin",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T21:59:56+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.996031746031746,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for tomato",
+      "description": "reaching for tomato"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.996031746031746,
+      "t_end_sec": 5.992063492063492,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "grasping the tomato",
+      "description": "grasping the tomato"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.992063492063492,
+      "t_end_sec": 8.988095238095239,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "moving tomato toward bin",
+      "description": "moving tomato toward bin"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.988095238095239,
+      "t_end_sec": 11.984126984126984,
+      "control_step_start": 453,
+      "control_step_end": 603,
+      "policy_step_start": 75,
+      "policy_step_end": 100,
+      "phrase": "placing tomato in bin",
+      "description": "placing tomato in bin"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 11.984126984126984,
+      "t_end_sec": 13.472222222222223,
+      "control_step_start": 604,
+      "control_step_end": 678,
+      "policy_step_start": 100,
+      "policy_step_end": 113,
+      "phrase": "retracting arm",
+      "description": "retracting arm"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/tomato/ep024/labels.npz b/tomato/ep024/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..65c282b789ce85271a208851f03ea2173db98326
--- /dev/null
+++ b/tomato/ep024/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:203830570bf21451d39ca517ff049e67bdcd5f89e8f39e91d721341d77ce43a8
+size 3136
diff --git a/tomato/ep025/example.md b/tomato/ep025/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..ef7ac95f2e78518fe1f2a657257a1b11b3aa941d
--- /dev/null
+++ b/tomato/ep025/example.md
@@ -0,0 +1,22 @@
+# t003__ep025__tomato
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the tomato in the bin
+- **success:** True
+- **control steps:** 463  (9.2s at 50 Hz)
+- **policy inferences:** 80  (replan_steps=6)
+- **video:** `tomato/ep025/rollout.mp4`
+- **features:** `80` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | reaching for tomato |
+| 3.0–6.0 | 151–301 | 25–50 | grasping and lifting tomato |
+| 6.0–9.0 | 302–452 | 50–75 | placing tomato into bin |
+| 9.0–9.2 | 453–462 | 75–77 | retracting arm |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/tomato/ep025/labels.json b/tomato/ep025/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..9efd49ea6305944b1894bdf9ef1f8a738ab94a20
--- /dev/null
+++ b/tomato/ep025/labels.json
@@ -0,0 +1,79 @@
+{
+  "rollout_id": "t003__ep025__tomato",
+  "scene_variant": "",
+  "task_id": 3,
+  "task_file": "",
+  "instruction": "Put the tomato in the bin",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T22:00:05+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.996031746031746,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for tomato",
+      "description": "reaching for tomato"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.996031746031746,
+      "t_end_sec": 5.992063492063492,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "grasping and lifting tomato",
+      "description": "grasping and lifting tomato"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.992063492063492,
+      "t_end_sec": 8.988095238095239,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "placing tomato into bin",
+      "description": "placing tomato into bin"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.988095238095239,
+      "t_end_sec": 9.186507936507937,
+      "control_step_start": 453,
+      "control_step_end": 462,
+      "policy_step_start": 75,
+      "policy_step_end": 77,
+      "phrase": "retracting arm",
+      "description": "retracting arm"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/tomato/ep025/labels.npz b/tomato/ep025/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..5db2fcbae0c3237ff29fe4107eac3cd52dbc6d49
--- /dev/null
+++ b/tomato/ep025/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d5e0d85aa2f81f948c34f6a52841a3d784a46470d2e30444e399565f8c5a7b27
+size 3103
diff --git a/tomato/ep026/example.md b/tomato/ep026/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..b8cf68519f891f3e9b620056b1fa6ded7f7f9ef5
--- /dev/null
+++ b/tomato/ep026/example.md
@@ -0,0 +1,22 @@
+# t003__ep026__tomato
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the tomato in the bin
+- **success:** True
+- **control steps:** 476  (9.5s at 50 Hz)
+- **policy inferences:** 82  (replan_steps=6)
+- **video:** `tomato/ep026/rollout.mp4`
+- **features:** `82` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | reaching for tomato |
+| 3.0–6.0 | 151–301 | 25–50 | grasping and lifting tomato |
+| 6.0–9.0 | 302–452 | 50–75 | placing tomato into bin |
+| 9.0–9.5 | 453–475 | 75–79 | retracting to idle position |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/tomato/ep026/labels.json b/tomato/ep026/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..b3690940116601aeaac6c443e3de295b0dba4f3f
--- /dev/null
+++ b/tomato/ep026/labels.json
@@ -0,0 +1,79 @@
+{
+  "rollout_id": "t003__ep026__tomato",
+  "scene_variant": "",
+  "task_id": 3,
+  "task_file": "",
+  "instruction": "Put the tomato in the bin",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T22:00:14+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00796812749004,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for tomato",
+      "description": "reaching for tomato"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00796812749004,
+      "t_end_sec": 6.01593625498008,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "grasping and lifting tomato",
+      "description": "grasping and lifting tomato"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01593625498008,
+      "t_end_sec": 9.02390438247012,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "placing tomato into bin",
+      "description": "placing tomato into bin"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.02390438247012,
+      "t_end_sec": 9.482071713147409,
+      "control_step_start": 453,
+      "control_step_end": 475,
+      "policy_step_start": 75,
+      "policy_step_end": 79,
+      "phrase": "retracting to idle position",
+      "description": "retracting to idle position"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/tomato/ep026/labels.npz b/tomato/ep026/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..b662c24b57f5475f2780366605d3500c68849377
--- /dev/null
+++ b/tomato/ep026/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f1b3280c1a22e3867aad9c6cf8813fa59c7650023a8eb9e719732931ffedcdc1
+size 3113
diff --git a/tomato/ep027/example.md b/tomato/ep027/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..5a19ea8590e3808185ec3ddaa67fc113d8c222a2
--- /dev/null
+++ b/tomato/ep027/example.md
@@ -0,0 +1,25 @@
+# t003__ep027__tomato
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the tomato in the bin
+- **success:** True
+- **control steps:** 909  (18.2s at 50 Hz)
+- **policy inferences:** 156  (replan_steps=6)
+- **video:** `tomato/ep027/rollout.mp4`
+- **features:** `156` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–149 | 0–24 | reaching for tomato |
+| 3.0–6.0 | 150–299 | 25–49 | lowering over tomato |
+| 6.0–9.0 | 300–449 | 50–74 | grasping tomato |
+| 9.0–12.0 | 450–599 | 75–99 | carrying tomato toward bin |
+| 12.0–15.0 | 600–749 | 100–124 | placing tomato in bin |
+| 15.0–18.0 | 750–899 | 125–149 | retracting arm |
+| 18.0–18.2 | 900–908 | 150–151 | idle / task complete |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/tomato/ep027/labels.json b/tomato/ep027/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..11cfe9c1d091dd53c4980ae7f21a99c27be56fb8
--- /dev/null
+++ b/tomato/ep027/labels.json
@@ -0,0 +1,112 @@
+{
+  "rollout_id": "t003__ep027__tomato",
+  "scene_variant": "",
+  "task_id": 3,
+  "task_file": "",
+  "instruction": "Put the tomato in the bin",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T22:00:24+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0,
+      "control_step_start": 0,
+      "control_step_end": 149,
+      "policy_step_start": 0,
+      "policy_step_end": 24,
+      "phrase": "reaching for tomato",
+      "description": "reaching for tomato"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0,
+      "t_end_sec": 6.0,
+      "control_step_start": 150,
+      "control_step_end": 299,
+      "policy_step_start": 25,
+      "policy_step_end": 49,
+      "phrase": "lowering over tomato",
+      "description": "lowering over tomato"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.0,
+      "t_end_sec": 9.0,
+      "control_step_start": 300,
+      "control_step_end": 449,
+      "policy_step_start": 50,
+      "policy_step_end": 74,
+      "phrase": "grasping tomato",
+      "description": "grasping tomato"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.0,
+      "t_end_sec": 12.0,
+      "control_step_start": 450,
+      "control_step_end": 599,
+      "policy_step_start": 75,
+      "policy_step_end": 99,
+      "phrase": "carrying tomato toward bin",
+      "description": "carrying tomato toward bin"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 12.0,
+      "t_end_sec": 15.0,
+      "control_step_start": 600,
+      "control_step_end": 749,
+      "policy_step_start": 100,
+      "policy_step_end": 124,
+      "phrase": "placing tomato in bin",
+      "description": "placing tomato in bin"
+    },
+    {
+      "segment_index": 5,
+      "t_start_sec": 15.0,
+      "t_end_sec": 18.0,
+      "control_step_start": 750,
+      "control_step_end": 899,
+      "policy_step_start": 125,
+      "policy_step_end": 149,
+      "phrase": "retracting arm",
+      "description": "retracting arm"
+    },
+    {
+      "segment_index": 6,
+      "t_start_sec": 18.0,
+      "t_end_sec": 18.18,
+      "control_step_start": 900,
+      "control_step_end": 908,
+      "policy_step_start": 150,
+      "policy_step_end": 151,
+      "phrase": "idle / task complete",
+      "description": "idle / task complete"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/tomato/ep027/labels.npz b/tomato/ep027/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..1508c17ba4517d8240024ecc591684630da2f473
--- /dev/null
+++ b/tomato/ep027/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8b72cdb6efc21ce11ea7ece173737a033e4572da0984a0ea772bbe82585488e3
+size 3209
diff --git a/tomato/ep028/example.md b/tomato/ep028/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..8972798c8675494fc368894fde4f0b51eaaccbc1
--- /dev/null
+++ b/tomato/ep028/example.md
@@ -0,0 +1,33 @@
+# t003__ep028__tomato
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the tomato in the bin
+- **success:** False
+- **control steps:** 404  (8.0s at 51 Hz)
+- **policy inferences:** 70  (replan_steps=6)
+- **video:** `tomato/ep028/rollout.mp4`
+- **features:** `70` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching down toward tomato |
+| 3.0–6.0 | 152–303 | 25–50 | closing gripper on tomato |
+| 6.0–8.0 | 304–403 | 50–67 | idle at table surface |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 3.08300395256917 s (00:03.1)
+- **frame / control_step:** 156 (policy_step=26, chunk_index=0)
+- **coarse onset:** 3.8 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot lowered its gripper onto the tomato and closed its fingers, but failed to establish a secure hold or lift the object, remaining stationary at the table surface.
+  Summary: failed grasp on tomato; unable to lift object
+- **recovery:** Open the gripper, re-align the end-effector over the center of the tomato, lower to the correct height, and grasp it securely before lifting.
+  Summary: re-align end-effector; grasp tomato securely
+- **prevention:** Position the gripper directly over the tomato's center at an appropriate height before closing the fingers to ensure a firm grip.
+  Summary: center gripper over tomato; execute firm grasp
+- **justification:** Around 3.8 seconds into the video, the gripper descends onto the tomato and closes its fingers. However, it fails to acquire a secure hold on the tomato and stops advancing the task, leaving the tomato on the table.
diff --git a/tomato/ep028/labels.json b/tomato/ep028/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..d70b79a7c4204af683c9e75e7c7a44ad10563d8f
--- /dev/null
+++ b/tomato/ep028/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t003__ep028__tomato",
+  "scene_variant": "",
+  "task_id": 3,
+  "task_file": "",
+  "instruction": "Put the tomato in the bin",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T22:00:58+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0039525691699605,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching down toward tomato",
+      "description": "reaching down toward tomato"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0039525691699605,
+      "t_end_sec": 6.007905138339921,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "closing gripper on tomato",
+      "description": "closing gripper on tomato"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.007905138339921,
+      "t_end_sec": 7.984189723320158,
+      "control_step_start": 304,
+      "control_step_end": 403,
+      "policy_step_start": 50,
+      "policy_step_end": 67,
+      "phrase": "idle at table surface",
+      "description": "idle at table surface"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 3.08300395256917,
+    "vlm_failure_onset_timestamp": "00:03.1",
+    "vlm_failure_onset_frame": 156,
+    "vlm_failure_onset_step": 156,
+    "vlm_coarse_failure_onset_seconds": 3.8,
+    "vlm_failure_window_start_seconds": 3.3,
+    "vlm_failure_window_end_seconds": 4.3,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot lowered its gripper onto the tomato and closed its fingers, but failed to establish a secure hold or lift the object, remaining stationary at the table surface.",
+    "vlm_failure_reason_summary": "failed grasp on tomato; unable to lift object",
+    "vlm_recovery_action": "Open the gripper, re-align the end-effector over the center of the tomato, lower to the correct height, and grasp it securely before lifting.",
+    "vlm_recovery_action_summary": "re-align end-effector; grasp tomato securely",
+    "vlm_prevention_action": "Position the gripper directly over the tomato's center at an appropriate height before closing the fingers to ensure a firm grip.",
+    "vlm_prevention_action_summary": "center gripper over tomato; execute firm grasp",
+    "vlm_justification": "Around 3.8 seconds into the video, the gripper descends onto the tomato and closes its fingers. However, it fails to acquire a secure hold on the tomato and stops advancing the task, leaving the tomato on the table.",
+    "vlm_temporal_justification": "At clip second 116 (01:56), the robot's end-effector completes its downward motion and closes its fingers around the tomato, but fails to secure or lift the object, remaining frozen at the table surface.",
+    "vlm_video_fps": 50.6,
+    "vlm_video_num_frames": 404,
+    "vlm_video_duration_seconds": 7.9644268774703555,
+    "vlm_usage": {
+      "input_tokens": 21576,
+      "output_tokens": 382,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 3.8,\n \"failure_window_start_seconds\": 3.3,\n \"failure_window_end_seconds\": 4.3,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot lowered its gripper onto the tomato and closed its fingers, but failed to establish a secure hold or lift the object, remaining stationary at the table surface.\",\n \"failure_reason_summary\": \"failed grasp on tomato; unable to lift object\",\n \"recovery_action\": \"Open the gripper, re-align the end-effector over the center of the tomato, lower to the correct height, and grasp it securely before lifting.\",\n \"recovery_action_summary\": \"re-align end-effector; grasp tomato securely\",\n \"prevention_action\": \"Position the gripper directly over the tomato's center at an appropriate height before closing the fingers to ensure a firm grip.\",\n \"prevention_action_summary\": \"center gripper over tomato; execute firm grasp\",\n \"justification\": \"Around 3.8 seconds into the video, the gripper descends onto the tomato and closes its fingers. However, it fails to acquire a secure hold on the tomato and stops advancing the task, leaving the tomato on the table.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 156,
+    "failure_sim_step": 156,
+    "failure_policy_step": 26,
+    "failure_chunk_index": 0,
+    "first_post_failure_policy_step": 27,
+    "failure_type": "grasp_failure",
+    "correction_action": "Open the gripper, re-align the end-effector over the center of the tomato, lower to the correct height, and grasp it securely before lifting."
+  }
+}
\ No newline at end of file
diff --git a/tomato/ep028/labels.npz b/tomato/ep028/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..181b2866f4206054f5649840d935575ed609d907
--- /dev/null
+++ b/tomato/ep028/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:79afa5f7e778f87905bcf6876be34ad8fa9e8c21f126a50ed2cd60fc7eb7fc5b
+size 3819
diff --git a/tomato/ep029/example.md b/tomato/ep029/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..b4bf109841ed2ef6bf1be4d89682f2848e02612f
--- /dev/null
+++ b/tomato/ep029/example.md
@@ -0,0 +1,34 @@
+# t003__ep029__tomato
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the tomato in the bin
+- **success:** False
+- **control steps:** 506  (10.0s at 51 Hz)
+- **policy inferences:** 87  (replan_steps=6)
+- **video:** `tomato/ep029/rollout.mp4`
+- **features:** `87` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | lowering gripper toward tomato |
+| 3.0–6.0 | 152–303 | 25–50 | attempting grasp on tomato |
+| 6.0–9.0 | 304–455 | 50–75 | attempting second tomato grasp |
+| 9.0–10.0 | 456–505 | 76–84 | lifting empty gripper |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 2.2440944881889764 s (00:02.2)
+- **frame / control_step:** 114 (policy_step=19, chunk_index=0)
+- **coarse onset:** 5.0 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot positions its gripper over the tomato and closes its fingers, but fails to establish a secure grasp on the object, leaving the tomato behind on the table.
+  Summary: failed grasp on tomato; incomplete grip closure
+- **recovery:** Re-align the gripper over the center of the tomato, lower to an appropriate height, and close the gripper securely before lifting.
+  Summary: realign over tomato; close gripper securely; pick up tomato
+- **prevention:** Ensure the gripper is properly centered and at the correct height relative to the tomato before closing the fingers to guarantee a firm grasp.
+  Summary: center gripper over tomato; achieve proper depth before grasping
+- **justification:** At around 5.0 seconds (visible in frame 00:05), the robot closes its gripper around the tomato but fails to securely hold it. It then lifts away empty and tries again at 00:07, failing a second time. This is a clear grasp_failure starting at the first failed attempt around 5.0s.
diff --git a/tomato/ep029/labels.json b/tomato/ep029/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..e63ce6775f41394c13a142ed314ab584d052554f
--- /dev/null
+++ b/tomato/ep029/labels.json
@@ -0,0 +1,109 @@
+{
+  "rollout_id": "t003__ep029__tomato",
+  "scene_variant": "",
+  "task_id": 3,
+  "task_file": "",
+  "instruction": "Put the tomato in the bin",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T22:01:33+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9921259842519685,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "lowering gripper toward tomato",
+      "description": "lowering gripper toward tomato"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9921259842519685,
+      "t_end_sec": 5.984251968503937,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "attempting grasp on tomato",
+      "description": "attempting grasp on tomato"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.984251968503937,
+      "t_end_sec": 8.976377952755906,
+      "control_step_start": 304,
+      "control_step_end": 455,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "attempting second tomato grasp",
+      "description": "attempting second tomato grasp"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.976377952755906,
+      "t_end_sec": 9.960629921259843,
+      "control_step_start": 456,
+      "control_step_end": 505,
+      "policy_step_start": 76,
+      "policy_step_end": 84,
+      "phrase": "lifting empty gripper",
+      "description": "lifting empty gripper"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 2.2440944881889764,
+    "vlm_failure_onset_timestamp": "00:02.2",
+    "vlm_failure_onset_frame": 114,
+    "vlm_failure_onset_step": 114,
+    "vlm_coarse_failure_onset_seconds": 5.0,
+    "vlm_failure_window_start_seconds": 4.5,
+    "vlm_failure_window_end_seconds": 5.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot positions its gripper over the tomato and closes its fingers, but fails to establish a secure grasp on the object, leaving the tomato behind on the table.",
+    "vlm_failure_reason_summary": "failed grasp on tomato; incomplete grip closure",
+    "vlm_recovery_action": "Re-align the gripper over the center of the tomato, lower to an appropriate height, and close the gripper securely before lifting.",
+    "vlm_recovery_action_summary": "realign over tomato; close gripper securely; pick up tomato",
+    "vlm_prevention_action": "Ensure the gripper is properly centered and at the correct height relative to the tomato before closing the fingers to guarantee a firm grasp.",
+    "vlm_prevention_action_summary": "center gripper over tomato; achieve proper depth before grasping",
+    "vlm_justification": "At around 5.0 seconds (visible in frame 00:05), the robot closes its gripper around the tomato but fails to securely hold it. It then lifts away empty and tries again at 00:07, failing a second time. This is a clear grasp_failure starting at the first failed attempt around 5.0s.",
+    "vlm_temporal_justification": "At second 02:12 (corresponding to frame 12 of the clip), the robot's end-effector begins closing its fingers around the tomato, but fails to properly align and secure it, leading directly to the failed grasp.",
+    "vlm_video_fps": 50.8,
+    "vlm_video_num_frames": 506,
+    "vlm_video_duration_seconds": 9.940944881889765,
+    "vlm_usage": {
+      "input_tokens": 21705,
+      "output_tokens": 409,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 5.0,\n \"failure_window_start_seconds\": 4.5,\n \"failure_window_end_seconds\": 5.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot positions its gripper over the tomato and closes its fingers, but fails to establish a secure grasp on the object, leaving the tomato behind on the table.\",\n \"failure_reason_summary\": \"failed grasp on tomato; incomplete grip closure\",\n \"recovery_action\": \"Re-align the gripper over the center of the tomato, lower to an appropriate height, and close the gripper securely before lifting.\",\n \"recovery_action_summary\": \"realign over tomato; close gripper securely; pick up tomato\",\n \"prevention_action\": \"Ensure the gripper is properly centered and at the correct height relative to the tomato before closing the fingers to guarantee a firm grasp.\",\n \"prevention_action_summary\": \"center gripper over tomato; achieve proper depth before grasping\",\n \"justification\": \"At around 5.0 seconds (visible in frame 00:05), the robot closes its gripper around the tomato but fails to securely hold it. It then lifts away empty and tries again at 00:07, failing a second time. This is a clear grasp_failure starting at the first failed attempt around 5.0s.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 114,
+    "failure_sim_step": 114,
+    "failure_policy_step": 19,
+    "failure_chunk_index": 0,
+    "first_post_failure_policy_step": 20,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-align the gripper over the center of the tomato, lower to an appropriate height, and close the gripper securely before lifting."
+  }
+}
\ No newline at end of file
diff --git a/tomato/ep029/labels.npz b/tomato/ep029/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..99f3682b4177421a464d78541261b4e749f4daeb
--- /dev/null
+++ b/tomato/ep029/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8bf1a269df8b03d1f29be74735213ba5278a9f653e5b2529ab118da336c6e94b
+size 3851
diff --git a/tomato/ep030/example.md b/tomato/ep030/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..e15d0c6a9a2918681c8fc7d445058a196bd8f3c3
--- /dev/null
+++ b/tomato/ep030/example.md
@@ -0,0 +1,34 @@
+# t003__ep030__tomato
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the tomato in the bin
+- **success:** False
+- **control steps:** 481  (9.5s at 50 Hz)
+- **policy inferences:** 83  (replan_steps=6)
+- **video:** `tomato/ep030/rollout.mp4`
+- **features:** `83` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching down toward tomato |
+| 3.0–6.0 | 152–303 | 25–50 | closing gripper on tomato |
+| 6.0–9.0 | 304–455 | 50–75 | moving empty arm toward bin |
+| 9.0–9.5 | 456–480 | 76–80 | hovering above bin empty |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 5.821782178217822 s (00:05.8)
+- **frame / control_step:** 294 (policy_step=49, chunk_index=0)
+- **coarse onset:** 5.0 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the tomato, but failed to establish a secure hold, causing the tomato to slip out from the fingers and roll across the tabletop.
+  Summary: failed grasp attempt; tomato slipping away
+- **recovery:** Re-align the gripper directly over the displaced tomato on the table, establish a secure grasp, and place it into the gray bin.
+  Summary: re-align over tomato; secure grasp; place in bin
+- **prevention:** Center the gripper precisely over the tomato before closing the fingers to ensure a stable grip prior to lifting.
+  Summary: center gripper over tomato; secure stable grasp
+- **justification:** At around 5.0 seconds (frame 00:05), the robot gripper closes around the tomato, but fails to grip it securely. By frame 00:06, as the arm rises, the tomato slips out and rolls away across the table, resulting in an unrecovered grasp failure.
diff --git a/tomato/ep030/labels.json b/tomato/ep030/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..9a4ed53b0129c68c59e52468aafd8fe5516004ce
--- /dev/null
+++ b/tomato/ep030/labels.json
@@ -0,0 +1,109 @@
+{
+  "rollout_id": "t003__ep030__tomato",
+  "scene_variant": "",
+  "task_id": 3,
+  "task_file": "",
+  "instruction": "Put the tomato in the bin",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T22:02:04+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00990099009901,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching down toward tomato",
+      "description": "reaching down toward tomato"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00990099009901,
+      "t_end_sec": 6.01980198019802,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "closing gripper on tomato",
+      "description": "closing gripper on tomato"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01980198019802,
+      "t_end_sec": 9.029702970297029,
+      "control_step_start": 304,
+      "control_step_end": 455,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "moving empty arm toward bin",
+      "description": "moving empty arm toward bin"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.029702970297029,
+      "t_end_sec": 9.524752475247524,
+      "control_step_start": 456,
+      "control_step_end": 480,
+      "policy_step_start": 76,
+      "policy_step_end": 80,
+      "phrase": "hovering above bin empty",
+      "description": "hovering above bin empty"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 5.821782178217822,
+    "vlm_failure_onset_timestamp": "00:05.8",
+    "vlm_failure_onset_frame": 294,
+    "vlm_failure_onset_step": 294,
+    "vlm_coarse_failure_onset_seconds": 5.0,
+    "vlm_failure_window_start_seconds": 4.5,
+    "vlm_failure_window_end_seconds": 5.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the tomato, but failed to establish a secure hold, causing the tomato to slip out from the fingers and roll across the tabletop.",
+    "vlm_failure_reason_summary": "failed grasp attempt; tomato slipping away",
+    "vlm_recovery_action": "Re-align the gripper directly over the displaced tomato on the table, establish a secure grasp, and place it into the gray bin.",
+    "vlm_recovery_action_summary": "re-align over tomato; secure grasp; place in bin",
+    "vlm_prevention_action": "Center the gripper precisely over the tomato before closing the fingers to ensure a stable grip prior to lifting.",
+    "vlm_prevention_action_summary": "center gripper over tomato; secure stable grasp",
+    "vlm_justification": "At around 5.0 seconds (frame 00:05), the robot gripper closes around the tomato, but fails to grip it securely. By frame 00:06, as the arm rises, the tomato slips out and rolls away across the table, resulting in an unrecovered grasp failure.",
+    "vlm_temporal_justification": "At timestamp 03:14 (194 seconds into the clip), the robot gripper closes onto the tomato but fails to secure it, causing the tomato to slip out from between the fingers and roll across the table.",
+    "vlm_video_fps": 50.5,
+    "vlm_video_num_frames": 481,
+    "vlm_video_duration_seconds": 9.504950495049505,
+    "vlm_usage": {
+      "input_tokens": 21704,
+      "output_tokens": 388,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 5.0,\n \"failure_window_start_seconds\": 4.5,\n \"failure_window_end_seconds\": 5.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the tomato, but failed to establish a secure hold, causing the tomato to slip out from the fingers and roll across the tabletop.\",\n \"failure_reason_summary\": \"failed grasp attempt; tomato slipping away\",\n \"recovery_action\": \"Re-align the gripper directly over the displaced tomato on the table, establish a secure grasp, and place it into the gray bin.\",\n \"recovery_action_summary\": \"re-align over tomato; secure grasp; place in bin\",\n \"prevention_action\": \"Center the gripper precisely over the tomato before closing the fingers to ensure a stable grip prior to lifting.\",\n \"prevention_action_summary\": \"center gripper over tomato; secure stable grasp\",\n \"justification\": \"At around 5.0 seconds (frame 00:05), the robot gripper closes around the tomato, but fails to grip it securely. By frame 00:06, as the arm rises, the tomato slips out and rolls away across the table, resulting in an unrecovered grasp failure.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 294,
+    "failure_sim_step": 294,
+    "failure_policy_step": 49,
+    "failure_chunk_index": 0,
+    "first_post_failure_policy_step": 50,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-align the gripper directly over the displaced tomato on the table, establish a secure grasp, and place it into the gray bin."
+  }
+}
\ No newline at end of file
diff --git a/tomato/ep030/labels.npz b/tomato/ep030/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..13458755cd5ca8522422c8b5e0381aef0c7b56e4
--- /dev/null
+++ b/tomato/ep030/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b9f056a6a172cfeb1ce6a8616af5761f025110561c9ba4ad10db883aee6b3285
+size 3794
diff --git a/tomato/ep031/example.md b/tomato/ep031/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..5be8e49e971dc45c14584ca2e196e69785c07539
--- /dev/null
+++ b/tomato/ep031/example.md
@@ -0,0 +1,36 @@
+# t003__ep031__tomato
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the tomato in the bin
+- **success:** False
+- **control steps:** 877  (17.5s at 50 Hz)
+- **policy inferences:** 150  (replan_steps=6)
+- **video:** `tomato/ep031/rollout.mp4`
+- **features:** `150` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–149 | 0–24 | reaching toward tomato |
+| 3.0–6.0 | 150–299 | 25–49 | attempting grasp on tomato |
+| 6.0–9.0 | 300–449 | 50–74 | pressing down on tomato |
+| 9.0–12.0 | 450–599 | 75–99 | repeatedly pushing tomato |
+| 12.0–15.0 | 600–749 | 100–124 | poking tomato with gripper |
+| 15.0–17.5 | 750–876 | 125–146 | hovering over tomato |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 2.5948103792415167 s (00:02.6)
+- **frame / control_step:** 130 (policy_step=21, chunk_index=4)
+- **coarse onset:** 3.5 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the tomato on the table but failed to secure its grip, repeatedly pressing down on the object without picking it up.
+  Summary: failed grasp attempt; unable to secure tomato
+- **recovery:** Re-open the gripper completely, position it centered over the tomato, descend to the proper grasp height, and close the fingers securely around the tomato before transporting it to the bin.
+  Summary: re-open gripper; align with tomato; grasp securely
+- **prevention:** Ensure accurate centering and sufficient grasp depth over the tomato before closing the fingers to ensure a firm grasp.
+  Summary: align gripper centered over tomato; close fingers firmly
+- **justification:** At approximately 3.5 seconds, the robot reaches down to grasp the tomato but fails to secure it between the gripper fingers. It repeatedly attempts to press down on the tomato without successfully gripping or lifting it for the remainder of the rollout.
diff --git a/tomato/ep031/labels.json b/tomato/ep031/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..6dcd4942d63a4bc0cc5710583178cef1f60b0325
--- /dev/null
+++ b/tomato/ep031/labels.json
@@ -0,0 +1,131 @@
+{
+  "rollout_id": "t003__ep031__tomato",
+  "scene_variant": "",
+  "task_id": 3,
+  "task_file": "",
+  "instruction": "Put the tomato in the bin",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T22:02:36+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9940119760479043,
+      "control_step_start": 0,
+      "control_step_end": 149,
+      "policy_step_start": 0,
+      "policy_step_end": 24,
+      "phrase": "reaching toward tomato",
+      "description": "reaching toward tomato"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9940119760479043,
+      "t_end_sec": 5.9880239520958085,
+      "control_step_start": 150,
+      "control_step_end": 299,
+      "policy_step_start": 25,
+      "policy_step_end": 49,
+      "phrase": "attempting grasp on tomato",
+      "description": "attempting grasp on tomato"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.9880239520958085,
+      "t_end_sec": 8.982035928143713,
+      "control_step_start": 300,
+      "control_step_end": 449,
+      "policy_step_start": 50,
+      "policy_step_end": 74,
+      "phrase": "pressing down on tomato",
+      "description": "pressing down on tomato"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.982035928143713,
+      "t_end_sec": 11.976047904191617,
+      "control_step_start": 450,
+      "control_step_end": 599,
+      "policy_step_start": 75,
+      "policy_step_end": 99,
+      "phrase": "repeatedly pushing tomato",
+      "description": "repeatedly pushing tomato"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 11.976047904191617,
+      "t_end_sec": 14.970059880239521,
+      "control_step_start": 600,
+      "control_step_end": 749,
+      "policy_step_start": 100,
+      "policy_step_end": 124,
+      "phrase": "poking tomato with gripper",
+      "description": "poking tomato with gripper"
+    },
+    {
+      "segment_index": 5,
+      "t_start_sec": 14.970059880239521,
+      "t_end_sec": 17.50499001996008,
+      "control_step_start": 750,
+      "control_step_end": 876,
+      "policy_step_start": 125,
+      "policy_step_end": 146,
+      "phrase": "hovering over tomato",
+      "description": "hovering over tomato"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 2.5948103792415167,
+    "vlm_failure_onset_timestamp": "00:02.6",
+    "vlm_failure_onset_frame": 130,
+    "vlm_failure_onset_step": 130,
+    "vlm_coarse_failure_onset_seconds": 3.5,
+    "vlm_failure_window_start_seconds": 2.5,
+    "vlm_failure_window_end_seconds": 5.0,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the tomato on the table but failed to secure its grip, repeatedly pressing down on the object without picking it up.",
+    "vlm_failure_reason_summary": "failed grasp attempt; unable to secure tomato",
+    "vlm_recovery_action": "Re-open the gripper completely, position it centered over the tomato, descend to the proper grasp height, and close the fingers securely around the tomato before transporting it to the bin.",
+    "vlm_recovery_action_summary": "re-open gripper; align with tomato; grasp securely",
+    "vlm_prevention_action": "Ensure accurate centering and sufficient grasp depth over the tomato before closing the fingers to ensure a firm grasp.",
+    "vlm_prevention_action_summary": "align gripper centered over tomato; close fingers firmly",
+    "vlm_justification": "At approximately 3.5 seconds, the robot reaches down to grasp the tomato but fails to secure it between the gripper fingers. It repeatedly attempts to press down on the tomato without successfully gripping or lifting it for the remainder of the rollout.",
+    "vlm_temporal_justification": "At timestamp 01:45, the robot arm's gripper closes on the tomato, but fails to establish a proper grip, resulting in a grasp failure as it merely presses against the object without securing it.",
+    "vlm_video_fps": 50.1,
+    "vlm_video_num_frames": 877,
+    "vlm_video_duration_seconds": 17.48502994011976,
+    "vlm_usage": {
+      "input_tokens": 21953,
+      "output_tokens": 377,
+      "cached_tokens": 1225
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 3.5,\n \"failure_window_start_seconds\": 2.5,\n \"failure_window_end_seconds\": 5.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the tomato on the table but failed to secure its grip, repeatedly pressing down on the object without picking it up.\",\n \"failure_reason_summary\": \"failed grasp attempt; unable to secure tomato\",\n \"recovery_action\": \"Re-open the gripper completely, position it centered over the tomato, descend to the proper grasp height, and close the fingers securely around the tomato before transporting it to the bin.\",\n \"recovery_action_summary\": \"re-open gripper; align with tomato; grasp securely\",\n \"prevention_action\": \"Ensure accurate centering and sufficient grasp depth over the tomato before closing the fingers to ensure a firm grasp.\",\n \"prevention_action_summary\": \"align gripper centered over tomato; close fingers firmly\",\n \"justification\": \"At approximately 3.5 seconds, the robot reaches down to grasp the tomato but fails to secure it between the gripper fingers. It repeatedly attempts to press down on the tomato without successfully gripping or lifting it for the remainder of the rollout.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 130,
+    "failure_sim_step": 130,
+    "failure_policy_step": 21,
+    "failure_chunk_index": 4,
+    "first_post_failure_policy_step": 22,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-open the gripper completely, position it centered over the tomato, descend to the proper grasp height, and close the fingers securely around the tomato before transporting it to the bin."
+  }
+}
\ No newline at end of file
diff --git a/tomato/ep031/labels.npz b/tomato/ep031/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..717068bbb18db4075dffcf15224dee8bfc3688d8
--- /dev/null
+++ b/tomato/ep031/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1f20981a95f363ccfb93cb6dcf04c9ba7f7de506165f9d76e37b783ac13a4c64
+size 3941
diff --git a/tomato/ep032/example.md b/tomato/ep032/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..6ee941847c11b62e74fd665bdd7796a2497633f1
--- /dev/null
+++ b/tomato/ep032/example.md
@@ -0,0 +1,34 @@
+# t003__ep032__tomato
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the tomato in the bin
+- **success:** False
+- **control steps:** 560  (11.0s at 51 Hz)
+- **policy inferences:** 95  (replan_steps=6)
+- **video:** `tomato/ep032/rollout.mp4`
+- **features:** `95` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | descending arm toward tomato |
+| 3.0–6.0 | 152–303 | 25–50 | closing gripper around tomato |
+| 6.0–9.0 | 304–455 | 50–75 | slipping off tomato without grasping |
+| 9.0–11.0 | 456–559 | 76–93 | hovering over tomato without progress |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 5.936883629191321 s (00:05.9)
+- **frame / control_step:** 301 (policy_step=50, chunk_index=1)
+- **coarse onset:** 4.5 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the tomato on the table but failed to establish a secure grip, leaving the tomato on the tabletop without lifting it.
+  Summary: failed grasp attempt; object not secured
+- **recovery:** Re-open the gripper, re-center over the tomato, and grasp it firmly before lifting it into the bin.
+  Summary: re-align gripper; regrasp tomato; lift to bin
+- **prevention:** Align the end-effector accurately over the center of the tomato and close the fingers securely around it prior to initiating the lift.
+  Summary: center gripper on tomato; secure grasp before lifting
+- **justification:** The robot reaches the tomato around 4.0 seconds and attempts to close its gripper around 4.5 seconds. However, it fails to acquire the tomato and stays near the surface for the remainder of the video without picking it up.
diff --git a/tomato/ep032/labels.json b/tomato/ep032/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..3d0722127fb34cc7ffcee25780adf47230e0d918
--- /dev/null
+++ b/tomato/ep032/labels.json
@@ -0,0 +1,109 @@
+{
+  "rollout_id": "t003__ep032__tomato",
+  "scene_variant": "",
+  "task_id": 3,
+  "task_file": "",
+  "instruction": "Put the tomato in the bin",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T22:03:11+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9980276134122286,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "descending arm toward tomato",
+      "description": "descending arm toward tomato"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9980276134122286,
+      "t_end_sec": 5.996055226824457,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "closing gripper around tomato",
+      "description": "closing gripper around tomato"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.996055226824457,
+      "t_end_sec": 8.994082840236686,
+      "control_step_start": 304,
+      "control_step_end": 455,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "slipping off tomato without grasping",
+      "description": "slipping off tomato without grasping"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.994082840236686,
+      "t_end_sec": 11.045364891518737,
+      "control_step_start": 456,
+      "control_step_end": 559,
+      "policy_step_start": 76,
+      "policy_step_end": 93,
+      "phrase": "hovering over tomato without progress",
+      "description": "hovering over tomato without progress"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 5.936883629191321,
+    "vlm_failure_onset_timestamp": "00:05.9",
+    "vlm_failure_onset_frame": 301,
+    "vlm_failure_onset_step": 301,
+    "vlm_coarse_failure_onset_seconds": 4.5,
+    "vlm_failure_window_start_seconds": 3.5,
+    "vlm_failure_window_end_seconds": 5.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the tomato on the table but failed to establish a secure grip, leaving the tomato on the tabletop without lifting it.",
+    "vlm_failure_reason_summary": "failed grasp attempt; object not secured",
+    "vlm_recovery_action": "Re-open the gripper, re-center over the tomato, and grasp it firmly before lifting it into the bin.",
+    "vlm_recovery_action_summary": "re-align gripper; regrasp tomato; lift to bin",
+    "vlm_prevention_action": "Align the end-effector accurately over the center of the tomato and close the fingers securely around it prior to initiating the lift.",
+    "vlm_prevention_action_summary": "center gripper on tomato; secure grasp before lifting",
+    "vlm_justification": "The robot reaches the tomato around 4.0 seconds and attempts to close its gripper around 4.5 seconds. However, it fails to acquire the tomato and stays near the surface for the remainder of the video without picking it up.",
+    "vlm_temporal_justification": "At timestamp 03:45, the robot's gripper fingers close completely without grasping the tomato, leaving the tomato resting on the table surface and making the grasp failure visibly apparent.",
+    "vlm_video_fps": 50.7,
+    "vlm_video_num_frames": 560,
+    "vlm_video_duration_seconds": 11.025641025641026,
+    "vlm_usage": {
+      "input_tokens": 21763,
+      "output_tokens": 364,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 4.5,\n \"failure_window_start_seconds\": 3.5,\n \"failure_window_end_seconds\": 5.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the tomato on the table but failed to establish a secure grip, leaving the tomato on the tabletop without lifting it.\",\n \"failure_reason_summary\": \"failed grasp attempt; object not secured\",\n \"recovery_action\": \"Re-open the gripper, re-center over the tomato, and grasp it firmly before lifting it into the bin.\",\n \"recovery_action_summary\": \"re-align gripper; regrasp tomato; lift to bin\",\n \"prevention_action\": \"Align the end-effector accurately over the center of the tomato and close the fingers securely around it prior to initiating the lift.\",\n \"prevention_action_summary\": \"center gripper on tomato; secure grasp before lifting\",\n \"justification\": \"The robot reaches the tomato around 4.0 seconds and attempts to close its gripper around 4.5 seconds. However, it fails to acquire the tomato and stays near the surface for the remainder of the video without picking it up.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 301,
+    "failure_sim_step": 301,
+    "failure_policy_step": 50,
+    "failure_chunk_index": 1,
+    "first_post_failure_policy_step": 51,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-open the gripper, re-center over the tomato, and grasp it firmly before lifting it into the bin."
+  }
+}
\ No newline at end of file
diff --git a/tomato/ep032/labels.npz b/tomato/ep032/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..aa72107648d3c8775c9a7bbad610ab02b911e163
--- /dev/null
+++ b/tomato/ep032/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:91b2819425d1300197d5b5fb1d2dcbf7dfe13efaa991f2d184699452e4c298ba
+size 3807
diff --git a/tomato/ep033/example.md b/tomato/ep033/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..bc9dc6689b2b5f1b623cd0ff932f0ea63ce7a0e7
--- /dev/null
+++ b/tomato/ep033/example.md
@@ -0,0 +1,37 @@
+# t003__ep033__tomato
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the tomato in the bin
+- **success:** False
+- **control steps:** 1010  (20.2s at 50 Hz)
+- **policy inferences:** 174  (replan_steps=6)
+- **video:** `tomato/ep033/rollout.mp4`
+- **features:** `174` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–149 | 0–24 | lowering gripper into bin |
+| 3.0–6.0 | 150–299 | 25–49 | hovering inside bin |
+| 6.0–9.0 | 300–449 | 50–74 | idle inside bin |
+| 9.0–12.0 | 450–599 | 75–99 | hovering inside bin |
+| 12.0–15.0 | 600–749 | 100–124 | idle inside bin |
+| 15.0–18.0 | 750–899 | 125–149 | hovering above bin |
+| 18.0–20.2 | 900–1009 | 150–168 | idle near bin |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 1.2974051896207583 s (00:01.3)
+- **frame / control_step:** 65 (policy_step=10, chunk_index=5)
+- **coarse onset:** 1.8 s (refined=True)
+- **mode:** wrong_object
+- **reason:** The robot moved towards and attempted to interact with the empty bin instead of targeting and picking up the tomato.
+  Summary: attempted interaction with bin instead of tomato
+- **recovery:** Move the arm away from the bin, navigate to the tomato, grasp it securely, and then carry it to the bin.
+  Summary: move away from bin; grasp tomato; place in bin
+- **prevention:** Target the red tomato first and grasp it cleanly before moving toward the bin.
+  Summary: grasp tomato first; avoid bin initially
+- **justification:** At approximately 1.8 seconds into the video, the robot arm bypasses the tomato located near the front edge of the table and lowers its gripper into the empty gray bin, closing its fingers on nothing. It remains stuck at the bin for the rest of the episode.
diff --git a/tomato/ep033/labels.json b/tomato/ep033/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..6c8a50c00d89a8cf4a398b93b37a7af0e9c3887b
--- /dev/null
+++ b/tomato/ep033/labels.json
@@ -0,0 +1,142 @@
+{
+  "rollout_id": "t003__ep033__tomato",
+  "scene_variant": "",
+  "task_id": 3,
+  "task_file": "",
+  "instruction": "Put the tomato in the bin",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T22:03:43+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9940119760479043,
+      "control_step_start": 0,
+      "control_step_end": 149,
+      "policy_step_start": 0,
+      "policy_step_end": 24,
+      "phrase": "lowering gripper into bin",
+      "description": "lowering gripper into bin"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9940119760479043,
+      "t_end_sec": 5.9880239520958085,
+      "control_step_start": 150,
+      "control_step_end": 299,
+      "policy_step_start": 25,
+      "policy_step_end": 49,
+      "phrase": "hovering inside bin",
+      "description": "hovering inside bin"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.9880239520958085,
+      "t_end_sec": 8.982035928143713,
+      "control_step_start": 300,
+      "control_step_end": 449,
+      "policy_step_start": 50,
+      "policy_step_end": 74,
+      "phrase": "idle inside bin",
+      "description": "idle inside bin"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.982035928143713,
+      "t_end_sec": 11.976047904191617,
+      "control_step_start": 450,
+      "control_step_end": 599,
+      "policy_step_start": 75,
+      "policy_step_end": 99,
+      "phrase": "hovering inside bin",
+      "description": "hovering inside bin"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 11.976047904191617,
+      "t_end_sec": 14.970059880239521,
+      "control_step_start": 600,
+      "control_step_end": 749,
+      "policy_step_start": 100,
+      "policy_step_end": 124,
+      "phrase": "idle inside bin",
+      "description": "idle inside bin"
+    },
+    {
+      "segment_index": 5,
+      "t_start_sec": 14.970059880239521,
+      "t_end_sec": 17.964071856287426,
+      "control_step_start": 750,
+      "control_step_end": 899,
+      "policy_step_start": 125,
+      "policy_step_end": 149,
+      "phrase": "hovering above bin",
+      "description": "hovering above bin"
+    },
+    {
+      "segment_index": 6,
+      "t_start_sec": 17.964071856287426,
+      "t_end_sec": 20.159680638722556,
+      "control_step_start": 900,
+      "control_step_end": 1009,
+      "policy_step_start": 150,
+      "policy_step_end": 168,
+      "phrase": "idle near bin",
+      "description": "idle near bin"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "wrong_object",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 1.2974051896207583,
+    "vlm_failure_onset_timestamp": "00:01.3",
+    "vlm_failure_onset_frame": 65,
+    "vlm_failure_onset_step": 65,
+    "vlm_coarse_failure_onset_seconds": 1.8,
+    "vlm_failure_window_start_seconds": 1.5,
+    "vlm_failure_window_end_seconds": 2.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot moved towards and attempted to interact with the empty bin instead of targeting and picking up the tomato.",
+    "vlm_failure_reason_summary": "attempted interaction with bin instead of tomato",
+    "vlm_recovery_action": "Move the arm away from the bin, navigate to the tomato, grasp it securely, and then carry it to the bin.",
+    "vlm_recovery_action_summary": "move away from bin; grasp tomato; place in bin",
+    "vlm_prevention_action": "Target the red tomato first and grasp it cleanly before moving toward the bin.",
+    "vlm_prevention_action_summary": "grasp tomato first; avoid bin initially",
+    "vlm_justification": "At approximately 1.8 seconds into the video, the robot arm bypasses the tomato located near the front edge of the table and lowers its gripper into the empty gray bin, closing its fingers on nothing. It remains stuck at the bin for the rest of the episode.",
+    "vlm_temporal_justification": "At second 65 (frame 65), the robot arm initiates a clear trajectory downwards toward the grey bin rather than moving forward toward the tomato located in the foreground.",
+    "vlm_video_fps": 50.1,
+    "vlm_video_num_frames": 1010,
+    "vlm_video_duration_seconds": 20.139720558882235,
+    "vlm_usage": {
+      "input_tokens": 18289,
+      "output_tokens": 349,
+      "cached_tokens": 1361
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"wrong_object\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 1.8,\n \"failure_window_start_seconds\": 1.5,\n \"failure_window_end_seconds\": 2.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot moved towards and attempted to interact with the empty bin instead of targeting and picking up the tomato.\",\n \"failure_reason_summary\": \"attempted interaction with bin instead of tomato\",\n \"recovery_action\": \"Move the arm away from the bin, navigate to the tomato, grasp it securely, and then carry it to the bin.\",\n \"recovery_action_summary\": \"move away from bin; grasp tomato; place in bin\",\n \"prevention_action\": \"Target the red tomato first and grasp it cleanly before moving toward the bin.\",\n \"prevention_action_summary\": \"grasp tomato first; avoid bin initially\",\n \"justification\": \"At approximately 1.8 seconds into the video, the robot arm bypasses the tomato located near the front edge of the table and lowers its gripper into the empty gray bin, closing its fingers on nothing. It remains stuck at the bin for the rest of the episode.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 65,
+    "failure_sim_step": 65,
+    "failure_policy_step": 10,
+    "failure_chunk_index": 5,
+    "first_post_failure_policy_step": 11,
+    "failure_type": "wrong_object",
+    "correction_action": "Move the arm away from the bin, navigate to the tomato, grasp it securely, and then carry it to the bin."
+  }
+}
\ No newline at end of file
diff --git a/tomato/ep033/labels.npz b/tomato/ep033/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..ab68e897858b7b7bfd37e800ba79639d80e3144b
--- /dev/null
+++ b/tomato/ep033/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d80c024139178fb63386d7ef7b3ec76115cffbb29d11a7384982d2981d8103e1
+size 3752
diff --git a/tomato/ep034/example.md b/tomato/ep034/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..f8a7bae79c6757133be48e635bf5bbc9a78202f3
--- /dev/null
+++ b/tomato/ep034/example.md
@@ -0,0 +1,36 @@
+# t003__ep034__tomato
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the tomato in the bin
+- **success:** False
+- **control steps:** 753  (15.0s at 50 Hz)
+- **policy inferences:** 131  (replan_steps=6)
+- **video:** `tomato/ep034/rollout.mp4`
+- **features:** `131` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–149 | 0–24 | lowering gripper into empty bin |
+| 3.0–6.0 | 150–299 | 25–49 | hovering inside empty bin |
+| 6.0–9.0 | 300–449 | 50–74 | remaining motionless in bin |
+| 9.0–12.0 | 450–599 | 75–99 | hovering inside empty bin |
+| 12.0–15.0 | 600–749 | 100–124 | remaining motionless in bin |
+| 15.0–15.0 | 750–752 | 125–125 | idle inside empty bin |
+
+## Failure
+
+- **yes**
+- **onset type:** operator_intervention
+- **onset:** 0.8383233532934131 s (00:00.8)
+- **frame / control_step:** 42 (policy_step=7, chunk_index=0)
+- **coarse onset:** 2.0 s (refined=True)
+- **mode:** stuck_or_no_progress
+- **reason:** The robot arm moved directly down into the empty bin instead of approaching the tomato, and remained hovering inside the bin for the rest of the rollout without attempting to grasp the target object.
+  Summary: robot moved into empty bin; stuck hovering without grasping tomato
+- **recovery:** Retract the end effector from the bin, reposition the arm over the tomato, grasp the tomato, and then move it into the bin.
+  Summary: retract arm from bin; grasp tomato; place tomato into bin
+- **prevention:** Plan a trajectory to approach and grasp the tomato on the table before navigating toward the target bin.
+  Summary: approach tomato first; grasp tomato securely; transport to bin
+- **justification:** Around 2.0 seconds into the rollout, the robot lowers its gripper into the empty bin without grasping the tomato on the table, remaining stuck in that position until the end of the video. An operator would need to intervene around 2.0s as no task progress is being made.
diff --git a/tomato/ep034/labels.json b/tomato/ep034/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..05fc25d8e32e881ef41f753a641b554533fa4f32
--- /dev/null
+++ b/tomato/ep034/labels.json
@@ -0,0 +1,131 @@
+{
+  "rollout_id": "t003__ep034__tomato",
+  "scene_variant": "",
+  "task_id": 3,
+  "task_file": "",
+  "instruction": "Put the tomato in the bin",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T22:04:15+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9940119760479043,
+      "control_step_start": 0,
+      "control_step_end": 149,
+      "policy_step_start": 0,
+      "policy_step_end": 24,
+      "phrase": "lowering gripper into empty bin",
+      "description": "lowering gripper into empty bin"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9940119760479043,
+      "t_end_sec": 5.9880239520958085,
+      "control_step_start": 150,
+      "control_step_end": 299,
+      "policy_step_start": 25,
+      "policy_step_end": 49,
+      "phrase": "hovering inside empty bin",
+      "description": "hovering inside empty bin"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.9880239520958085,
+      "t_end_sec": 8.982035928143713,
+      "control_step_start": 300,
+      "control_step_end": 449,
+      "policy_step_start": 50,
+      "policy_step_end": 74,
+      "phrase": "remaining motionless in bin",
+      "description": "remaining motionless in bin"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.982035928143713,
+      "t_end_sec": 11.976047904191617,
+      "control_step_start": 450,
+      "control_step_end": 599,
+      "policy_step_start": 75,
+      "policy_step_end": 99,
+      "phrase": "hovering inside empty bin",
+      "description": "hovering inside empty bin"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 11.976047904191617,
+      "t_end_sec": 14.970059880239521,
+      "control_step_start": 600,
+      "control_step_end": 749,
+      "policy_step_start": 100,
+      "policy_step_end": 124,
+      "phrase": "remaining motionless in bin",
+      "description": "remaining motionless in bin"
+    },
+    {
+      "segment_index": 5,
+      "t_start_sec": 14.970059880239521,
+      "t_end_sec": 15.029940119760479,
+      "control_step_start": 750,
+      "control_step_end": 752,
+      "policy_step_start": 125,
+      "policy_step_end": 125,
+      "phrase": "idle inside empty bin",
+      "description": "idle inside empty bin"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "stuck_or_no_progress",
+    "vlm_failure_onset_type": "operator_intervention",
+    "vlm_failure_onset_seconds": 0.8383233532934131,
+    "vlm_failure_onset_timestamp": "00:00.8",
+    "vlm_failure_onset_frame": 42,
+    "vlm_failure_onset_step": 42,
+    "vlm_coarse_failure_onset_seconds": 2.0,
+    "vlm_failure_window_start_seconds": 1.5,
+    "vlm_failure_window_end_seconds": 3.0,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot arm moved directly down into the empty bin instead of approaching the tomato, and remained hovering inside the bin for the rest of the rollout without attempting to grasp the target object.",
+    "vlm_failure_reason_summary": "robot moved into empty bin; stuck hovering without grasping tomato",
+    "vlm_recovery_action": "Retract the end effector from the bin, reposition the arm over the tomato, grasp the tomato, and then move it into the bin.",
+    "vlm_recovery_action_summary": "retract arm from bin; grasp tomato; place tomato into bin",
+    "vlm_prevention_action": "Plan a trajectory to approach and grasp the tomato on the table before navigating toward the target bin.",
+    "vlm_prevention_action_summary": "approach tomato first; grasp tomato securely; transport to bin",
+    "vlm_justification": "Around 2.0 seconds into the rollout, the robot lowers its gripper into the empty bin without grasping the tomato on the table, remaining stuck in that position until the end of the video. An operator would need to intervene around 2.0s as no task progress is being made.",
+    "vlm_temporal_justification": "At second 42 (frame 42), the robot arm initiates its movement, proceeding directly down and towards the empty container on the left rather than moving forward towards the target tomato.",
+    "vlm_video_fps": 50.1,
+    "vlm_video_num_frames": 753,
+    "vlm_video_duration_seconds": 15.009980039920158,
+    "vlm_usage": {
+      "input_tokens": 18621,
+      "output_tokens": 390,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"stuck_or_no_progress\",\n \"onset_type\": \"operator_intervention\",\n \"failure_onset_seconds\": 2.0,\n \"failure_window_start_seconds\": 1.5,\n \"failure_window_end_seconds\": 3.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot arm moved directly down into the empty bin instead of approaching the tomato, and remained hovering inside the bin for the rest of the rollout without attempting to grasp the target object.\",\n \"failure_reason_summary\": \"robot moved into empty bin; stuck hovering without grasping tomato\",\n \"recovery_action\": \"Retract the end effector from the bin, reposition the arm over the tomato, grasp the tomato, and then move it into the bin.\",\n \"recovery_action_summary\": \"retract arm from bin; grasp tomato; place tomato into bin\",\n \"prevention_action\": \"Plan a trajectory to approach and grasp the tomato on the table before navigating toward the target bin.\",\n \"prevention_action_summary\": \"approach tomato first; grasp tomato securely; transport to bin\",\n \"justification\": \"Around 2.0 seconds into the rollout, the robot lowers its gripper into the empty bin without grasping the tomato on the table, remaining stuck in that position until the end of the video. An operator would need to intervene around 2.0s as no task progress is being made.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 42,
+    "failure_sim_step": 42,
+    "failure_policy_step": 7,
+    "failure_chunk_index": 0,
+    "first_post_failure_policy_step": 8,
+    "failure_type": "stuck_or_no_progress",
+    "correction_action": "Retract the end effector from the bin, reposition the arm over the tomato, grasp the tomato, and then move it into the bin."
+  }
+}
\ No newline at end of file
diff --git a/tomato/ep034/labels.npz b/tomato/ep034/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..e95d644759868d1c8926efc93477c4d625be1690
--- /dev/null
+++ b/tomato/ep034/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8d77a3ca83655f778fed57fa1b6cc99418abd2ef5b498feb878858b062fbbb08
+size 3870
diff --git a/tomato/ep035/example.md b/tomato/ep035/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..7954fe04da30b3f39ce0594afa03d23d876e73b9
--- /dev/null
+++ b/tomato/ep035/example.md
@@ -0,0 +1,36 @@
+# t003__ep035__tomato
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the tomato in the bin
+- **success:** False
+- **control steps:** 807  (16.0s at 50 Hz)
+- **policy inferences:** 139  (replan_steps=6)
+- **video:** `tomato/ep035/rollout.mp4`
+- **features:** `139` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | descending towards tomato |
+| 3.0–6.0 | 151–301 | 25–50 | attempting grasp on tomato |
+| 6.0–9.0 | 302–452 | 50–75 | stuck pressing on tomato |
+| 9.0–12.0 | 453–603 | 75–100 | idle above tomato |
+| 12.0–15.0 | 604–754 | 100–125 | remaining stationary near tomato |
+| 15.0–16.0 | 755–806 | 125–134 | idle near tomato |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 4.027777777777778 s (00:04.0)
+- **frame / control_step:** 203 (policy_step=33, chunk_index=5)
+- **coarse onset:** 3.3 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot lowered its end effector onto the tomato but failed to secure a firm grip, causing the gripper to pinch incorrectly and remain stuck on the table without lifting the tomato.
+  Summary: failed to grasp tomato; gripper missed grasp
+- **recovery:** Open the gripper, reposition the end effector directly above the tomato, close the gripper firmly around it, and proceed to carry it to the bin.
+  Summary: reposition gripper; re-grasp tomato; place in bin
+- **prevention:** Position the gripper precisely over the tomato and ensure a stable grasp is established before initiating any lifting motion.
+  Summary: align gripper over tomato; grasp securely; move to bin
+- **justification:** At around 3.3 seconds (between 00:03 and 00:04), the robot attempts to close its gripper on the tomato but fails to acquire a proper grip. For the remainder of the video, it remains stuck at the tomato position without lifting it.
diff --git a/tomato/ep035/labels.json b/tomato/ep035/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..cea99ac049cc041c9d0f8114e0a6fbb6fa105c39
--- /dev/null
+++ b/tomato/ep035/labels.json
@@ -0,0 +1,131 @@
+{
+  "rollout_id": "t003__ep035__tomato",
+  "scene_variant": "",
+  "task_id": 3,
+  "task_file": "",
+  "instruction": "Put the tomato in the bin",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T22:04:52+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.996031746031746,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "descending towards tomato",
+      "description": "descending towards tomato"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.996031746031746,
+      "t_end_sec": 5.992063492063492,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "attempting grasp on tomato",
+      "description": "attempting grasp on tomato"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.992063492063492,
+      "t_end_sec": 8.988095238095239,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "stuck pressing on tomato",
+      "description": "stuck pressing on tomato"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.988095238095239,
+      "t_end_sec": 11.984126984126984,
+      "control_step_start": 453,
+      "control_step_end": 603,
+      "policy_step_start": 75,
+      "policy_step_end": 100,
+      "phrase": "idle above tomato",
+      "description": "idle above tomato"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 11.984126984126984,
+      "t_end_sec": 14.980158730158731,
+      "control_step_start": 604,
+      "control_step_end": 754,
+      "policy_step_start": 100,
+      "policy_step_end": 125,
+      "phrase": "remaining stationary near tomato",
+      "description": "remaining stationary near tomato"
+    },
+    {
+      "segment_index": 5,
+      "t_start_sec": 14.980158730158731,
+      "t_end_sec": 16.011904761904763,
+      "control_step_start": 755,
+      "control_step_end": 806,
+      "policy_step_start": 125,
+      "policy_step_end": 134,
+      "phrase": "idle near tomato",
+      "description": "idle near tomato"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 4.027777777777778,
+    "vlm_failure_onset_timestamp": "00:04.0",
+    "vlm_failure_onset_frame": 203,
+    "vlm_failure_onset_step": 203,
+    "vlm_coarse_failure_onset_seconds": 3.3,
+    "vlm_failure_window_start_seconds": 3.0,
+    "vlm_failure_window_end_seconds": 4.0,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot lowered its end effector onto the tomato but failed to secure a firm grip, causing the gripper to pinch incorrectly and remain stuck on the table without lifting the tomato.",
+    "vlm_failure_reason_summary": "failed to grasp tomato; gripper missed grasp",
+    "vlm_recovery_action": "Open the gripper, reposition the end effector directly above the tomato, close the gripper firmly around it, and proceed to carry it to the bin.",
+    "vlm_recovery_action_summary": "reposition gripper; re-grasp tomato; place in bin",
+    "vlm_prevention_action": "Position the gripper precisely over the tomato and ensure a stable grasp is established before initiating any lifting motion.",
+    "vlm_prevention_action_summary": "align gripper over tomato; grasp securely; move to bin",
+    "vlm_justification": "At around 3.3 seconds (between 00:03 and 00:04), the robot attempts to close its gripper on the tomato but fails to acquire a proper grip. For the remainder of the video, it remains stuck at the tomato position without lifting it.",
+    "vlm_temporal_justification": "At clip second 03:08 (second 188), the robot's gripper finishes closing around the tomato but fails to secure a proper grip under the object, resulting in an incomplete and failed grasp.",
+    "vlm_video_fps": 50.4,
+    "vlm_video_num_frames": 807,
+    "vlm_video_duration_seconds": 15.992063492063492,
+    "vlm_usage": {
+      "input_tokens": 21959,
+      "output_tokens": 391,
+      "cached_tokens": 1089
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 3.3,\n \"failure_window_start_seconds\": 3.0,\n \"failure_window_end_seconds\": 4.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot lowered its end effector onto the tomato but failed to secure a firm grip, causing the gripper to pinch incorrectly and remain stuck on the table without lifting the tomato.\",\n \"failure_reason_summary\": \"failed to grasp tomato; gripper missed grasp\",\n \"recovery_action\": \"Open the gripper, reposition the end effector directly above the tomato, close the gripper firmly around it, and proceed to carry it to the bin.\",\n \"recovery_action_summary\": \"reposition gripper; re-grasp tomato; place in bin\",\n \"prevention_action\": \"Position the gripper precisely over the tomato and ensure a stable grasp is established before initiating any lifting motion.\",\n \"prevention_action_summary\": \"align gripper over tomato; grasp securely; move to bin\",\n \"justification\": \"At around 3.3 seconds (between 00:03 and 00:04), the robot attempts to close its gripper on the tomato but fails to acquire a proper grip. For the remainder of the video, it remains stuck at the tomato position without lifting it.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 203,
+    "failure_sim_step": 203,
+    "failure_policy_step": 33,
+    "failure_chunk_index": 5,
+    "first_post_failure_policy_step": 34,
+    "failure_type": "grasp_failure",
+    "correction_action": "Open the gripper, reposition the end effector directly above the tomato, close the gripper firmly around it, and proceed to carry it to the bin."
+  }
+}
\ No newline at end of file
diff --git a/tomato/ep035/labels.npz b/tomato/ep035/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..6619bcd46a7bff26a4c76b887e574b55f3c54bcc
--- /dev/null
+++ b/tomato/ep035/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3c56664709037bfe24c7f5e7c3fadf380b39c906fcc57e1533b32ffae656dcbe
+size 3911
diff --git a/tomato/ep036/example.md b/tomato/ep036/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..4873866e570f431892ae2dfd35bf5595bb20d388
--- /dev/null
+++ b/tomato/ep036/example.md
@@ -0,0 +1,37 @@
+# t003__ep036__tomato
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the tomato in the bin
+- **success:** False
+- **control steps:** 925  (18.4s at 50 Hz)
+- **policy inferences:** 161  (replan_steps=6)
+- **video:** `tomato/ep036/rollout.mp4`
+- **features:** `161` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | reaching down toward tomato |
+| 3.0–6.0 | 151–301 | 25–50 | closing gripper and dropping tomato |
+| 6.0–9.0 | 302–452 | 50–75 | adjusting arm position over tomato |
+| 9.0–12.0 | 453–603 | 75–100 | lowering gripper for second grasp |
+| 12.0–15.0 | 604–754 | 100–125 | clamping tomato onto table surface |
+| 15.0–18.0 | 755–905 | 125–150 | holding closed gripper on tomato |
+| 18.0–18.4 | 906–924 | 151–154 | remaining stationary above tomato |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 6.175298804780876 s (00:06.2)
+- **frame / control_step:** 310 (policy_step=51, chunk_index=4)
+- **coarse onset:** 5.5 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the tomato, but failed to establish a firm grip, causing the tomato to slip out of the gripper as soon as the arm lifted.
+  Summary: improper grasp alignment; tomato slipping from gripper
+- **recovery:** Re-center the gripper directly over the tomato, close the fingers firmly around it, and lift it clear of the table before moving it into the bin.
+  Summary: re-align gripper; execute firm grasp; transport tomato to bin
+- **prevention:** Position the gripper precisely around the center of the tomato and verify a secure grip before initiating the upward lifting motion.
+  Summary: center gripper over tomato; secure firm hold before lifting
+- **justification:** At around 5.5 seconds, the robot closes its gripper on the tomato and begins lifting, but the tomato immediately slips out back onto the table. Subsequent attempts to re-grasp also fail to securely lift the object.
diff --git a/tomato/ep036/labels.json b/tomato/ep036/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..6abf7e2a4a25db49e36dabc35fb7dea9bc297f28
--- /dev/null
+++ b/tomato/ep036/labels.json
@@ -0,0 +1,142 @@
+{
+  "rollout_id": "t003__ep036__tomato",
+  "scene_variant": "",
+  "task_id": 3,
+  "task_file": "",
+  "instruction": "Put the tomato in the bin",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T22:05:30+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00796812749004,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching down toward tomato",
+      "description": "reaching down toward tomato"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00796812749004,
+      "t_end_sec": 6.01593625498008,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "closing gripper and dropping tomato",
+      "description": "closing gripper and dropping tomato"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01593625498008,
+      "t_end_sec": 9.02390438247012,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "adjusting arm position over tomato",
+      "description": "adjusting arm position over tomato"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.02390438247012,
+      "t_end_sec": 12.03187250996016,
+      "control_step_start": 453,
+      "control_step_end": 603,
+      "policy_step_start": 75,
+      "policy_step_end": 100,
+      "phrase": "lowering gripper for second grasp",
+      "description": "lowering gripper for second grasp"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 12.03187250996016,
+      "t_end_sec": 15.0398406374502,
+      "control_step_start": 604,
+      "control_step_end": 754,
+      "policy_step_start": 100,
+      "policy_step_end": 125,
+      "phrase": "clamping tomato onto table surface",
+      "description": "clamping tomato onto table surface"
+    },
+    {
+      "segment_index": 5,
+      "t_start_sec": 15.0398406374502,
+      "t_end_sec": 18.04780876494024,
+      "control_step_start": 755,
+      "control_step_end": 905,
+      "policy_step_start": 125,
+      "policy_step_end": 150,
+      "phrase": "holding closed gripper on tomato",
+      "description": "holding closed gripper on tomato"
+    },
+    {
+      "segment_index": 6,
+      "t_start_sec": 18.04780876494024,
+      "t_end_sec": 18.426294820717132,
+      "control_step_start": 906,
+      "control_step_end": 924,
+      "policy_step_start": 151,
+      "policy_step_end": 154,
+      "phrase": "remaining stationary above tomato",
+      "description": "remaining stationary above tomato"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 6.175298804780876,
+    "vlm_failure_onset_timestamp": "00:06.2",
+    "vlm_failure_onset_frame": 310,
+    "vlm_failure_onset_step": 310,
+    "vlm_coarse_failure_onset_seconds": 5.5,
+    "vlm_failure_window_start_seconds": 4.5,
+    "vlm_failure_window_end_seconds": 6.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the tomato, but failed to establish a firm grip, causing the tomato to slip out of the gripper as soon as the arm lifted.",
+    "vlm_failure_reason_summary": "improper grasp alignment; tomato slipping from gripper",
+    "vlm_recovery_action": "Re-center the gripper directly over the tomato, close the fingers firmly around it, and lift it clear of the table before moving it into the bin.",
+    "vlm_recovery_action_summary": "re-align gripper; execute firm grasp; transport tomato to bin",
+    "vlm_prevention_action": "Position the gripper precisely around the center of the tomato and verify a secure grip before initiating the upward lifting motion.",
+    "vlm_prevention_action_summary": "center gripper over tomato; secure firm hold before lifting",
+    "vlm_justification": "At around 5.5 seconds, the robot closes its gripper on the tomato and begins lifting, but the tomato immediately slips out back onto the table. Subsequent attempts to re-grasp also fail to securely lift the object.",
+    "vlm_temporal_justification": "At 03:05 in the clip, the robotic arm begins lifting upward, but fails to secure the tomato, causing it to slip out and remain on the table.",
+    "vlm_video_fps": 50.2,
+    "vlm_video_num_frames": 925,
+    "vlm_video_duration_seconds": 18.40637450199203,
+    "vlm_usage": {
+      "input_tokens": 22086,
+      "output_tokens": 371,
+      "cached_tokens": 1225
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 5.5,\n \"failure_window_start_seconds\": 4.5,\n \"failure_window_end_seconds\": 6.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the tomato, but failed to establish a firm grip, causing the tomato to slip out of the gripper as soon as the arm lifted.\",\n \"failure_reason_summary\": \"improper grasp alignment; tomato slipping from gripper\",\n \"recovery_action\": \"Re-center the gripper directly over the tomato, close the fingers firmly around it, and lift it clear of the table before moving it into the bin.\",\n \"recovery_action_summary\": \"re-align gripper; execute firm grasp; transport tomato to bin\",\n \"prevention_action\": \"Position the gripper precisely around the center of the tomato and verify a secure grip before initiating the upward lifting motion.\",\n \"prevention_action_summary\": \"center gripper over tomato; secure firm hold before lifting\",\n \"justification\": \"At around 5.5 seconds, the robot closes its gripper on the tomato and begins lifting, but the tomato immediately slips out back onto the table. Subsequent attempts to re-grasp also fail to securely lift the object.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 310,
+    "failure_sim_step": 310,
+    "failure_policy_step": 51,
+    "failure_chunk_index": 4,
+    "first_post_failure_policy_step": 52,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-center the gripper directly over the tomato, close the fingers firmly around it, and lift it clear of the table before moving it into the bin."
+  }
+}
\ No newline at end of file
diff --git a/tomato/ep036/labels.npz b/tomato/ep036/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..d4fdade4d8ca6c9bf9d8d1fdc131c163f7a727b0
--- /dev/null
+++ b/tomato/ep036/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c72c97c3fcb39041e38827d47ee9eddf43f40659a357e0ed81b9e1620f4a3687
+size 4035
diff --git a/tomato/ep037/example.md b/tomato/ep037/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..7acac84de2469af4a2743c8ddb3461eaf1ab150e
--- /dev/null
+++ b/tomato/ep037/example.md
@@ -0,0 +1,34 @@
+# t003__ep037__tomato
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the tomato in the bin
+- **success:** False
+- **control steps:** 488  (9.6s at 51 Hz)
+- **policy inferences:** 83  (replan_steps=6)
+- **video:** `tomato/ep037/rollout.mp4`
+- **features:** `83` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching down toward tomato |
+| 3.0–6.0 | 152–303 | 25–50 | grasping and lifting tomato |
+| 6.0–9.0 | 304–455 | 50–75 | dropping tomato and moving toward bin |
+| 9.0–9.6 | 456–487 | 76–81 | hovering empty gripper over bin |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 3.5629921259842523 s (00:03.6)
+- **frame / control_step:** 181 (policy_step=30, chunk_index=1)
+- **coarse onset:** 6.5 s (refined=True)
+- **mode:** object_displacement
+- **reason:** The robot grasped the tomato and began lifting it, but lost its grip, dropping the tomato back onto the table.
+  Summary: losing grip on tomato; dropping object onto table
+- **recovery:** Re-align the end-effector over the dropped tomato, execute a firm grasp, and carry it into the bin.
+  Summary: re-grasp tomato; carry tomato to bin
+- **prevention:** Apply and maintain sufficient gripping force when lifting the tomato to prevent it from slipping out during transport.
+  Summary: secure grip on tomato; maintain grasp while lifting
+- **justification:** At around 6.5 seconds, as the robot lifts the tomato off the table, the object slips from the gripper fingers and falls back onto the desk. The robot arm continues moving to the bin with empty grippers, failing the task.
diff --git a/tomato/ep037/labels.json b/tomato/ep037/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..fb4314d7729683665c116a873e2d162d8ca2a85d
--- /dev/null
+++ b/tomato/ep037/labels.json
@@ -0,0 +1,109 @@
+{
+  "rollout_id": "t003__ep037__tomato",
+  "scene_variant": "",
+  "task_id": 3,
+  "task_file": "",
+  "instruction": "Put the tomato in the bin",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T22:05:58+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9921259842519685,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching down toward tomato",
+      "description": "reaching down toward tomato"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9921259842519685,
+      "t_end_sec": 5.984251968503937,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "grasping and lifting tomato",
+      "description": "grasping and lifting tomato"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.984251968503937,
+      "t_end_sec": 8.976377952755906,
+      "control_step_start": 304,
+      "control_step_end": 455,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "dropping tomato and moving toward bin",
+      "description": "dropping tomato and moving toward bin"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.976377952755906,
+      "t_end_sec": 9.606299212598426,
+      "control_step_start": 456,
+      "control_step_end": 487,
+      "policy_step_start": 76,
+      "policy_step_end": 81,
+      "phrase": "hovering empty gripper over bin",
+      "description": "hovering empty gripper over bin"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "object_displacement",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 3.5629921259842523,
+    "vlm_failure_onset_timestamp": "00:03.6",
+    "vlm_failure_onset_frame": 181,
+    "vlm_failure_onset_step": 181,
+    "vlm_coarse_failure_onset_seconds": 6.5,
+    "vlm_failure_window_start_seconds": 6.0,
+    "vlm_failure_window_end_seconds": 7.0,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot grasped the tomato and began lifting it, but lost its grip, dropping the tomato back onto the table.",
+    "vlm_failure_reason_summary": "losing grip on tomato; dropping object onto table",
+    "vlm_recovery_action": "Re-align the end-effector over the dropped tomato, execute a firm grasp, and carry it into the bin.",
+    "vlm_recovery_action_summary": "re-grasp tomato; carry tomato to bin",
+    "vlm_prevention_action": "Apply and maintain sufficient gripping force when lifting the tomato to prevent it from slipping out during transport.",
+    "vlm_prevention_action_summary": "secure grip on tomato; maintain grasp while lifting",
+    "vlm_justification": "At around 6.5 seconds, as the robot lifts the tomato off the table, the object slips from the gripper fingers and falls back onto the desk. The robot arm continues moving to the bin with empty grippers, failing the task.",
+    "vlm_temporal_justification": "The robot lifts the tomato off the table, but at clip timestamp 03:17 the tomato slips out from between the gripper fingers and falls back onto the table surface.",
+    "vlm_video_fps": 50.8,
+    "vlm_video_num_frames": 488,
+    "vlm_video_duration_seconds": 9.586614173228346,
+    "vlm_usage": {
+      "input_tokens": 21696,
+      "output_tokens": 353,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"object_displacement\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 6.5,\n \"failure_window_start_seconds\": 6.0,\n \"failure_window_end_seconds\": 7.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot grasped the tomato and began lifting it, but lost its grip, dropping the tomato back onto the table.\",\n \"failure_reason_summary\": \"losing grip on tomato; dropping object onto table\",\n \"recovery_action\": \"Re-align the end-effector over the dropped tomato, execute a firm grasp, and carry it into the bin.\",\n \"recovery_action_summary\": \"re-grasp tomato; carry tomato to bin\",\n \"prevention_action\": \"Apply and maintain sufficient gripping force when lifting the tomato to prevent it from slipping out during transport.\",\n \"prevention_action_summary\": \"secure grip on tomato; maintain grasp while lifting\",\n \"justification\": \"At around 6.5 seconds, as the robot lifts the tomato off the table, the object slips from the gripper fingers and falls back onto the desk. The robot arm continues moving to the bin with empty grippers, failing the task.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 181,
+    "failure_sim_step": 181,
+    "failure_policy_step": 30,
+    "failure_chunk_index": 1,
+    "first_post_failure_policy_step": 31,
+    "failure_type": "object_displacement",
+    "correction_action": "Re-align the end-effector over the dropped tomato, execute a firm grasp, and carry it into the bin."
+  }
+}
\ No newline at end of file
diff --git a/tomato/ep037/labels.npz b/tomato/ep037/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..5617e69128d9bd975d5b1ee471c4949355033a73
--- /dev/null
+++ b/tomato/ep037/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b36356bb475b74bc70ef36d37e042024f80b6f04c803865e1313da65bdc51ee5
+size 3766
diff --git a/tomato/ep038/example.md b/tomato/ep038/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..843be26f0e85883411b226bcfcd4a09ad6ed524f
--- /dev/null
+++ b/tomato/ep038/example.md
@@ -0,0 +1,24 @@
+# t003__ep038__tomato
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the tomato in the bin
+- **success:** True
+- **control steps:** 764  (15.2s at 50 Hz)
+- **policy inferences:** 132  (replan_steps=6)
+- **video:** `tomato/ep038/rollout.mp4`
+- **features:** `132` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | reaching for tomato |
+| 3.0–6.0 | 151–301 | 25–50 | grasping tomato |
+| 6.0–9.0 | 302–452 | 50–75 | lifting tomato |
+| 9.0–12.0 | 453–603 | 75–100 | moving tomato toward bin |
+| 12.0–15.0 | 604–754 | 100–125 | releasing tomato in bin |
+| 15.0–15.2 | 755–763 | 125–127 | retracting arm |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/tomato/ep038/labels.json b/tomato/ep038/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..225868c3730d231799323b1e29a9e69da9d0eb1c
--- /dev/null
+++ b/tomato/ep038/labels.json
@@ -0,0 +1,101 @@
+{
+  "rollout_id": "t003__ep038__tomato",
+  "scene_variant": "",
+  "task_id": 3,
+  "task_file": "",
+  "instruction": "Put the tomato in the bin",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T22:06:07+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.001988071570577,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for tomato",
+      "description": "reaching for tomato"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.001988071570577,
+      "t_end_sec": 6.003976143141154,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "grasping tomato",
+      "description": "grasping tomato"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.003976143141154,
+      "t_end_sec": 9.00596421471173,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "lifting tomato",
+      "description": "lifting tomato"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.00596421471173,
+      "t_end_sec": 12.007952286282308,
+      "control_step_start": 453,
+      "control_step_end": 603,
+      "policy_step_start": 75,
+      "policy_step_end": 100,
+      "phrase": "moving tomato toward bin",
+      "description": "moving tomato toward bin"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 12.007952286282308,
+      "t_end_sec": 15.009940357852884,
+      "control_step_start": 604,
+      "control_step_end": 754,
+      "policy_step_start": 100,
+      "policy_step_end": 125,
+      "phrase": "releasing tomato in bin",
+      "description": "releasing tomato in bin"
+    },
+    {
+      "segment_index": 5,
+      "t_start_sec": 15.009940357852884,
+      "t_end_sec": 15.188866799204773,
+      "control_step_start": 755,
+      "control_step_end": 763,
+      "policy_step_start": 125,
+      "policy_step_end": 127,
+      "phrase": "retracting arm",
+      "description": "retracting arm"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/tomato/ep038/labels.npz b/tomato/ep038/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..ccb8c1593b9e29731992e29fd77a9bf4ef9d811d
--- /dev/null
+++ b/tomato/ep038/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9c21560db336c9823b783cca5b3b4fe800f3265bab1247689ff4c94da5baae96
+size 3156
diff --git a/tomato/ep039/example.md b/tomato/ep039/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..e814266d10b358829d424f6b9404afeb0207f697
--- /dev/null
+++ b/tomato/ep039/example.md
@@ -0,0 +1,23 @@
+# t003__ep039__tomato
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the tomato in the bin
+- **success:** True
+- **control steps:** 626  (12.3s at 51 Hz)
+- **policy inferences:** 107  (replan_steps=6)
+- **video:** `tomato/ep039/rollout.mp4`
+- **features:** `107` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching for tomato |
+| 3.0–6.0 | 152–303 | 25–50 | grasping tomato |
+| 6.0–9.0 | 304–455 | 50–75 | moving tomato to bin |
+| 9.0–12.0 | 456–607 | 76–101 | releasing tomato into bin |
+| 12.0–12.3 | 608–625 | 101–104 | idle / holding position |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/tomato/ep039/labels.json b/tomato/ep039/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..5322336ed47a561a66fbeba8b1415a11da092475
--- /dev/null
+++ b/tomato/ep039/labels.json
@@ -0,0 +1,90 @@
+{
+  "rollout_id": "t003__ep039__tomato",
+  "scene_variant": "",
+  "task_id": 3,
+  "task_file": "",
+  "instruction": "Put the tomato in the bin",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T22:06:16+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9921259842519685,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for tomato",
+      "description": "reaching for tomato"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9921259842519685,
+      "t_end_sec": 5.984251968503937,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "grasping tomato",
+      "description": "grasping tomato"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.984251968503937,
+      "t_end_sec": 8.976377952755906,
+      "control_step_start": 304,
+      "control_step_end": 455,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "moving tomato to bin",
+      "description": "moving tomato to bin"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.976377952755906,
+      "t_end_sec": 11.968503937007874,
+      "control_step_start": 456,
+      "control_step_end": 607,
+      "policy_step_start": 76,
+      "policy_step_end": 101,
+      "phrase": "releasing tomato into bin",
+      "description": "releasing tomato into bin"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 11.968503937007874,
+      "t_end_sec": 12.322834645669293,
+      "control_step_start": 608,
+      "control_step_end": 625,
+      "policy_step_start": 101,
+      "policy_step_end": 104,
+      "phrase": "idle / holding position",
+      "description": "idle / holding position"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/tomato/ep039/labels.npz b/tomato/ep039/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..08f4ba9a8f053b6406d20b4fa3612e5d6820575e
--- /dev/null
+++ b/tomato/ep039/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:638c067954fc498197abbf54239293d287905f0e60c920e6432e336a63fe5499
+size 3128
diff --git a/tomato/ep040/example.md b/tomato/ep040/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..c789372f9ce5fda581b4e5360d48f24a85fb5e2e
--- /dev/null
+++ b/tomato/ep040/example.md
@@ -0,0 +1,23 @@
+# t003__ep040__tomato
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the tomato in the bin
+- **success:** True
+- **control steps:** 678  (13.4s at 50 Hz)
+- **policy inferences:** 117  (replan_steps=6)
+- **video:** `tomato/ep040/rollout.mp4`
+- **features:** `117` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching for tomato |
+| 3.0–6.0 | 152–303 | 25–50 | grasping tomato |
+| 6.0–9.0 | 304–455 | 50–75 | lifting and moving tomato |
+| 9.0–12.0 | 456–607 | 76–101 | placing tomato into bin |
+| 12.0–13.4 | 608–677 | 101–112 | retracting arm |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/tomato/ep040/labels.json b/tomato/ep040/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..e8ec7c5df902ec24ea037c2e684f8f8e8cfaac79
--- /dev/null
+++ b/tomato/ep040/labels.json
@@ -0,0 +1,90 @@
+{
+  "rollout_id": "t003__ep040__tomato",
+  "scene_variant": "",
+  "task_id": 3,
+  "task_file": "",
+  "instruction": "Put the tomato in the bin",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T22:06:24+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00990099009901,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for tomato",
+      "description": "reaching for tomato"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00990099009901,
+      "t_end_sec": 6.01980198019802,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "grasping tomato",
+      "description": "grasping tomato"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01980198019802,
+      "t_end_sec": 9.029702970297029,
+      "control_step_start": 304,
+      "control_step_end": 455,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "lifting and moving tomato",
+      "description": "lifting and moving tomato"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.029702970297029,
+      "t_end_sec": 12.03960396039604,
+      "control_step_start": 456,
+      "control_step_end": 607,
+      "policy_step_start": 76,
+      "policy_step_end": 101,
+      "phrase": "placing tomato into bin",
+      "description": "placing tomato into bin"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 12.03960396039604,
+      "t_end_sec": 13.425742574257425,
+      "control_step_start": 608,
+      "control_step_end": 677,
+      "policy_step_start": 101,
+      "policy_step_end": 112,
+      "phrase": "retracting arm",
+      "description": "retracting arm"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/tomato/ep040/labels.npz b/tomato/ep040/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..3bebd8beef8b1676a07149cc531514ff428fdd9d
--- /dev/null
+++ b/tomato/ep040/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:904cb24fc9805949efbeb29757ca983a8f97d80d154a6a9d7df5bb765253a9b3
+size 3132
diff --git a/tomato/ep041/example.md b/tomato/ep041/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..fcd46b44c6dd00e81413c5643d5fb4349bc3b64c
--- /dev/null
+++ b/tomato/ep041/example.md
@@ -0,0 +1,22 @@
+# t003__ep041__tomato
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the tomato in the bin
+- **success:** True
+- **control steps:** 458  (9.1s at 50 Hz)
+- **policy inferences:** 80  (replan_steps=6)
+- **video:** `tomato/ep041/rollout.mp4`
+- **features:** `80` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | reaching for the tomato |
+| 3.0–6.0 | 151–301 | 25–50 | grasping and lifting tomato |
+| 6.0–9.0 | 302–452 | 50–75 | placing tomato in bin |
+| 9.0–9.1 | 453–457 | 75–76 | retracting arm and idle |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/tomato/ep041/labels.json b/tomato/ep041/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..f962c64b2b0d7fb70f6b416310f56ee63b82fe52
--- /dev/null
+++ b/tomato/ep041/labels.json
@@ -0,0 +1,79 @@
+{
+  "rollout_id": "t003__ep041__tomato",
+  "scene_variant": "",
+  "task_id": 3,
+  "task_file": "",
+  "instruction": "Put the tomato in the bin",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T22:06:33+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.001988071570577,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for the tomato",
+      "description": "reaching for the tomato"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.001988071570577,
+      "t_end_sec": 6.003976143141154,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "grasping and lifting tomato",
+      "description": "grasping and lifting tomato"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.003976143141154,
+      "t_end_sec": 9.00596421471173,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "placing tomato in bin",
+      "description": "placing tomato in bin"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.00596421471173,
+      "t_end_sec": 9.105367793240557,
+      "control_step_start": 453,
+      "control_step_end": 457,
+      "policy_step_start": 75,
+      "policy_step_end": 76,
+      "phrase": "retracting arm and idle",
+      "description": "retracting arm and idle"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/tomato/ep041/labels.npz b/tomato/ep041/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..17c5768f44b48945022f6f2474973a35fcc5d1d7
--- /dev/null
+++ b/tomato/ep041/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3c1a66f033d70a17c5b79e2cbe20deb2204ab23e20ceeb290134d379a907c40e
+size 3110
diff --git a/tomato/ep042/example.md b/tomato/ep042/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..6cc0182b546d8a0a639262bb4228a9554ea880bf
--- /dev/null
+++ b/tomato/ep042/example.md
@@ -0,0 +1,22 @@
+# t003__ep042__tomato
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the tomato in the bin
+- **success:** True
+- **control steps:** 496  (9.9s at 50 Hz)
+- **policy inferences:** 85  (replan_steps=6)
+- **video:** `tomato/ep042/rollout.mp4`
+- **features:** `85` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | reaching for the tomato |
+| 3.0–6.0 | 151–301 | 25–50 | grasping the tomato |
+| 6.0–9.0 | 302–452 | 50–75 | placing tomato in the bin |
+| 9.0–9.9 | 453–495 | 75–82 | retracting to home position |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/tomato/ep042/labels.json b/tomato/ep042/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..8bb098df3447526294ee41e21190a4a35bc948e5
--- /dev/null
+++ b/tomato/ep042/labels.json
@@ -0,0 +1,79 @@
+{
+  "rollout_id": "t003__ep042__tomato",
+  "scene_variant": "",
+  "task_id": 3,
+  "task_file": "",
+  "instruction": "Put the tomato in the bin",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T22:06:42+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00796812749004,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for the tomato",
+      "description": "reaching for the tomato"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00796812749004,
+      "t_end_sec": 6.01593625498008,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "grasping the tomato",
+      "description": "grasping the tomato"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01593625498008,
+      "t_end_sec": 9.02390438247012,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "placing tomato in the bin",
+      "description": "placing tomato in the bin"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.02390438247012,
+      "t_end_sec": 9.880478087649402,
+      "control_step_start": 453,
+      "control_step_end": 495,
+      "policy_step_start": 75,
+      "policy_step_end": 82,
+      "phrase": "retracting to home position",
+      "description": "retracting to home position"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/tomato/ep042/labels.npz b/tomato/ep042/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..ddc4cb42bb5691f95f2ecccd6e40e15a965e785e
--- /dev/null
+++ b/tomato/ep042/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:83164588f9d5d2be87d9601dcbfea5f2a78fb088d350a8609cba39de77b28332
+size 3108
diff --git a/tomato/ep043/example.md b/tomato/ep043/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..687f616e9b8bdf2c307d26a499fe6127f74f4396
--- /dev/null
+++ b/tomato/ep043/example.md
@@ -0,0 +1,33 @@
+# t003__ep043__tomato
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the tomato in the bin
+- **success:** False
+- **control steps:** 432  (8.5s at 51 Hz)
+- **policy inferences:** 73  (replan_steps=6)
+- **video:** `tomato/ep043/rollout.mp4`
+- **features:** `73` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–152 | 0–25 | reaching toward tomato |
+| 3.0–6.0 | 153–305 | 25–50 | closing gripper and lifting empty |
+| 6.0–8.5 | 306–431 | 51–71 | moving empty gripper over bin |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 6.2426614481409 s (00:06.2)
+- **frame / control_step:** 319 (policy_step=53, chunk_index=1)
+- **coarse onset:** 4.5 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the tomato, but closed its gripper loosely or too high, leaving the tomato on the table as the arm lifted away.
+  Summary: failed grasp on tomato; lifting empty gripper
+- **recovery:** Lower the end effector back over the tomato, close the gripper securely around it, and carry it to the bin.
+  Summary: re-align gripper; regrasp tomato; move to bin
+- **prevention:** Ensure the gripper descends sufficiently low over the center of the tomato before closing to establish a firm grasp.
+  Summary: lower gripper further; close firmly on tomato
+- **justification:** At around 4.5 seconds, the gripper closes around the tomato but fails to secure it. When the arm lifts at 5.0 seconds, the tomato remains on the table while the arm moves toward the bin empty-handed.
diff --git a/tomato/ep043/labels.json b/tomato/ep043/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..5aed2479da55f2221823d7c6c48cca9f195c1587
--- /dev/null
+++ b/tomato/ep043/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t003__ep043__tomato",
+  "scene_variant": "",
+  "task_id": 3,
+  "task_file": "",
+  "instruction": "Put the tomato in the bin",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T22:07:18+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.99412915851272,
+      "control_step_start": 0,
+      "control_step_end": 152,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching toward tomato",
+      "description": "reaching toward tomato"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.99412915851272,
+      "t_end_sec": 5.98825831702544,
+      "control_step_start": 153,
+      "control_step_end": 305,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "closing gripper and lifting empty",
+      "description": "closing gripper and lifting empty"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.98825831702544,
+      "t_end_sec": 8.454011741682974,
+      "control_step_start": 306,
+      "control_step_end": 431,
+      "policy_step_start": 51,
+      "policy_step_end": 71,
+      "phrase": "moving empty gripper over bin",
+      "description": "moving empty gripper over bin"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 6.2426614481409,
+    "vlm_failure_onset_timestamp": "00:06.2",
+    "vlm_failure_onset_frame": 319,
+    "vlm_failure_onset_step": 319,
+    "vlm_coarse_failure_onset_seconds": 4.5,
+    "vlm_failure_window_start_seconds": 4.0,
+    "vlm_failure_window_end_seconds": 5.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the tomato, but closed its gripper loosely or too high, leaving the tomato on the table as the arm lifted away.",
+    "vlm_failure_reason_summary": "failed grasp on tomato; lifting empty gripper",
+    "vlm_recovery_action": "Lower the end effector back over the tomato, close the gripper securely around it, and carry it to the bin.",
+    "vlm_recovery_action_summary": "re-align gripper; regrasp tomato; move to bin",
+    "vlm_prevention_action": "Ensure the gripper descends sufficiently low over the center of the tomato before closing to establish a firm grasp.",
+    "vlm_prevention_action_summary": "lower gripper further; close firmly on tomato",
+    "vlm_justification": "At around 4.5 seconds, the gripper closes around the tomato but fails to secure it. When the arm lifts at 5.0 seconds, the tomato remains on the table while the arm moves toward the bin empty-handed.",
+    "vlm_temporal_justification": "At clip second 242 (04:02), the robot finishes closing its gripper fingers entirely without grasping the tomato, leaving the tomato sitting untouched on the table.",
+    "vlm_video_fps": 51.1,
+    "vlm_video_num_frames": 432,
+    "vlm_video_duration_seconds": 8.434442270058709,
+    "vlm_usage": {
+      "input_tokens": 21699,
+      "output_tokens": 357,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 4.5,\n \"failure_window_start_seconds\": 4.0,\n \"failure_window_end_seconds\": 5.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the tomato, but closed its gripper loosely or too high, leaving the tomato on the table as the arm lifted away.\",\n \"failure_reason_summary\": \"failed grasp on tomato; lifting empty gripper\",\n \"recovery_action\": \"Lower the end effector back over the tomato, close the gripper securely around it, and carry it to the bin.\",\n \"recovery_action_summary\": \"re-align gripper; regrasp tomato; move to bin\",\n \"prevention_action\": \"Ensure the gripper descends sufficiently low over the center of the tomato before closing to establish a firm grasp.\",\n \"prevention_action_summary\": \"lower gripper further; close firmly on tomato\",\n \"justification\": \"At around 4.5 seconds, the gripper closes around the tomato but fails to secure it. When the arm lifts at 5.0 seconds, the tomato remains on the table while the arm moves toward the bin empty-handed.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 319,
+    "failure_sim_step": 319,
+    "failure_policy_step": 53,
+    "failure_chunk_index": 1,
+    "first_post_failure_policy_step": 54,
+    "failure_type": "grasp_failure",
+    "correction_action": "Lower the end effector back over the tomato, close the gripper securely around it, and carry it to the bin."
+  }
+}
\ No newline at end of file
diff --git a/tomato/ep043/labels.npz b/tomato/ep043/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..c22498bb194029b0851b9fec861d047fbf4547f2
--- /dev/null
+++ b/tomato/ep043/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1fe26337b9d616c0b022413906974b32406428fd7896857df4bb7b97c433f14b
+size 3756
diff --git a/tomato/ep044/example.md b/tomato/ep044/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..e059d39b407c959b263e62d7aabdd42b737807da
--- /dev/null
+++ b/tomato/ep044/example.md
@@ -0,0 +1,33 @@
+# t003__ep044__tomato
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the tomato in the bin
+- **success:** False
+- **control steps:** 403  (8.1s at 50 Hz)
+- **policy inferences:** 70  (replan_steps=6)
+- **video:** `tomato/ep044/rollout.mp4`
+- **features:** `70` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–149 | 0–24 | lowering gripper toward tomato |
+| 3.0–6.0 | 150–299 | 25–49 | attempting grasp; slipping off tomato |
+| 6.0–8.1 | 300–402 | 50–67 | holding closed gripper near object |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 4.46 s (00:04.5)
+- **frame / control_step:** 223 (policy_step=37, chunk_index=1)
+- **coarse onset:** 4.0 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the tomato, but the gripper failed to establish a secure hold, causing the tomato to slip out as the fingers closed.
+  Summary: failed grasp attempt; tomato slipped from gripper
+- **recovery:** Re-open the gripper, re-align the end effector directly over the center of the tomato, and attempt to grasp it again before moving to the bin.
+  Summary: re-align gripper over tomato; retry grasp
+- **prevention:** Lower the gripper further and center it precisely over the tomato prior to closing the fingers to ensure a successful hold.
+  Summary: center gripper properly; secure firm grip
+- **justification:** At around 4.0 seconds, the robot closes its gripper on the tomato, but fails to grasp it firmly. The tomato rolls out of the fingers as they close, leaving the robot without the object.
diff --git a/tomato/ep044/labels.json b/tomato/ep044/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..5a882b03bf0a7a2fa764b5cf7ed48e64bb60b95b
--- /dev/null
+++ b/tomato/ep044/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t003__ep044__tomato",
+  "scene_variant": "",
+  "task_id": 3,
+  "task_file": "",
+  "instruction": "Put the tomato in the bin",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T22:07:50+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.0,
+      "control_step_start": 0,
+      "control_step_end": 149,
+      "policy_step_start": 0,
+      "policy_step_end": 24,
+      "phrase": "lowering gripper toward tomato",
+      "description": "lowering gripper toward tomato"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.0,
+      "t_end_sec": 6.0,
+      "control_step_start": 150,
+      "control_step_end": 299,
+      "policy_step_start": 25,
+      "policy_step_end": 49,
+      "phrase": "attempting grasp; slipping off tomato",
+      "description": "attempting grasp; slipping off tomato"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.0,
+      "t_end_sec": 8.06,
+      "control_step_start": 300,
+      "control_step_end": 402,
+      "policy_step_start": 50,
+      "policy_step_end": 67,
+      "phrase": "holding closed gripper near object",
+      "description": "holding closed gripper near object"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 4.46,
+    "vlm_failure_onset_timestamp": "00:04.5",
+    "vlm_failure_onset_frame": 223,
+    "vlm_failure_onset_step": 223,
+    "vlm_coarse_failure_onset_seconds": 4.0,
+    "vlm_failure_window_start_seconds": 3.5,
+    "vlm_failure_window_end_seconds": 5.0,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the tomato, but the gripper failed to establish a secure hold, causing the tomato to slip out as the fingers closed.",
+    "vlm_failure_reason_summary": "failed grasp attempt; tomato slipped from gripper",
+    "vlm_recovery_action": "Re-open the gripper, re-align the end effector directly over the center of the tomato, and attempt to grasp it again before moving to the bin.",
+    "vlm_recovery_action_summary": "re-align gripper over tomato; retry grasp",
+    "vlm_prevention_action": "Lower the gripper further and center it precisely over the tomato prior to closing the fingers to ensure a successful hold.",
+    "vlm_prevention_action_summary": "center gripper properly; secure firm grip",
+    "vlm_justification": "At around 4.0 seconds, the robot closes its gripper on the tomato, but fails to grasp it firmly. The tomato rolls out of the fingers as they close, leaving the robot without the object.",
+    "vlm_temporal_justification": "The robot lowers its end-effector over the tomato and begins closing its gripper fingers around 02:52 (172s). By 02:53 (173s), the gripper attempts to secure the tomato, but fails to establish a hold as the fingers close improperly around it.",
+    "vlm_video_fps": 50.0,
+    "vlm_video_num_frames": 403,
+    "vlm_video_duration_seconds": 8.04,
+    "vlm_usage": {
+      "input_tokens": 21321,
+      "output_tokens": 386,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 4.0,\n \"failure_window_start_seconds\": 3.5,\n \"failure_window_end_seconds\": 5.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the tomato, but the gripper failed to establish a secure hold, causing the tomato to slip out as the fingers closed.\",\n \"failure_reason_summary\": \"failed grasp attempt; tomato slipped from gripper\",\n \"recovery_action\": \"Re-open the gripper, re-align the end effector directly over the center of the tomato, and attempt to grasp it again before moving to the bin.\",\n \"recovery_action_summary\": \"re-align gripper over tomato; retry grasp\",\n \"prevention_action\": \"Lower the gripper further and center it precisely over the tomato prior to closing the fingers to ensure a successful hold.\",\n \"prevention_action_summary\": \"center gripper properly; secure firm grip\",\n \"justification\": \"At around 4.0 seconds, the robot closes its gripper on the tomato, but fails to grasp it firmly. The tomato rolls out of the fingers as they close, leaving the robot without the object.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 223,
+    "failure_sim_step": 223,
+    "failure_policy_step": 37,
+    "failure_chunk_index": 1,
+    "first_post_failure_policy_step": 38,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-open the gripper, re-align the end effector directly over the center of the tomato, and attempt to grasp it again before moving to the bin."
+  }
+}
\ No newline at end of file
diff --git a/tomato/ep044/labels.npz b/tomato/ep044/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..d4ebb1ea4733429c53a78718419b51632a72508c
--- /dev/null
+++ b/tomato/ep044/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ecc45ff499db1fe00d9db5c739effc1271ba33d364e86664a092d386fd70ef5e
+size 3774
diff --git a/tomato/ep045/example.md b/tomato/ep045/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..5fd75088e2cb4121c371dcdf2f5be5cb7487c3cd
--- /dev/null
+++ b/tomato/ep045/example.md
@@ -0,0 +1,34 @@
+# t003__ep045__tomato
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the tomato in the bin
+- **success:** False
+- **control steps:** 532  (10.5s at 50 Hz)
+- **policy inferences:** 92  (replan_steps=6)
+- **video:** `tomato/ep045/rollout.mp4`
+- **features:** `92` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–151 | 0–25 | reaching toward the tomato |
+| 3.0–6.0 | 152–303 | 25–50 | attempting grasp on tomato |
+| 6.0–9.0 | 304–455 | 50–75 | lifting empty gripper up |
+| 9.0–10.5 | 456–531 | 76–88 | retracting arm without tomato |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 4.653465346534653 s (00:04.7)
+- **frame / control_step:** 235 (policy_step=39, chunk_index=1)
+- **coarse onset:** 5.0 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot descended to grasp the tomato but failed to securely enclose it, closing its gripper empty and lifting away without the object.
+  Summary: failed grasp; gripper closed empty
+- **recovery:** Lower the end-effector back down over the tomato, align the open fingers around its center, close the gripper firmly around the tomato, and lift it into the bin.
+  Summary: re-align gripper; regrasp tomato; place in bin
+- **prevention:** Center the gripper directly over the tomato and maintain proper depth before closing fingers to ensure a secure hold before lifting.
+  Summary: center gripper over tomato; secure grip before lifting
+- **justification:** Around 5.0 seconds into the rollout, the robot attempts to pick up the tomato on the table, but the gripper closes without grasping the object and retracts empty.
diff --git a/tomato/ep045/labels.json b/tomato/ep045/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..b0c8f22e4d38987bb7961412121877e07ee87759
--- /dev/null
+++ b/tomato/ep045/labels.json
@@ -0,0 +1,109 @@
+{
+  "rollout_id": "t003__ep045__tomato",
+  "scene_variant": "",
+  "task_id": 3,
+  "task_file": "",
+  "instruction": "Put the tomato in the bin",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T22:08:22+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00990099009901,
+      "control_step_start": 0,
+      "control_step_end": 151,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching toward the tomato",
+      "description": "reaching toward the tomato"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00990099009901,
+      "t_end_sec": 6.01980198019802,
+      "control_step_start": 152,
+      "control_step_end": 303,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "attempting grasp on tomato",
+      "description": "attempting grasp on tomato"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01980198019802,
+      "t_end_sec": 9.029702970297029,
+      "control_step_start": 304,
+      "control_step_end": 455,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "lifting empty gripper up",
+      "description": "lifting empty gripper up"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.029702970297029,
+      "t_end_sec": 10.534653465346535,
+      "control_step_start": 456,
+      "control_step_end": 531,
+      "policy_step_start": 76,
+      "policy_step_end": 88,
+      "phrase": "retracting arm without tomato",
+      "description": "retracting arm without tomato"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 4.653465346534653,
+    "vlm_failure_onset_timestamp": "00:04.7",
+    "vlm_failure_onset_frame": 235,
+    "vlm_failure_onset_step": 235,
+    "vlm_coarse_failure_onset_seconds": 5.0,
+    "vlm_failure_window_start_seconds": 4.5,
+    "vlm_failure_window_end_seconds": 6.0,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot descended to grasp the tomato but failed to securely enclose it, closing its gripper empty and lifting away without the object.",
+    "vlm_failure_reason_summary": "failed grasp; gripper closed empty",
+    "vlm_recovery_action": "Lower the end-effector back down over the tomato, align the open fingers around its center, close the gripper firmly around the tomato, and lift it into the bin.",
+    "vlm_recovery_action_summary": "re-align gripper; regrasp tomato; place in bin",
+    "vlm_prevention_action": "Center the gripper directly over the tomato and maintain proper depth before closing fingers to ensure a secure hold before lifting.",
+    "vlm_prevention_action_summary": "center gripper over tomato; secure grip before lifting",
+    "vlm_justification": "Around 5.0 seconds into the rollout, the robot attempts to pick up the tomato on the table, but the gripper closes without grasping the object and retracts empty.",
+    "vlm_temporal_justification": "The robot gripper descends to grasp the tomato and closes completely without grasping it at 02:15 (frame 135), leaving the tomato sitting untouched on the table.",
+    "vlm_video_fps": 50.5,
+    "vlm_video_num_frames": 532,
+    "vlm_video_duration_seconds": 10.514851485148515,
+    "vlm_usage": {
+      "input_tokens": 21763,
+      "output_tokens": 353,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 5.0,\n \"failure_window_start_seconds\": 4.5,\n \"failure_window_end_seconds\": 6.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot descended to grasp the tomato but failed to securely enclose it, closing its gripper empty and lifting away without the object.\",\n \"failure_reason_summary\": \"failed grasp; gripper closed empty\",\n \"recovery_action\": \"Lower the end-effector back down over the tomato, align the open fingers around its center, close the gripper firmly around the tomato, and lift it into the bin.\",\n \"recovery_action_summary\": \"re-align gripper; regrasp tomato; place in bin\",\n \"prevention_action\": \"Center the gripper directly over the tomato and maintain proper depth before closing fingers to ensure a secure hold before lifting.\",\n \"prevention_action_summary\": \"center gripper over tomato; secure grip before lifting\",\n \"justification\": \"Around 5.0 seconds into the rollout, the robot attempts to pick up the tomato on the table, but the gripper closes without grasping the object and retracts empty.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 235,
+    "failure_sim_step": 235,
+    "failure_policy_step": 39,
+    "failure_chunk_index": 1,
+    "first_post_failure_policy_step": 40,
+    "failure_type": "grasp_failure",
+    "correction_action": "Lower the end-effector back down over the tomato, align the open fingers around its center, close the gripper firmly around the tomato, and lift it into the bin."
+  }
+}
\ No newline at end of file
diff --git a/tomato/ep045/labels.npz b/tomato/ep045/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..9227b8847eb6268d00b78d01ee353a7d96a6b1bb
--- /dev/null
+++ b/tomato/ep045/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:116d72da556e16c53b43368e7c318657831b55427cf982d3df78217147b454ca
+size 3822
diff --git a/tomato/ep046/example.md b/tomato/ep046/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..a1ede908f55969c3e823dad59e12fa5f419fe08b
--- /dev/null
+++ b/tomato/ep046/example.md
@@ -0,0 +1,35 @@
+# t003__ep046__tomato
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the tomato in the bin
+- **success:** False
+- **control steps:** 606  (12.1s at 50 Hz)
+- **policy inferences:** 106  (replan_steps=6)
+- **video:** `tomato/ep046/rollout.mp4`
+- **features:** `106` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–149 | 0–24 | reaching toward tomato |
+| 3.0–6.0 | 150–299 | 25–49 | attempting grasp; tomato slipping out |
+| 6.0–9.0 | 300–449 | 50–74 | attempting second grasp on tomato |
+| 9.0–12.0 | 450–599 | 75–99 | retracting arm; remaining idle |
+| 12.0–12.1 | 600–605 | 100–100 | idle above table |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 5.189620758483033 s (00:05.2)
+- **frame / control_step:** 260 (policy_step=43, chunk_index=2)
+- **coarse onset:** 4.0 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot attempted to grasp the tomato, but failed to secure a firm grip, causing the tomato to slip out of the gripper as soon as the arm lifted.
+  Summary: failed gripper contact; tomato slipping out
+- **recovery:** Re-align the gripper over the center of the tomato, close the fingers firmly around it to ensure a secure hold, lift, and place it into the bin.
+  Summary: re-align gripper; securely grasp tomato; place in bin
+- **prevention:** Ensure the gripper is properly centered and positioned lower over the tomato before closing the fingers to achieve a stable grasp prior to lifting.
+  Summary: center gripper over tomato; secure grasp before lifting
+- **justification:** At approximately 00:04 (4.0s), the robot closes its gripper on the tomato, but upon lifting at 00:05, the tomato immediately slips out and remains on the tabletop. A second grasp attempt at 00:08 fails similarly, resulting in task failure.
diff --git a/tomato/ep046/labels.json b/tomato/ep046/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..b53addd166e792625193d1cdd8a1041763d95502
--- /dev/null
+++ b/tomato/ep046/labels.json
@@ -0,0 +1,120 @@
+{
+  "rollout_id": "t003__ep046__tomato",
+  "scene_variant": "",
+  "task_id": 3,
+  "task_file": "",
+  "instruction": "Put the tomato in the bin",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T22:08:55+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9940119760479043,
+      "control_step_start": 0,
+      "control_step_end": 149,
+      "policy_step_start": 0,
+      "policy_step_end": 24,
+      "phrase": "reaching toward tomato",
+      "description": "reaching toward tomato"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9940119760479043,
+      "t_end_sec": 5.9880239520958085,
+      "control_step_start": 150,
+      "control_step_end": 299,
+      "policy_step_start": 25,
+      "policy_step_end": 49,
+      "phrase": "attempting grasp; tomato slipping out",
+      "description": "attempting grasp; tomato slipping out"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.9880239520958085,
+      "t_end_sec": 8.982035928143713,
+      "control_step_start": 300,
+      "control_step_end": 449,
+      "policy_step_start": 50,
+      "policy_step_end": 74,
+      "phrase": "attempting second grasp on tomato",
+      "description": "attempting second grasp on tomato"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.982035928143713,
+      "t_end_sec": 11.976047904191617,
+      "control_step_start": 450,
+      "control_step_end": 599,
+      "policy_step_start": 75,
+      "policy_step_end": 99,
+      "phrase": "retracting arm; remaining idle",
+      "description": "retracting arm; remaining idle"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 11.976047904191617,
+      "t_end_sec": 12.095808383233532,
+      "control_step_start": 600,
+      "control_step_end": 605,
+      "policy_step_start": 100,
+      "policy_step_end": 100,
+      "phrase": "idle above table",
+      "description": "idle above table"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 5.189620758483033,
+    "vlm_failure_onset_timestamp": "00:05.2",
+    "vlm_failure_onset_frame": 260,
+    "vlm_failure_onset_step": 260,
+    "vlm_coarse_failure_onset_seconds": 4.0,
+    "vlm_failure_window_start_seconds": 3.5,
+    "vlm_failure_window_end_seconds": 5.0,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot attempted to grasp the tomato, but failed to secure a firm grip, causing the tomato to slip out of the gripper as soon as the arm lifted.",
+    "vlm_failure_reason_summary": "failed gripper contact; tomato slipping out",
+    "vlm_recovery_action": "Re-align the gripper over the center of the tomato, close the fingers firmly around it to ensure a secure hold, lift, and place it into the bin.",
+    "vlm_recovery_action_summary": "re-align gripper; securely grasp tomato; place in bin",
+    "vlm_prevention_action": "Ensure the gripper is properly centered and positioned lower over the tomato before closing the fingers to achieve a stable grasp prior to lifting.",
+    "vlm_prevention_action_summary": "center gripper over tomato; secure grasp before lifting",
+    "vlm_justification": "At approximately 00:04 (4.0s), the robot closes its gripper on the tomato, but upon lifting at 00:05, the tomato immediately slips out and remains on the tabletop. A second grasp attempt at 00:08 fails similarly, resulting in task failure.",
+    "vlm_temporal_justification": "At clip timestamp 03:30 (corresponding to frame 260 of the original video), the gripper closes around the tomato but fails to secure it, causing the tomato to slip out from between the fingers as they close.",
+    "vlm_video_fps": 50.1,
+    "vlm_video_num_frames": 606,
+    "vlm_video_duration_seconds": 12.075848303393213,
+    "vlm_usage": {
+      "input_tokens": 21578,
+      "output_tokens": 403,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 4.0,\n \"failure_window_start_seconds\": 3.5,\n \"failure_window_end_seconds\": 5.0,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot attempted to grasp the tomato, but failed to secure a firm grip, causing the tomato to slip out of the gripper as soon as the arm lifted.\",\n \"failure_reason_summary\": \"failed gripper contact; tomato slipping out\",\n \"recovery_action\": \"Re-align the gripper over the center of the tomato, close the fingers firmly around it to ensure a secure hold, lift, and place it into the bin.\",\n \"recovery_action_summary\": \"re-align gripper; securely grasp tomato; place in bin\",\n \"prevention_action\": \"Ensure the gripper is properly centered and positioned lower over the tomato before closing the fingers to achieve a stable grasp prior to lifting.\",\n \"prevention_action_summary\": \"center gripper over tomato; secure grasp before lifting\",\n \"justification\": \"At approximately 00:04 (4.0s), the robot closes its gripper on the tomato, but upon lifting at 00:05, the tomato immediately slips out and remains on the tabletop. A second grasp attempt at 00:08 fails similarly, resulting in task failure.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 260,
+    "failure_sim_step": 260,
+    "failure_policy_step": 43,
+    "failure_chunk_index": 2,
+    "first_post_failure_policy_step": 44,
+    "failure_type": "grasp_failure",
+    "correction_action": "Re-align the gripper over the center of the tomato, close the fingers firmly around it to ensure a secure hold, lift, and place it into the bin."
+  }
+}
\ No newline at end of file
diff --git a/tomato/ep046/labels.npz b/tomato/ep046/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..4d71f509616c628f6061a3e276872f2c49849453
--- /dev/null
+++ b/tomato/ep046/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2fbb9f7a72594b54bad347e9f45f0619d47d9db6044cd4a36f7c875b6978069d
+size 3905
diff --git a/tomato/ep047/example.md b/tomato/ep047/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..eab5a4be8f17adf6075004cdb969e62587266bdf
--- /dev/null
+++ b/tomato/ep047/example.md
@@ -0,0 +1,23 @@
+# t003__ep047__tomato
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the tomato in the bin
+- **success:** True
+- **control steps:** 632  (12.6s at 50 Hz)
+- **policy inferences:** 108  (replan_steps=6)
+- **video:** `tomato/ep047/rollout.mp4`
+- **features:** `108` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | reaching for tomato |
+| 3.0–6.0 | 151–301 | 25–50 | grasping tomato |
+| 6.0–9.0 | 302–452 | 50–75 | lifting and moving tomato |
+| 9.0–12.0 | 453–603 | 75–100 | placing tomato in bin |
+| 12.0–12.6 | 604–631 | 100–105 | retracting robot arm |
+
+## Failure
+
+No. Episode succeeded; Dan failure VLM was not run.
diff --git a/tomato/ep047/labels.json b/tomato/ep047/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..74bf7f6e48d4b3ec1b716293ce3de58ad5df86c6
--- /dev/null
+++ b/tomato/ep047/labels.json
@@ -0,0 +1,90 @@
+{
+  "rollout_id": "t003__ep047__tomato",
+  "scene_variant": "",
+  "task_id": 3,
+  "task_file": "",
+  "instruction": "Put the tomato in the bin",
+  "success": true,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T22:09:05+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 3.00796812749004,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching for tomato",
+      "description": "reaching for tomato"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 3.00796812749004,
+      "t_end_sec": 6.01593625498008,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "grasping tomato",
+      "description": "grasping tomato"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 6.01593625498008,
+      "t_end_sec": 9.02390438247012,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "lifting and moving tomato",
+      "description": "lifting and moving tomato"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 9.02390438247012,
+      "t_end_sec": 12.03187250996016,
+      "control_step_start": 453,
+      "control_step_end": 603,
+      "policy_step_start": 75,
+      "policy_step_end": 100,
+      "phrase": "placing tomato in bin",
+      "description": "placing tomato in bin"
+    },
+    {
+      "segment_index": 4,
+      "t_start_sec": 12.03187250996016,
+      "t_end_sec": 12.589641434262948,
+      "control_step_start": 604,
+      "control_step_end": 631,
+      "policy_step_start": 100,
+      "policy_step_end": 105,
+      "phrase": "retracting robot arm",
+      "description": "retracting robot arm"
+    }
+  ],
+  "vlm_failure": {},
+  "failure_annotation": {
+    "failure_control_step": null,
+    "failure_sim_step": null,
+    "failure_policy_step": null,
+    "failure_chunk_index": null,
+    "first_post_failure_policy_step": null,
+    "failure_type": null,
+    "correction_action": null
+  }
+}
\ No newline at end of file
diff --git a/tomato/ep047/labels.npz b/tomato/ep047/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..f24ca9bb8190fcf79cdbb30526a7541f5f3024e0
--- /dev/null
+++ b/tomato/ep047/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:533bacca2b0f8acb6097bb6354822404e808ddb78b07bb6257ad08424d0cb4eb
+size 3139
diff --git a/tomato/ep048/example.md b/tomato/ep048/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..98cd7d5b788edbeae5dad09cc10db400cf4e283d
--- /dev/null
+++ b/tomato/ep048/example.md
@@ -0,0 +1,33 @@
+# t003__ep048__tomato
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the tomato in the bin
+- **success:** False
+- **control steps:** 376  (7.5s at 50 Hz)
+- **policy inferences:** 67  (replan_steps=6)
+- **video:** `tomato/ep048/rollout.mp4`
+- **features:** `67` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–149 | 0–24 | reaching down toward grey bin |
+| 3.0–6.0 | 150–299 | 25–49 | closing gripper inside empty bin |
+| 6.0–7.5 | 300–375 | 50–62 | retracting arm away from bin |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 1.6966067864271457 s (00:01.7)
+- **frame / control_step:** 85 (policy_step=14, chunk_index=1)
+- **coarse onset:** 3.5 s (refined=True)
+- **mode:** wrong_object
+- **reason:** The robot targeted and attempted to grasp inside the grey bin instead of targeting the red tomato on the table.
+  Summary: attempting grasp on bin instead of tomato
+- **recovery:** Move end-effector away from the bin, locate and pick up the red tomato from the table, then place it inside the bin.
+  Summary: re-orient to tomato; grasp tomato; place in bin
+- **prevention:** Identify the red tomato on the table first, descend to grasp it securely, and then move to the bin to place it.
+  Summary: target tomato; grasp tomato; transport to bin
+- **justification:** The robot moved directly toward the grey bin and attempted to close its gripper inside it at around 3.5 seconds, failing to target the red tomato required for the task.
diff --git a/tomato/ep048/labels.json b/tomato/ep048/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..d033e6402a2e5b9336ccce9ff42b7dd542583166
--- /dev/null
+++ b/tomato/ep048/labels.json
@@ -0,0 +1,98 @@
+{
+  "rollout_id": "t003__ep048__tomato",
+  "scene_variant": "",
+  "task_id": 3,
+  "task_file": "",
+  "instruction": "Put the tomato in the bin",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T22:09:36+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.9940119760479043,
+      "control_step_start": 0,
+      "control_step_end": 149,
+      "policy_step_start": 0,
+      "policy_step_end": 24,
+      "phrase": "reaching down toward grey bin",
+      "description": "reaching down toward grey bin"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.9940119760479043,
+      "t_end_sec": 5.9880239520958085,
+      "control_step_start": 150,
+      "control_step_end": 299,
+      "policy_step_start": 25,
+      "policy_step_end": 49,
+      "phrase": "closing gripper inside empty bin",
+      "description": "closing gripper inside empty bin"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.9880239520958085,
+      "t_end_sec": 7.504990019960079,
+      "control_step_start": 300,
+      "control_step_end": 375,
+      "policy_step_start": 50,
+      "policy_step_end": 62,
+      "phrase": "retracting arm away from bin",
+      "description": "retracting arm away from bin"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "wrong_object",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 1.6966067864271457,
+    "vlm_failure_onset_timestamp": "00:01.7",
+    "vlm_failure_onset_frame": 85,
+    "vlm_failure_onset_step": 85,
+    "vlm_coarse_failure_onset_seconds": 3.5,
+    "vlm_failure_window_start_seconds": 3.0,
+    "vlm_failure_window_end_seconds": 4.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot targeted and attempted to grasp inside the grey bin instead of targeting the red tomato on the table.",
+    "vlm_failure_reason_summary": "attempting grasp on bin instead of tomato",
+    "vlm_recovery_action": "Move end-effector away from the bin, locate and pick up the red tomato from the table, then place it inside the bin.",
+    "vlm_recovery_action_summary": "re-orient to tomato; grasp tomato; place in bin",
+    "vlm_prevention_action": "Identify the red tomato on the table first, descend to grasp it securely, and then move to the bin to place it.",
+    "vlm_prevention_action_summary": "target tomato; grasp tomato; transport to bin",
+    "vlm_justification": "The robot moved directly toward the grey bin and attempted to close its gripper inside it at around 3.5 seconds, failing to target the red tomato required for the task.",
+    "vlm_temporal_justification": "At 01:00 (second 60 in the clip), the robot arm begins moving from its resting position and immediately heads towards the grey bin rather than reaching for the red tomato on the table.",
+    "vlm_video_fps": 50.1,
+    "vlm_video_num_frames": 376,
+    "vlm_video_duration_seconds": 7.485029940119761,
+    "vlm_usage": {
+      "input_tokens": 21313,
+      "output_tokens": 352,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"wrong_object\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 3.5,\n \"failure_window_start_seconds\": 3.0,\n \"failure_window_end_seconds\": 4.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot targeted and attempted to grasp inside the grey bin instead of targeting the red tomato on the table.\",\n \"failure_reason_summary\": \"attempting grasp on bin instead of tomato\",\n \"recovery_action\": \"Move end-effector away from the bin, locate and pick up the red tomato from the table, then place it inside the bin.\",\n \"recovery_action_summary\": \"re-orient to tomato; grasp tomato; place in bin\",\n \"prevention_action\": \"Identify the red tomato on the table first, descend to grasp it securely, and then move to the bin to place it.\",\n \"prevention_action_summary\": \"target tomato; grasp tomato; transport to bin\",\n \"justification\": \"The robot moved directly toward the grey bin and attempted to close its gripper inside it at around 3.5 seconds, failing to target the red tomato required for the task.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 85,
+    "failure_sim_step": 85,
+    "failure_policy_step": 14,
+    "failure_chunk_index": 1,
+    "first_post_failure_policy_step": 15,
+    "failure_type": "wrong_object",
+    "correction_action": "Move end-effector away from the bin, locate and pick up the red tomato from the table, then place it inside the bin."
+  }
+}
\ No newline at end of file
diff --git a/tomato/ep048/labels.npz b/tomato/ep048/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..f0e2105979d5c16a8e10d4b4e84d355950025b52
--- /dev/null
+++ b/tomato/ep048/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7d6e0d109d3a0157a2f72224985a7755ecefeb71c54b7680f96ca04609fb4d0b
+size 3690
diff --git a/tomato/ep049/example.md b/tomato/ep049/example.md
new file mode 100644
index 0000000000000000000000000000000000000000..5025de53e22d96ee0deac4b7896d3fb989d16a5b
--- /dev/null
+++ b/tomato/ep049/example.md
@@ -0,0 +1,34 @@
+# t003__ep049__tomato
+
+- **suite / variant:** `` / ``
+- **instruction:** Put the tomato in the bin
+- **success:** False
+- **control steps:** 562  (11.2s at 50 Hz)
+- **policy inferences:** 97  (replan_steps=6)
+- **video:** `tomato/ep049/rollout.mp4`
+- **features:** `97` prefix tensors (`prefix`; model=``)
+
+## Semantic timeline
+
+| seconds | control_step | policy_step | phrase |
+|---|---|---|---|
+| 0.0–3.0 | 0–150 | 0–25 | reaching toward tomato |
+| 3.0–6.0 | 151–301 | 25–50 | closing gripper on tomato |
+| 6.0–9.0 | 302–452 | 50–75 | holding position above tomato |
+| 9.0–11.2 | 453–561 | 75–93 | remaining idle near tomato |
+
+## Failure
+
+- **yes**
+- **onset type:** obvious_mistake
+- **onset:** 3.6507936507936507 s (00:03.7)
+- **frame / control_step:** 184 (policy_step=30, chunk_index=4)
+- **coarse onset:** 4.5 s (refined=True)
+- **mode:** grasp_failure
+- **reason:** The robot aligned over the tomato and closed its gripper, but failed to establish a secure grip, leaving the tomato on the table while the end-effector remained closed over/beside it.
+  Summary: failed grasp attempt; object not secured
+- **recovery:** Reopen the gripper, center the end-effector over the tomato, lower to the correct height, close the gripper to securely grasp the tomato, and transport it to the bin.
+  Summary: reopen gripper; re-align and grasp tomato
+- **prevention:** Ensure precise end-effector positioning and height calibration over the center of the tomato before closing the fingers to guarantee a solid grasp.
+  Summary: align gripper centered over tomato; secure grasp before lifting
+- **justification:** At around 4.5 seconds, the robot's gripper closes onto the tomato but fails to acquire a hold on it. The robot continues to stay over the object without lifting it for the remainder of the video.
diff --git a/tomato/ep049/labels.json b/tomato/ep049/labels.json
new file mode 100644
index 0000000000000000000000000000000000000000..23e1aae35ec0b4d878513d6f23b6e9bfe31880c9
--- /dev/null
+++ b/tomato/ep049/labels.json
@@ -0,0 +1,109 @@
+{
+  "rollout_id": "t003__ep049__tomato",
+  "scene_variant": "",
+  "task_id": 3,
+  "task_file": "",
+  "instruction": "Put the tomato in the bin",
+  "success": false,
+  "model_id": "",
+  "labeler": {
+    "backend": "GeminiBackend",
+    "model": "gemini-3.5-flash",
+    "refine": true,
+    "labeled_at": "2026-10-22T22:10:14+00:00",
+    "pipeline": "scripts/semantic_failure",
+    "prompts": {
+      "keyword_phrases": "You are reviewing the SAME complete robot manipulation rollout.\n\nThe video is {hz:.0f} FPS. Each frame is exactly one executed robot action.\nEpisode duration is {duration:.1f} seconds ({n_control} control steps).\n\nTask instruction:\n\"{instruction}\"\n{failure_context}\nFor EACH numbered 3-second window, write a SHORT KEYWORD PHRASE (2\u20136 words)\nfor what the manipulator is doing, like \"reaching for mug\" or \"closing gripper\"\nor \"idle / no progress\". Not a full sentence. Only what is visible in THAT window.\n\nDo not copy a failure mode, failure reason, recovery, or prevention into these phrases.\nThose are stored in a separate array. If a window happens to show a mistake,\ndescribe the visible motion (e.g. \"dropping can\"), not the taxonomy name.\n\nWindows:\n{windows}\n\nReturn ONLY a JSON object:\n{{\"segments\": [{{\"segment_index\": 0, \"phrase\": \"...\"}}, ...]}}\nInclude every segment_index listed above.\n",
+      "failure_coarse": "You are reviewing the COMPLETE video of a failed robot manipulation rollout in LIBERO-X.\n\nThe video begins at rollout time 0.0 seconds and has a duration of approximately {duration_seconds:.2f} seconds.\n\nTask instruction:\n\"{task_desc}\"\n\nThe simulator reports that this goal predicate was NOT satisfied at the end of the episode:\n{failing_predicate}\n\nAdditional simulator information:\n{detail}\n\nThe simulator predicate tells you WHAT goal condition was unsatisfied at the end. It does NOT tell you when the behavior that caused the failure began.\n\nYour job is to determine:\n\n1. the primary visible failure mode,\n2. when the failure behavior begins,\n3. why the episode failed (a long reason AND a short keyword summary),\n4. a recovery action for after the failure has already happened (long AND short),\n5. a prevention action for before the failure has happened (long AND short).\n\nFAILURE ONSET DEFINITION\n\nUse exactly one of the following onset types:\n\nobvious_mistake:\n Use this when there is an identifiable task-relevant mistake such as\n dropping an object, knocking an object away, manipulating the wrong\n object, making a failed placement, or otherwise performing an action\n that contributes to the eventual failure.\n\n The onset should be the EARLIEST point where the mistake responsible\n for the eventual failure becomes apparent.\n\noperator_intervention:\n Use this when there is not one discrete mistake, but the robot reaches\n a point where a reasonable human operator would decide that the policy\n needs assistance.\n\n Examples include repeatedly attempting an ineffective action, becoming\n stuck, moving erratically, or clearly ceasing to make useful progress.\n\ntimeout:\n Use this when there is no defensible earlier failure point. If the robot\n continues behaving plausibly but simply does not finish before the time\n limit, the failure onset is the END of the rollout.\n\nIMPORTANT RULES\n\n- Do NOT treat the goal predicate being false early in the rollout as failure.\n Most goal predicates are naturally false until the robot completes the task.\n\n- Do NOT label a temporary mistake as failure onset if the robot later\n successfully recovers from that mistake.\n\n- Use the earliest mistake or intervention point that actually explains the\n eventual failed outcome.\n\n- Do not use hindsight to label normal task execution as failure merely because\n you know the episode eventually fails.\n\n- If there is no identifiable earlier point, use onset_type=\"timeout\" and set\n failure_onset_seconds to approximately {duration_seconds:.2f}.\n\nFAILURE TAXONOMY\n\nChoose exactly one:\n\n{taxonomy}\n\nREASON, RECOVERY, PREVENTION \u2014 TWO VERSIONS EACH\n\nEvery language field has a LONG form and a SHORT keyword summary.\n\nLONG: one or two sentences. Specific enough for a human to audit.\n\nSHORT: 2\u20138 words, semicolon-separated clauses, NOT a full sentence.\n Same style as the 3-second action phrases. Good for later alignment.\n Examples: \"collision with object; dropping gripped object\"\n\nfailure_reason / failure_reason_summary:\n Why this episode failed. Example long: the robot hit the basket while\n carrying the moka pot, dropped it, and knocked the basket over.\n Example short: \"collision with object; dropping gripped object\"\n\nRECOVERY (failure HAS already happened):\n What the robot should do at or immediately after onset, now that the\n scene is already broken. Identify object, action, and the condition to\n restore before continuing.\n Example long: move the basket away from the hot plate, then place the\n moka pot on the hot-plate.\n Example short: \"clear basket; place moka on stove\"\n\nPREVENTION (failure has NOT yet happened):\n The counterfactual at or before onset: what the robot should have done\n so the mistake never occurs. Do not describe post-drop cleanup.\n Example long: place the moka pot on the stove; do not approach the box.\n Example short: \"place moka on stove; avoid box\"\n\nDo not copy taxonomy names into these strings.\n\nReturn ONLY a JSON object with this structure:\n\n{{\n \"failure_mode\": \" \",\n \"onset_type\": \"obvious_mistake|operator_intervention|timeout\",\n \"failure_onset_seconds\":,\n \"failure_window_start_seconds\":,\n \"failure_window_end_seconds\":,\n \"confidence\": \"high|medium|low\",\n \"failure_reason\": \" \",\n \"failure_reason_summary\": \" \",\n \"recovery_action\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \",\n \"justification\": \" \"\n}}\n",
+      "failure_refine": "You are reviewing a TEMPORALLY MAGNIFIED clip from a failed robot rollout.\n\nThe complete rollout was already reviewed by another pass of the same VLM.\n\nThe coarse analysis identified:\n\nFailure mode:\n{failure_mode}\n\nFailure onset type:\n{onset_type}\n\nFailure reason:\n{failure_reason}\n\nCoarse onset estimate in the original rollout:\n{coarse_seconds:.2f} seconds\n\nTask instruction:\n\"{task_desc}\"\n\nFailed simulator predicate:\n{failing_predicate}\n\nTEMPORAL MAPPING\n\nThis clip has been deliberately slowed down.\n\nEach SECOND of this clip corresponds to exactly ONE FRAME from the original\nrollout.\n\nTherefore:\n\n refined clip second 0 = original rollout frame {start_frame}\n refined clip second 1 = original rollout frame {start_frame_plus_one}\n refined clip second 2 = original rollout frame {start_frame_plus_two}\n ...\n\nThe clip contains {num_frames} original rollout frames.\n\nYour job is ONLY to refine the temporal location of the failure.\n\nFind the EARLIEST frame in this clip where the failure identified above becomes\nvisibly apparent or where operator intervention becomes justified.\n\nRemember:\n\n- Do not select a temporary mistake if the robot recovers from it.\n- Do not select a frame merely because the task is not complete yet.\n- Select the first frame associated with the behavior responsible for the\n eventual failure.\n- If the coarse event is not actually visible in this clip, return null.\n\nReturn ONLY:\n\n{{\n \"refined_second\":,\n \"confidence\": \"high|medium|low\",\n \"justification\": \" \"\n}}\n",
+      "failure_language": "You already analyzed this failed robot rollout. The onset, mode, and taxonomy are FIXED. Do not change them.\n\nTask instruction:\n\"{task_desc}\"\n\nExisting labels:\n- failure_mode: {failure_mode}\n- onset_type: {onset_type}\n- failure_onset_seconds: {onset_seconds}\n- failure_reason: {failure_reason}\n- recovery_action: {recovery_action}\n\nFill the missing language fields.\n\nSHORT STYLE\n2\u20138 words, semicolon-separated clauses, not a full sentence.\nExample: \"collision with object; dropping gripped object\"\n\nRECOVERY vs PREVENTION\n- recovery_action_summary: keyword form of the EXISTING recovery (failure already happened).\n- prevention_action: long, what the robot should have done at or before onset so the failure never happens.\n- prevention_action_summary: keyword form of that prevention.\n\nReturn ONLY:\n\n{{\n \"failure_reason_summary\": \" \",\n \"recovery_action_summary\": \" \",\n \"prevention_action\": \" \",\n \"prevention_action_summary\": \" \"\n}}\n",
+      "failure_taxonomy": "- wrong_object: the robot acts on (grasps, pushes, presses, turns, opens, or closes) an object or control other than the one the task specifies, for example picking up the wrong item, pressing the wrong button, or opening the wrong drawer. Use this whenever the robot targets the wrong entity, even if the action itself is also unsuccessful.\n- wrong_target: the robot manipulates the correct object but takes it to the wrong destination, location, or receptacle.\n- press_failure: the robot targets the correct button or knob but does not achieve the required result, for example pressing without triggering it, not pressing far enough, or not turning the knob to the required position. Does not apply if the wrong button or knob was targeted (use wrong_object).\n- open_close_failure: the robot targets the correct container (drawer, door, lid, box, etc.) but does not achieve the required open or closed state, for example a partial opening, a lid left ajar, or a door handle or drawer handle that is grabbed but the door or drawer does not move.\n- grasp_failure: the robot attempts to grasp the correct object but never acquires it securely, including repeated unsuccessful grasp attempts. Does not apply if the object was held and then lost (use object_displacement).\n- object_displacement: the robot unintentionally knocks over, pushes away, or drops an object, including dropping an object it had successfully grasped, in a way that causes task failure.\n- placement_or_insertion_failure: the robot holds the correct object and brings it to the correct target, but the final placement, insertion, position, or orientation is incorrect (e.g. misaligned peg, object resting off-target, tilted or upside down).\n- stuck_or_no_progress: the robot shows visibly unproductive behavior, such as looping, oscillating, repeating the same ineffective motion, hovering without acting, or freezing. The robot is active or idle in a way that clearly is not advancing the task. Does not apply if the robot is moving purposefully but slowly.\n- timeout_or_insufficient_progress: the robot moves purposefully toward completing the task and makes no identifiable mistake, but is too slow to finish before the episode ends. Use only when no other label applies.\n- other: the failure fits none of the labels above. Briefly explain what went wrong.\n"
+    }
+  },
+  "semantic_timeline": [
+    {
+      "segment_index": 0,
+      "t_start_sec": 0.0,
+      "t_end_sec": 2.996031746031746,
+      "control_step_start": 0,
+      "control_step_end": 150,
+      "policy_step_start": 0,
+      "policy_step_end": 25,
+      "phrase": "reaching toward tomato",
+      "description": "reaching toward tomato"
+    },
+    {
+      "segment_index": 1,
+      "t_start_sec": 2.996031746031746,
+      "t_end_sec": 5.992063492063492,
+      "control_step_start": 151,
+      "control_step_end": 301,
+      "policy_step_start": 25,
+      "policy_step_end": 50,
+      "phrase": "closing gripper on tomato",
+      "description": "closing gripper on tomato"
+    },
+    {
+      "segment_index": 2,
+      "t_start_sec": 5.992063492063492,
+      "t_end_sec": 8.988095238095239,
+      "control_step_start": 302,
+      "control_step_end": 452,
+      "policy_step_start": 50,
+      "policy_step_end": 75,
+      "phrase": "holding position above tomato",
+      "description": "holding position above tomato"
+    },
+    {
+      "segment_index": 3,
+      "t_start_sec": 8.988095238095239,
+      "t_end_sec": 11.15079365079365,
+      "control_step_start": 453,
+      "control_step_end": 561,
+      "policy_step_start": 75,
+      "policy_step_end": 93,
+      "phrase": "remaining idle near tomato",
+      "description": "remaining idle near tomato"
+    }
+  ],
+  "vlm_failure": {
+    "vlm_failure_mode": "grasp_failure",
+    "vlm_failure_onset_type": "obvious_mistake",
+    "vlm_failure_onset_seconds": 3.6507936507936507,
+    "vlm_failure_onset_timestamp": "00:03.7",
+    "vlm_failure_onset_frame": 184,
+    "vlm_failure_onset_step": 184,
+    "vlm_coarse_failure_onset_seconds": 4.5,
+    "vlm_failure_window_start_seconds": 3.5,
+    "vlm_failure_window_end_seconds": 5.5,
+    "vlm_temporal_refined": true,
+    "vlm_confidence": "high",
+    "vlm_temporal_confidence": "high",
+    "vlm_failure_reason": "The robot aligned over the tomato and closed its gripper, but failed to establish a secure grip, leaving the tomato on the table while the end-effector remained closed over/beside it.",
+    "vlm_failure_reason_summary": "failed grasp attempt; object not secured",
+    "vlm_recovery_action": "Reopen the gripper, center the end-effector over the tomato, lower to the correct height, close the gripper to securely grasp the tomato, and transport it to the bin.",
+    "vlm_recovery_action_summary": "reopen gripper; re-align and grasp tomato",
+    "vlm_prevention_action": "Ensure precise end-effector positioning and height calibration over the center of the tomato before closing the fingers to guarantee a solid grasp.",
+    "vlm_prevention_action_summary": "align gripper centered over tomato; secure grasp before lifting",
+    "vlm_justification": "At around 4.5 seconds, the robot's gripper closes onto the tomato but fails to acquire a hold on it. The robot continues to stay over the object without lifting it for the remainder of the video.",
+    "vlm_temporal_justification": "At 01:48 (108 seconds into the clip), the robot gripper finishes closing completely without securing the tomato, leaving it sitting ungrasped on the table surface.",
+    "vlm_video_fps": 50.4,
+    "vlm_video_num_frames": 562,
+    "vlm_video_duration_seconds": 11.130952380952381,
+    "vlm_usage": {
+      "input_tokens": 21647,
+      "output_tokens": 381,
+      "cached_tokens": 0
+    },
+    "vlm_raw_response": "```json\n{\n \"failure_mode\": \"grasp_failure\",\n \"onset_type\": \"obvious_mistake\",\n \"failure_onset_seconds\": 4.5,\n \"failure_window_start_seconds\": 3.5,\n \"failure_window_end_seconds\": 5.5,\n \"confidence\": \"high\",\n \"failure_reason\": \"The robot aligned over the tomato and closed its gripper, but failed to establish a secure grip, leaving the tomato on the table while the end-effector remained closed over/beside it.\",\n \"failure_reason_summary\": \"failed grasp attempt; object not secured\",\n \"recovery_action\": \"Reopen the gripper, center the end-effector over the tomato, lower to the correct height, close the gripper to securely grasp the tomato, and transport it to the bin.\",\n \"recovery_action_summary\": \"reopen gripper; re-align and grasp tomato\",\n \"prevention_action\": \"Ensure precise end-effector positioning and height calibration over the center of the tomato before closing the fingers to guarantee a solid grasp.\",\n \"prevention_action_summary\": \"align gripper centered over tomato; secure grasp before lifting\",\n \"justification\": \"At around 4.5 seconds, the robot's gripper closes onto the tomato but fails to acquire a hold on it. The robot continues to stay over the object without lifting it for the remainder of the video.\"\n}\n```"
+  },
+  "failure_annotation": {
+    "failure_control_step": 184,
+    "failure_sim_step": 184,
+    "failure_policy_step": 30,
+    "failure_chunk_index": 4,
+    "first_post_failure_policy_step": 31,
+    "failure_type": "grasp_failure",
+    "correction_action": "Reopen the gripper, center the end-effector over the tomato, lower to the correct height, close the gripper to securely grasp the tomato, and transport it to the bin."
+  }
+}
\ No newline at end of file
diff --git a/tomato/ep049/labels.npz b/tomato/ep049/labels.npz
new file mode 100644
index 0000000000000000000000000000000000000000..f223116991116944288cf18c1caeeed536de1d91
--- /dev/null
+++ b/tomato/ep049/labels.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6d5d01eec8a786fc27b4e57b092ea7ee5889009273af0cfd1e5e9a9f0067c068
+size 3883