AutoBio masters: 9 task videos (6 agent successes, 3 reference solutions) + teaser + highlights

masters/autobio/highlights.json

@@ -0,0 +1,11 @@
+{
+ "teaser": "screw_tighten",
+ "highlights": [
+  "screw_tighten",
+  "insert_centrifuge_5430",
+  "thermal_mixer",
+  "pipette",
+  "screw_loose"
+ ],
+ "why_fewer": ""
+}

masters/autobio/insert/meta.json

@@ -0,0 +1,74 @@
+{
+ "suite": "autobio",
+ "task": "insert",
+ "task_dir": "autobio-insert-i00-privileged",
+ "source": "oracle",
+ "agent": "reference solution (oracle)",
+ "run": null,
+ "mode": "privileged",
+ "trial": null,
+ "trial_choice": "no counted success by any agent other than GPT-6 Luna, whose trials are never shown: the task's reference solution (AutoBio's scripted expert, solution/oracle.npz)",
+ "needs_owner_ok": true,
+ "renderer": "Blender 4.2.23 LTS Cycles path tracing (GPU OptiX), adaptive sampling (threshold 0.01, 64-128 spp), OpenImageDenoise with albedo + normal passes; kinematic playback of the exact MuJoCo replay poses (MuJoCo's own compiled meshes, primitives, colours and UV textures; AutoBio's liquid-surface disc and the thermal mixer's display drawn as AutoBio's own renderer draws them); seamless studio floor (cyclorama), one large soft key area light, a weak overhead softbox, a rim light, a dim sky dome (Poly Haven studio_small_09 HDRI, CC0); AgX view transform, Medium High Contrast look; fixed 3/4 camera with a slow eased 10-degree orbit; native 2560x1440",
+ "rendering": "path-tracing",
+ "media_standard": "v2 (2026-10-16)",
+ "width": 2560,
+ "height": 1440,
+ "fps": 30,
+ "frames": 311,
+ "seconds": 10.37,
+ "speed": 1.0,
+ "success_check": true,
+ "success_check_how": "the task's reference solution (solution/oracle.npz) replayed from the frozen instance with the task's own tooling (rcb_autobio.Session, as tests/grade.py replays it): success by AutoBio's check; a second plain replay (Session.replay) ends in the same end-state hash",
+ "end_check": {
+  "success": true,
+  "rule": "first",
+  "reason": "success",
+  "score": 1.0,
+  "sim_time_s": 8.324
+ },
+ "replay": {
+  "control_hz": 500.0,
+  "n_actions": 4162,
+  "control_steps_replayed": 4162,
+  "sim_time_s": 8.324,
+  "frames_every_control_steps": 16.6667,
+  "final_state_sha": "6bd9f204cf6c8786",
+  "plain_replay_state_sha": "6bd9f204cf6c8786",
+  "replays_agree": true,
+  "model_sha": "3be600fd69c42e20",
+  "init_sha": "26cdc6033fab7de6",
+  "instance_ok": true,
+  "state_hashes_every_control_steps": 1000,
+  "trajectory_sha256": "fdd886ade4db180f1872f91ebfc174fb05e7db9ca39beaf274540b4a5c91af30",
+  "mujoco": "3.3.0"
+ },
+ "visual_changes": [
+  "studio floor and cove instead of MuJoCo's checker ground and skybox"
+ ],
+ "camera": {
+  "mode": "fixed",
+  "frames_on": "the instruments and objects of the task and the arms' end effectors",
+  "az_deg": -35.0,
+  "el_deg": 30.0,
+  "fovy_deg": 34.0,
+  "fill": 0.78,
+  "target": [
+   0.316,
+   -0.0168,
+   1.0079
+  ],
+  "distance_m": 1.126,
+  "orbit_deg": 10.0
+ },
+ "holds": {
+  "start_s": 0.5,
+  "end_s": 1.5,
+  "note": "first / last frame held, no extra simulation"
+ },
+ "image": "efe-private:autobio-0.1.0",
+ "encode": "libx264 preset slow CRF 16, yuv420p, bt709, faststart",
+ "poster": "last frame (solved state)",
+ "seconds_per_frame": 3.98,
+ "created": "2026-10-23T01:08:02-0500"
+}

masters/autobio/insert/video.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b2ea5b1329ccb1d095605db92bc50c004ad8f8f5a9a368c196d26c9cd72eb87f
+size 5790582

masters/autobio/insert_centrifuge_5430/meta.json

@@ -0,0 +1,76 @@
+{
+ "suite": "autobio",
+ "task": "insert_centrifuge_5430",
+ "task_dir": "autobio-insert-centrifuge-5430-i00-standard",
+ "source": "agent-success:claude_code-opus_5_5-high-autobio-insert-centrifuge-5430-i00-standard/autobio-insert-centrifuge-5430-i__ZkE8x7J",
+ "agent": "claude code + Claude Opus 5.5 (high)",
+ "run": "claude_code-opus_5_5-high",
+ "mode": "standard",
+ "trial": "claude_code-opus_5_5-high-autobio-insert-centrifuge-5430-i00-standard/autobio-insert-centrifuge-5430-i__ZkE8x7J",
+ "trial_choice": "a counted success, never GPT-6 Luna's: privileged before standard, GPT-6 Astra first, then Claude Opus 5.5, Gemini 3.8 Flash, DeepSeek, GLM",
+ "renderer": "Blender 4.2.23 LTS Cycles path tracing (GPU OptiX), adaptive sampling (threshold 0.01, 64-128 spp), OpenImageDenoise with albedo + normal passes; kinematic playback of the exact MuJoCo replay poses (MuJoCo's own compiled meshes, primitives, colours and UV textures; AutoBio's liquid-surface disc and the thermal mixer's display drawn as AutoBio's own renderer draws them); seamless studio floor (cyclorama), one large soft key area light, a weak overhead softbox, a rim light, a dim sky dome (Poly Haven studio_small_09 HDRI, CC0); AgX view transform, Medium High Contrast look; fixed 3/4 camera with a slow eased 10-degree orbit; native 2560x1440",
+ "rendering": "path-tracing",
+ "media_standard": "v2 (2026-10-16)",
+ "width": 2560,
+ "height": 1440,
+ "fps": 30,
+ "frames": 762,
+ "seconds": 25.4,
+ "speed": 12.0,
+ "success_check": true,
+ "success_check_how": "standard-mode trial (graded by its live episode): re-rendered only because the replay is bit-exact. The recorded rows were replayed from the task's frozen instance with the task's own tooling (rcb_autobio.Session, as the sim service runs them); the replay ends in AutoBio's success check with an end-state hash equal to the live episode's recorded hash (8c3894a0c03d4075), and a second plain replay (Session.replay) ends in the same hash",
+ "end_check": {
+  "success": true,
+  "rule": "first",
+  "reason": "success",
+  "score": 1.0,
+  "sim_time_s": 280.286
+ },
+ "replay": {
+  "control_hz": 500.0,
+  "n_actions": 140143,
+  "control_steps_replayed": 140143,
+  "sim_time_s": 280.286,
+  "frames_every_control_steps": 200.0,
+  "final_state_sha": "8c3894a0c03d4075",
+  "plain_replay_state_sha": "8c3894a0c03d4075",
+  "replays_agree": true,
+  "model_sha": "a45ad43e4c568577",
+  "init_sha": "843c5cc8cbffee91",
+  "instance_ok": true,
+  "state_hashes_every_control_steps": 1000,
+  "trajectory_sha256": "9c9fcd83354209badc50a6350ec8b38c9b1b18455240381fc827b08f864f3e6b",
+  "mujoco": "3.3.0",
+  "recorded_live_state_sha": "8c3894a0c03d4075",
+  "matches_live_episode": true
+ },
+ "visual_changes": [
+  "studio floor and cove instead of MuJoCo's checker ground and skybox"
+ ],
+ "camera": {
+  "mode": "fixed",
+  "frames_on": "the instruments and objects of the task and the arms' end effectors",
+  "ee_percentile_kept_in_frame": 97.0,
+  "az_deg": -35.0,
+  "el_deg": 30.0,
+  "fovy_deg": 34.0,
+  "fill": 0.78,
+  "target": [
+   0.0666,
+   0.1797,
+   1.1018
+  ],
+  "distance_m": 1.7024,
+  "orbit_deg": 10.0
+ },
+ "holds": {
+  "start_s": 0.5,
+  "end_s": 1.5,
+  "note": "first / last frame held, no extra simulation"
+ },
+ "image": "efe-private:autobio-0.1.0",
+ "encode": "libx264 preset slow CRF 16, yuv420p, bt709, faststart",
+ "poster": "last frame (solved state)",
+ "seconds_per_frame": 6.06,
+ "created": "2026-10-23T02:02:20-0500"
+}

masters/autobio/insert_centrifuge_5430/video.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0ac2f4289b1a10d9bf21bd23941f2f38d828011b8c1baa37817b98e05ada0d8f
+size 15373584

masters/autobio/pickup/meta.json

@@ -0,0 +1,74 @@
+{
+ "suite": "autobio",
+ "task": "pickup",
+ "task_dir": "autobio-pickup-i00-privileged",
+ "source": "oracle",
+ "agent": "reference solution (oracle)",
+ "run": null,
+ "mode": "privileged",
+ "trial": null,
+ "trial_choice": "no counted success by any agent other than GPT-6 Luna, whose trials are never shown: the task's reference solution (AutoBio's scripted expert, solution/oracle.npz)",
+ "needs_owner_ok": true,
+ "renderer": "Blender 4.2.23 LTS Cycles path tracing (GPU OptiX), adaptive sampling (threshold 0.01, 64-128 spp), OpenImageDenoise with albedo + normal passes; kinematic playback of the exact MuJoCo replay poses (MuJoCo's own compiled meshes, primitives, colours and UV textures; AutoBio's liquid-surface disc and the thermal mixer's display drawn as AutoBio's own renderer draws them); seamless studio floor (cyclorama), one large soft key area light, a weak overhead softbox, a rim light, a dim sky dome (Poly Haven studio_small_09 HDRI, CC0); AgX view transform, Medium High Contrast look; fixed 3/4 camera with a slow eased 10-degree orbit; native 2560x1440",
+ "rendering": "path-tracing",
+ "media_standard": "v2 (2026-10-16)",
+ "width": 2560,
+ "height": 1440,
+ "fps": 30,
+ "frames": 298,
+ "seconds": 9.93,
+ "speed": 1.0,
+ "success_check": true,
+ "success_check_how": "the task's reference solution (solution/oracle.npz) replayed from the frozen instance with the task's own tooling (rcb_autobio.Session, as tests/grade.py replays it): success by AutoBio's check; a second plain replay (Session.replay) ends in the same end-state hash",
+ "end_check": {
+  "success": true,
+  "rule": "first",
+  "reason": "success",
+  "score": 1.0,
+  "sim_time_s": 7.89
+ },
+ "replay": {
+  "control_hz": 500.0,
+  "n_actions": 3945,
+  "control_steps_replayed": 3945,
+  "sim_time_s": 7.89,
+  "frames_every_control_steps": 16.6667,
+  "final_state_sha": "e167e764260a6fe6",
+  "plain_replay_state_sha": "e167e764260a6fe6",
+  "replays_agree": true,
+  "model_sha": "d3406a0e2ed85218",
+  "init_sha": "42f230966a7b71a8",
+  "instance_ok": true,
+  "state_hashes_every_control_steps": 1000,
+  "trajectory_sha256": "8644bd51893125bee7c398c8907b2ed4a5653c4f1db09ce9c5e555e85a65ded2",
+  "mujoco": "3.3.0"
+ },
+ "visual_changes": [
+  "studio floor and cove instead of MuJoCo's checker ground and skybox"
+ ],
+ "camera": {
+  "mode": "fixed",
+  "frames_on": "the instruments and objects of the task and the arms' end effectors",
+  "az_deg": 35.0,
+  "el_deg": 30.0,
+  "fovy_deg": 34.0,
+  "fill": 0.78,
+  "target": [
+   0.2673,
+   0.0206,
+   0.9935
+  ],
+  "distance_m": 0.8529,
+  "orbit_deg": 10.0
+ },
+ "holds": {
+  "start_s": 0.5,
+  "end_s": 1.5,
+  "note": "first / last frame held, no extra simulation"
+ },
+ "image": "efe-private:autobio-0.1.0",
+ "encode": "libx264 preset slow CRF 16, yuv420p, bt709, faststart",
+ "poster": "last frame (solved state)",
+ "seconds_per_frame": 4.0,
+ "created": "2026-10-23T00:50:44-0500"
+}

masters/autobio/pickup/video.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0c262bc4b9785dd9699104d5c90dfc979dd6452e74dd28a4209ee5f614dabc4b
+size 5554000

masters/autobio/pipette/meta.json

@@ -0,0 +1,76 @@
+{
+ "suite": "autobio",
+ "task": "pipette",
+ "task_dir": "autobio-pipette-i00-standard",
+ "source": "agent-success:codex-gpt6_astra-high-autobio-pipette-i00-standard/autobio-pipette-i00-standard__cx5TLkL",
+ "agent": "codex + GPT-6 Astra (high)",
+ "run": "codex-gpt6_astra-high",
+ "mode": "standard",
+ "trial": "codex-gpt6_astra-high-autobio-pipette-i00-standard/autobio-pipette-i00-standard__cx5TLkL",
+ "trial_choice": "a counted success, never GPT-6 Luna's: privileged before standard, GPT-6 Astra first, then Claude Opus 5.5, Gemini 3.8 Flash, DeepSeek, GLM",
+ "renderer": "Blender 4.2.23 LTS Cycles path tracing (GPU OptiX), adaptive sampling (threshold 0.01, 64-128 spp), OpenImageDenoise with albedo + normal passes; kinematic playback of the exact MuJoCo replay poses (MuJoCo's own compiled meshes, primitives, colours and UV textures; AutoBio's liquid-surface disc and the thermal mixer's display drawn as AutoBio's own renderer draws them); seamless studio floor (cyclorama), one large soft key area light, a weak overhead softbox, a rim light, a dim sky dome (Poly Haven studio_small_09 HDRI, CC0); AgX view transform, Medium High Contrast look; fixed 3/4 camera with a slow eased 10-degree orbit; native 2560x1440",
+ "rendering": "path-tracing",
+ "media_standard": "v2 (2026-10-16)",
+ "width": 2560,
+ "height": 1440,
+ "fps": 30,
+ "frames": 679,
+ "seconds": 22.63,
+ "speed": 6.0,
+ "success_check": true,
+ "success_check_how": "standard-mode trial (graded by its live episode): re-rendered only because the replay is bit-exact. The recorded rows were replayed from the task's frozen instance with the task's own tooling (rcb_autobio.Session, as the sim service runs them); the replay ends in AutoBio's success check with an end-state hash equal to the live episode's recorded hash (921980ff5d1447ef), and a second plain replay (Session.replay) ends in the same hash",
+ "end_check": {
+  "success": true,
+  "rule": "first",
+  "reason": "success",
+  "score": 1.0,
+  "sim_time_s": 123.572
+ },
+ "replay": {
+  "control_hz": 500.0,
+  "n_actions": 61786,
+  "control_steps_replayed": 61786,
+  "sim_time_s": 123.572,
+  "frames_every_control_steps": 100.0,
+  "final_state_sha": "921980ff5d1447ef",
+  "plain_replay_state_sha": "921980ff5d1447ef",
+  "replays_agree": true,
+  "model_sha": "40806bc5f581ca3e",
+  "init_sha": "228ededc6af93404",
+  "instance_ok": true,
+  "state_hashes_every_control_steps": 1000,
+  "trajectory_sha256": "444169325d9c62feab697881c638a6202458201f50479e2e939e7809e8f0b2ea",
+  "mujoco": "3.3.0",
+  "recorded_live_state_sha": "921980ff5d1447ef",
+  "matches_live_episode": true
+ },
+ "visual_changes": [
+  "studio floor and cove instead of MuJoCo's checker ground and skybox",
+  "AutoBio's liquid surface (a disc) drawn as tinted water"
+ ],
+ "camera": {
+  "mode": "fixed",
+  "frames_on": "the instruments and objects of the task and the arms' end effectors",
+  "az_deg": 35.0,
+  "el_deg": 30.0,
+  "fovy_deg": 34.0,
+  "fill": 0.9,
+  "target": [
+   0.0175,
+   -0.1287,
+   1.1347
+  ],
+  "distance_m": 1.9599,
+  "orbit_deg": 10.0
+ },
+ "holds": {
+  "start_s": 0.5,
+  "end_s": 1.5,
+  "note": "first / last frame held, no extra simulation"
+ },
+ "image": "efe-private:autobio-0.1.0",
+ "encode": "libx264 preset slow CRF 16, yuv420p, bt709, faststart",
+ "poster": "last frame (solved state)",
+ "seconds_per_frame": 6.7,
+ "created": "2026-10-23T02:01:45-0500"
+}

masters/autobio/pipette/video.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f675be0a769ba96eb44936a774a3669bd8f99a2156a08a406867b0726911f798
+size 12397547

masters/autobio/screw_loose/meta.json

@@ -0,0 +1,74 @@
+{
+ "suite": "autobio",
+ "task": "screw_loose",
+ "task_dir": "autobio-screw-loose-i00-privileged",
+ "source": "agent-success:codex-gpt6_astra-high-autobio-screw-loose-i00-privileged/autobio-screw-loose-i00-privileg__CBd4zLh",
+ "agent": "codex + GPT-6 Astra (high)",
+ "run": "codex-gpt6_astra-high",
+ "mode": "privileged",
+ "trial": "codex-gpt6_astra-high-autobio-screw-loose-i00-privileged/autobio-screw-loose-i00-privileg__CBd4zLh",
+ "trial_choice": "a counted success, never GPT-6 Luna's: privileged before standard, GPT-6 Astra first, then Claude Opus 5.5, Gemini 3.8 Flash, DeepSeek, GLM",
+ "renderer": "Blender 4.2.23 LTS Cycles path tracing (GPU OptiX), adaptive sampling (threshold 0.01, 64-128 spp), OpenImageDenoise with albedo + normal passes; kinematic playback of the exact MuJoCo replay poses (MuJoCo's own compiled meshes, primitives, colours and UV textures; AutoBio's liquid-surface disc and the thermal mixer's display drawn as AutoBio's own renderer draws them); seamless studio floor (cyclorama), one large soft key area light, a weak overhead softbox, a rim light, a dim sky dome (Poly Haven studio_small_09 HDRI, CC0); AgX view transform, Medium High Contrast look; fixed 3/4 camera with a slow eased 10-degree orbit; native 2560x1440",
+ "rendering": "path-tracing",
+ "media_standard": "v2 (2026-10-16)",
+ "width": 2560,
+ "height": 1440,
+ "fps": 30,
+ "frames": 197,
+ "seconds": 6.57,
+ "speed": 1.0,
+ "success_check": true,
+ "success_check_how": "privileged trial: the recorded rows replayed from the task's frozen instance with the task's own tooling (rcb_autobio.Session, as tests/grade.py replays them); success by AutoBio's check, the end-state hash equals both grading replays' hash (5f61815939bcdd8c), and a second plain replay (Session.replay) ends in the same hash",
+ "end_check": {
+  "success": true,
+  "rule": "first",
+  "reason": "success",
+  "score": 1.0,
+  "sim_time_s": 4.532
+ },
+ "replay": {
+  "control_hz": 500.0,
+  "n_actions": 2266,
+  "control_steps_replayed": 2266,
+  "sim_time_s": 4.532,
+  "frames_every_control_steps": 16.6667,
+  "final_state_sha": "5f61815939bcdd8c",
+  "plain_replay_state_sha": "5f61815939bcdd8c",
+  "replays_agree": true,
+  "model_sha": "73c3c34b2e6aac62",
+  "init_sha": "5334d17d50180476",
+  "instance_ok": true,
+  "state_hashes_every_control_steps": 1000,
+  "trajectory_sha256": "7dee866956730b97f6d64369580afbbc3eb2983932c98b70652b069f45f50bc0",
+  "mujoco": "3.3.0",
+  "recorded_grader_state_sha": "5f61815939bcdd8c"
+ },
+ "visual_changes": [
+  "studio floor and cove instead of MuJoCo's checker ground and skybox"
+ ],
+ "camera": {
+  "mode": "fixed",
+  "frames_on": "the instruments and objects of the task and the arms' end effectors",
+  "az_deg": 35.0,
+  "el_deg": 30.0,
+  "fovy_deg": 34.0,
+  "fill": 0.78,
+  "target": [
+   -0.0824,
+   0.0902,
+   0.9592
+  ],
+  "distance_m": 1.18,
+  "orbit_deg": 10.0
+ },
+ "holds": {
+  "start_s": 0.5,
+  "end_s": 1.5,
+  "note": "first / last frame held, no extra simulation"
+ },
+ "image": "efe-private:autobio-0.1.0",
+ "encode": "libx264 preset slow CRF 16, yuv420p, bt709, faststart",
+ "poster": "last frame (solved state)",
+ "seconds_per_frame": 3.94,
+ "created": "2026-10-23T01:36:36-0500"
+}

masters/autobio/screw_loose/video.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:09154d27949ef23d96fb5bbfbd856cc055a607b046538692c905d264e34349f4
+size 4462319

masters/autobio/screw_tighten/meta.json

@@ -0,0 +1,75 @@
+{
+ "suite": "autobio",
+ "task": "screw_tighten",
+ "task_dir": "autobio-screw-tighten-i00-standard",
+ "source": "agent-success:codex-gpt6_astra-high-autobio-screw-tighten-i00-standard/autobio-screw-tighten-i00-standa__ky4Chis",
+ "agent": "codex + GPT-6 Astra (high)",
+ "run": "codex-gpt6_astra-high",
+ "mode": "standard",
+ "trial": "codex-gpt6_astra-high-autobio-screw-tighten-i00-standard/autobio-screw-tighten-i00-standa__ky4Chis",
+ "trial_choice": "a counted success, never GPT-6 Luna's: privileged before standard, GPT-6 Astra first, then Claude Opus 5.5, Gemini 3.8 Flash, DeepSeek, GLM",
+ "renderer": "Blender 4.2.23 LTS Cycles path tracing (GPU OptiX), adaptive sampling (threshold 0.01, 64-128 spp), OpenImageDenoise with albedo + normal passes; kinematic playback of the exact MuJoCo replay poses (MuJoCo's own compiled meshes, primitives, colours and UV textures; AutoBio's liquid-surface disc and the thermal mixer's display drawn as AutoBio's own renderer draws them); seamless studio floor (cyclorama), one large soft key area light, a weak overhead softbox, a rim light, a dim sky dome (Poly Haven studio_small_09 HDRI, CC0); AgX view transform, Medium High Contrast look; fixed 3/4 camera with a slow eased 10-degree orbit; native 2560x1440",
+ "rendering": "path-tracing",
+ "media_standard": "v2 (2026-10-16)",
+ "width": 2560,
+ "height": 1440,
+ "fps": 30,
+ "frames": 658,
+ "seconds": 21.93,
+ "speed": 20.0,
+ "success_check": true,
+ "success_check_how": "standard-mode trial (graded by its live episode): re-rendered only because the replay is bit-exact. The recorded rows were replayed from the task's frozen instance with the task's own tooling (rcb_autobio.Session, as the sim service runs them); the replay ends in AutoBio's success check with an end-state hash equal to the live episode's recorded hash (6cd6e7f8df565581), and a second plain replay (Session.replay) ends in the same hash",
+ "end_check": {
+  "success": true,
+  "rule": "first",
+  "reason": "success",
+  "score": 1.0,
+  "sim_time_s": 397.736
+ },
+ "replay": {
+  "control_hz": 500.0,
+  "n_actions": 198868,
+  "control_steps_replayed": 198868,
+  "sim_time_s": 397.736,
+  "frames_every_control_steps": 333.3333,
+  "final_state_sha": "6cd6e7f8df565581",
+  "plain_replay_state_sha": "6cd6e7f8df565581",
+  "replays_agree": true,
+  "model_sha": "f168bb2de7a282c0",
+  "init_sha": "2667e5f48d914481",
+  "instance_ok": true,
+  "state_hashes_every_control_steps": 1000,
+  "trajectory_sha256": "05ee7882d9bb19bbe528f7863136eed8a5c59740bab7938f5c33b9ceafff3c0e",
+  "mujoco": "3.3.0",
+  "recorded_live_state_sha": "6cd6e7f8df565581",
+  "matches_live_episode": true
+ },
+ "visual_changes": [
+  "studio floor and cove instead of MuJoCo's checker ground and skybox"
+ ],
+ "camera": {
+  "mode": "fixed",
+  "frames_on": "the instruments and objects of the task and the arms' end effectors",
+  "az_deg": 35.0,
+  "el_deg": 30.0,
+  "fovy_deg": 34.0,
+  "fill": 0.78,
+  "target": [
+   -0.0711,
+   0.0363,
+   0.9587
+  ],
+  "distance_m": 1.3831,
+  "orbit_deg": 10.0
+ },
+ "holds": {
+  "start_s": 0.5,
+  "end_s": 1.5,
+  "note": "first / last frame held, no extra simulation"
+ },
+ "image": "efe-private:autobio-0.1.0",
+ "encode": "libx264 preset slow CRF 16, yuv420p, bt709, faststart",
+ "poster": "last frame (solved state)",
+ "seconds_per_frame": 5.64,
+ "created": "2026-10-23T02:01:53-0500"
+}

masters/autobio/screw_tighten/video.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a115dc6a109b3e19f5e7a02acb10ab7330d706e5498e3b14249288203fdc46c6
+size 14088623

masters/autobio/teaser.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a115dc6a109b3e19f5e7a02acb10ab7330d706e5498e3b14249288203fdc46c6
+size 14088623

masters/autobio/thermal_cycler_close/meta.json

@@ -0,0 +1,74 @@
+{
+ "suite": "autobio",
+ "task": "thermal_cycler_close",
+ "task_dir": "autobio-thermal-cycler-close-i00-privileged",
+ "source": "oracle",
+ "agent": "reference solution (oracle)",
+ "run": null,
+ "mode": "privileged",
+ "trial": null,
+ "trial_choice": "no counted success by any agent other than GPT-6 Luna, whose trials are never shown: the task's reference solution (AutoBio's scripted expert, solution/oracle.npz)",
+ "needs_owner_ok": true,
+ "renderer": "Blender 4.2.23 LTS Cycles path tracing (GPU OptiX), adaptive sampling (threshold 0.01, 64-128 spp), OpenImageDenoise with albedo + normal passes; kinematic playback of the exact MuJoCo replay poses (MuJoCo's own compiled meshes, primitives, colours and UV textures; AutoBio's liquid-surface disc and the thermal mixer's display drawn as AutoBio's own renderer draws them); seamless studio floor (cyclorama), one large soft key area light, a weak overhead softbox, a rim light, a dim sky dome (Poly Haven studio_small_09 HDRI, CC0); AgX view transform, Medium High Contrast look; fixed 3/4 camera with a slow eased 10-degree orbit; native 2560x1440",
+ "rendering": "path-tracing",
+ "media_standard": "v2 (2026-10-16)",
+ "width": 2560,
+ "height": 1440,
+ "fps": 30,
+ "frames": 366,
+ "seconds": 12.2,
+ "speed": 1.0,
+ "success_check": true,
+ "success_check_how": "the task's reference solution (solution/oracle.npz) replayed from the frozen instance with the task's own tooling (rcb_autobio.Session, as tests/grade.py replays it): success by AutoBio's check; a second plain replay (Session.replay) ends in the same end-state hash",
+ "end_check": {
+  "success": true,
+  "rule": "first",
+  "reason": "success",
+  "score": 1.0,
+  "sim_time_s": 10.136
+ },
+ "replay": {
+  "control_hz": 500.0,
+  "n_actions": 5068,
+  "control_steps_replayed": 5068,
+  "sim_time_s": 10.136,
+  "frames_every_control_steps": 16.6667,
+  "final_state_sha": "a95a7852dc40a74a",
+  "plain_replay_state_sha": "a95a7852dc40a74a",
+  "replays_agree": true,
+  "model_sha": "67478954d8efa27b",
+  "init_sha": "3c4a2c9c1b48659b",
+  "instance_ok": true,
+  "state_hashes_every_control_steps": 1000,
+  "trajectory_sha256": "2e7d2c68a85f36814002246f48d483afb3a93c6d7b68aedd099e03f86fa34ca6",
+  "mujoco": "3.3.0"
+ },
+ "visual_changes": [
+  "studio floor and cove instead of MuJoCo's checker ground and skybox"
+ ],
+ "camera": {
+  "mode": "fixed",
+  "frames_on": "the instruments and objects of the task and the arms' end effectors",
+  "az_deg": -35.0,
+  "el_deg": 30.0,
+  "fovy_deg": 34.0,
+  "fill": 0.78,
+  "target": [
+   0.0971,
+   0.2205,
+   1.1554
+  ],
+  "distance_m": 1.7781,
+  "orbit_deg": 10.0
+ },
+ "holds": {
+  "start_s": 0.5,
+  "end_s": 1.5,
+  "note": "first / last frame held, no extra simulation"
+ },
+ "image": "efe-private:autobio-0.1.0",
+ "encode": "libx264 preset slow CRF 16, yuv420p, bt709, faststart",
+ "poster": "last frame (solved state)",
+ "seconds_per_frame": 3.81,
+ "created": "2026-10-23T01:27:27-0500"
+}

masters/autobio/thermal_cycler_close/video.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0a1bb88c785f4abc236a75b6081a0f6c2319c2c80c0e14a648f83b1495a93455
+size 6024061

masters/autobio/thermal_cycler_open/meta.json

@@ -0,0 +1,75 @@
+{
+ "suite": "autobio",
+ "task": "thermal_cycler_open",
+ "task_dir": "autobio-thermal-cycler-open-i00-standard",
+ "source": "agent-success:codex-gpt6_astra-high-autobio-thermal-cycler-open-i00-standard/autobio-thermal-cycler-open-i00__5AQRomv",
+ "agent": "codex + GPT-6 Astra (high)",
+ "run": "codex-gpt6_astra-high",
+ "mode": "standard",
+ "trial": "codex-gpt6_astra-high-autobio-thermal-cycler-open-i00-standard/autobio-thermal-cycler-open-i00__5AQRomv",
+ "trial_choice": "a counted success, never GPT-6 Luna's: privileged before standard, GPT-6 Astra first, then Claude Opus 5.5, Gemini 3.8 Flash, DeepSeek, GLM",
+ "renderer": "Blender 4.2.23 LTS Cycles path tracing (GPU OptiX), adaptive sampling (threshold 0.01, 64-128 spp), OpenImageDenoise with albedo + normal passes; kinematic playback of the exact MuJoCo replay poses (MuJoCo's own compiled meshes, primitives, colours and UV textures; AutoBio's liquid-surface disc and the thermal mixer's display drawn as AutoBio's own renderer draws them); seamless studio floor (cyclorama), one large soft key area light, a weak overhead softbox, a rim light, a dim sky dome (Poly Haven studio_small_09 HDRI, CC0); AgX view transform, Medium High Contrast look; fixed 3/4 camera with a slow eased 10-degree orbit; native 2560x1440",
+ "rendering": "path-tracing",
+ "media_standard": "v2 (2026-10-16)",
+ "width": 2560,
+ "height": 1440,
+ "fps": 30,
+ "frames": 710,
+ "seconds": 23.67,
+ "speed": 2.0,
+ "success_check": true,
+ "success_check_how": "standard-mode trial (graded by its live episode): re-rendered only because the replay is bit-exact. The recorded rows were replayed from the task's frozen instance with the task's own tooling (rcb_autobio.Session, as the sim service runs them); the replay ends in AutoBio's success check with an end-state hash equal to the live episode's recorded hash (624e6793c4a6861d), and a second plain replay (Session.replay) ends in the same hash",
+ "end_check": {
+  "success": true,
+  "rule": "first",
+  "reason": "success",
+  "score": 1.0,
+  "sim_time_s": 43.236
+ },
+ "replay": {
+  "control_hz": 500.0,
+  "n_actions": 21618,
+  "control_steps_replayed": 21618,
+  "sim_time_s": 43.236,
+  "frames_every_control_steps": 33.3333,
+  "final_state_sha": "624e6793c4a6861d",
+  "plain_replay_state_sha": "624e6793c4a6861d",
+  "replays_agree": true,
+  "model_sha": "67478954d8efa27b",
+  "init_sha": "af233b90fd79689a",
+  "instance_ok": true,
+  "state_hashes_every_control_steps": 1000,
+  "trajectory_sha256": "aff6d1e532964fe592db1b4bf4baf82b8757490a06bcc94961c1674b540c6dba",
+  "mujoco": "3.3.0",
+  "recorded_live_state_sha": "624e6793c4a6861d",
+  "matches_live_episode": true
+ },
+ "visual_changes": [
+  "studio floor and cove instead of MuJoCo's checker ground and skybox"
+ ],
+ "camera": {
+  "mode": "fixed",
+  "frames_on": "the instruments and objects of the task and the arms' end effectors",
+  "az_deg": -35.0,
+  "el_deg": 30.0,
+  "fovy_deg": 34.0,
+  "fill": 0.78,
+  "target": [
+   0.1294,
+   0.2156,
+   1.1425
+  ],
+  "distance_m": 1.651,
+  "orbit_deg": 10.0
+ },
+ "holds": {
+  "start_s": 0.5,
+  "end_s": 1.5,
+  "note": "first / last frame held, no extra simulation"
+ },
+ "image": "efe-private:autobio-0.1.0",
+ "encode": "libx264 preset slow CRF 16, yuv420p, bt709, faststart",
+ "poster": "last frame (solved state)",
+ "seconds_per_frame": 6.09,
+ "created": "2026-10-23T02:02:11-0500"
+}

masters/autobio/thermal_cycler_open/video.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3039a6c7508a26686c870719dcb2754313e50104d852dd530cf045efb73c5b45
+size 11712923

masters/autobio/thermal_mixer/meta.json

@@ -0,0 +1,76 @@
+{
+ "suite": "autobio",
+ "task": "thermal_mixer",
+ "task_dir": "autobio-thermal-mixer-i00-standard",
+ "source": "agent-success:codex-gpt6_astra-high-autobio-thermal-mixer-i00-standard/autobio-thermal-mixer-i00-standa__Qe3kNyh",
+ "agent": "codex + GPT-6 Astra (high)",
+ "run": "codex-gpt6_astra-high",
+ "mode": "standard",
+ "trial": "codex-gpt6_astra-high-autobio-thermal-mixer-i00-standard/autobio-thermal-mixer-i00-standa__Qe3kNyh",
+ "trial_choice": "a counted success, never GPT-6 Luna's: privileged before standard, GPT-6 Astra first, then Claude Opus 5.5, Gemini 3.8 Flash, DeepSeek, GLM",
+ "renderer": "Blender 4.2.23 LTS Cycles path tracing (GPU OptiX), adaptive sampling (threshold 0.01, 64-128 spp), OpenImageDenoise with albedo + normal passes; kinematic playback of the exact MuJoCo replay poses (MuJoCo's own compiled meshes, primitives, colours and UV textures; AutoBio's liquid-surface disc and the thermal mixer's display drawn as AutoBio's own renderer draws them); seamless studio floor (cyclorama), one large soft key area light, a weak overhead softbox, a rim light, a dim sky dome (Poly Haven studio_small_09 HDRI, CC0); AgX view transform, Medium High Contrast look; fixed 3/4 camera with a slow eased 10-degree orbit; native 2560x1440",
+ "rendering": "path-tracing",
+ "media_standard": "v2 (2026-10-16)",
+ "width": 2560,
+ "height": 1440,
+ "fps": 30,
+ "frames": 688,
+ "seconds": 22.93,
+ "speed": 15.0,
+ "success_check": true,
+ "success_check_how": "standard-mode trial (graded by its live episode): re-rendered only because the replay is bit-exact. The recorded rows were replayed from the task's frozen instance with the task's own tooling (rcb_autobio.Session, as the sim service runs them); the replay ends in AutoBio's success check with an end-state hash equal to the live episode's recorded hash (3898bbd99a13ba14), and a second plain replay (Session.replay) ends in the same hash",
+ "end_check": {
+  "success": true,
+  "rule": "final",
+  "reason": null,
+  "score": 1.0,
+  "sim_time_s": 313.15
+ },
+ "replay": {
+  "control_hz": 500.0,
+  "n_actions": 156575,
+  "control_steps_replayed": 156575,
+  "sim_time_s": 313.15,
+  "frames_every_control_steps": 250.0,
+  "final_state_sha": "3898bbd99a13ba14",
+  "plain_replay_state_sha": "3898bbd99a13ba14",
+  "replays_agree": true,
+  "model_sha": "9ad959e8120cbbea",
+  "init_sha": "40d5857ebf38fb99",
+  "instance_ok": true,
+  "state_hashes_every_control_steps": 1000,
+  "trajectory_sha256": "2ca7e5381cb2f94662454c9a33e7d13e110570210f754d5ec4526633764df5bc",
+  "mujoco": "3.3.0",
+  "recorded_live_state_sha": "3898bbd99a13ba14",
+  "matches_live_episode": true
+ },
+ "visual_changes": [
+  "studio floor and cove instead of MuJoCo's checker ground and skybox",
+  "the thermal mixer's display drawn from AutoBio's own panel drawing, per frame"
+ ],
+ "camera": {
+  "mode": "fixed",
+  "frames_on": "the instruments and objects of the task and the arms' end effectors",
+  "az_deg": -35.0,
+  "el_deg": 30.0,
+  "fovy_deg": 34.0,
+  "fill": 0.78,
+  "target": [
+   0.2759,
+   0.2839,
+   1.0209
+  ],
+  "distance_m": 1.28,
+  "orbit_deg": 10.0
+ },
+ "holds": {
+  "start_s": 0.5,
+  "end_s": 1.5,
+  "note": "first / last frame held, no extra simulation"
+ },
+ "image": "efe-private:autobio-0.1.0",
+ "encode": "libx264 preset slow CRF 16, yuv420p, bt709, faststart",
+ "poster": "last frame (solved state)",
+ "seconds_per_frame": 6.35,
+ "created": "2026-10-23T02:02:03-0500"
+}

masters/autobio/thermal_mixer/video.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5c8ff26511d9380d74147383658a3a9647acb52db7003b2f978b79c0ce4774bb
+size 15018322