ABC Sim: 29 task masters, teaser, highlights (media v2)

masters/abc-sim/ball_tray/meta.json

@@ -0,0 +1,78 @@
+{
+ "suite": "abc-sim",
+ "task": "ball_tray",
+ "task_dir": "abc-sim-ball-tray-i00-privileged",
+ "source": "agent-success:codex-gpt6_luna-xhigh-abc-sim-ball-tray-i00-privileged/abc-sim-ball-tray-i00-privileged__BpBt4SA",
+ "agent": "codex + GPT-6 Luna (xhigh)",
+ "run": "codex-gpt6_luna-xhigh",
+ "mode": "privileged",
+ "trial": "codex-gpt6_luna-xhigh-abc-sim-ball-tray-i00-privileged/abc-sim-ball-tray-i00-privileged__BpBt4SA",
+ "trial_choice": "GPT-6 Luna's counted success; privileged before standard (a privileged trial is the submitted trajectory, replayed after the run)",
+ "renderer": "Blender 4.5.3 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); 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": 551,
+ "seconds": 18.37,
+ "speed": 1.0,
+ "success_check": true,
+ "success_check_how": "privileged trial: the submitted trajectory (trajectory.npz) replayed from the task's frozen instance with the task's own tooling (rcb_abc.Session, as tests' grade.py replays it): success by the task's rule (upstream's evaluator held 30 control steps; ball tray: the countdown completed with the ball on the tray and both handles held), the end-state hash equals both grading replays' hash (22dc6e5abac334f2), and a second plain replay in a fresh session (Session.replay) ends in the same hash",
+ "end_check": {
+  "success": true,
+  "rule": "tray",
+  "reason": "success",
+  "steps": 480,
+  "success_run": 0,
+  "hold_steps": 30
+ },
+ "replay": {
+  "control_hz": 29.4118,
+  "control_dt_s": 0.034,
+  "n_actions": 480,
+  "control_steps_replayed": 480,
+  "sim_time_s": 16.32,
+  "frames_every_control_steps": 0.9804,
+  "final_state_sha": "22dc6e5abac334f2",
+  "plain_replay_state_sha": "22dc6e5abac334f2",
+  "replays_agree": true,
+  "instance_sha": "6cdaa86a450092e7",
+  "init_sha": "6cdaa86a450092e7",
+  "instance_ok": true,
+  "state_hashes_every_control_steps": 1000,
+  "trajectory_sha256": "a5043076c687d194e7a0aea36610fd08dfbe9de32550c62449dd74e0f265ef02",
+  "mujoco": "3.6.0",
+  "recorded_grader_state_sha": "22dc6e5abac334f2"
+ },
+ "visual_changes": [
+  "studio floor and cove instead of MuJoCo's floor and the booth's three walls",
+  "the overhead camera gate (bar and camera) not drawn",
+  "the table top drawn as a solid light-wood block down to the floor, extended under the arms' base mounts (top face unchanged; the scene's table is a thin white slab)"
+ ],
+ "camera": {
+  "mode": "fixed",
+  "frames_on": "both arms and every object on the table",
+  "az_deg": 35.0,
+  "el_deg": 32.0,
+  "fovy_deg": 34.0,
+  "fill": 0.8,
+  "target": [
+   0.4355,
+   -0.0149,
+   0.8922
+  ],
+  "distance_m": 1.6022,
+  "orbit_deg": 10.0
+ },
+ "holds": {
+  "start_s": 0.5,
+  "end_s": 1.5,
+  "note": "first / last frame held, no extra simulation"
+ },
+ "image": "efe:abc-sim-0.1.0",
+ "encode": "libx264 preset slow CRF 16, yuv420p, bt709, faststart",
+ "poster": "last frame (solved state)",
+ "seconds_per_frame": 8.65,
+ "created": "2026-10-31T05:25:26-0500"
+}

masters/abc-sim/ball_tray/video.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:915da35c5ea292015d09be089871d3946b6b9a6d2f93f528cd9efd381c4affb7
+size 9130694

masters/abc-sim/chess/meta.json

@@ -0,0 +1,79 @@
+{
+ "suite": "abc-sim",
+ "task": "chess",
+ "task_dir": "abc-sim-chess-i00-privileged",
+ "source": "agent-success:codex-gpt6_luna-xhigh-abc-sim-chess-i00-privileged/abc-sim-chess-i00-privileged__C6wgrLQ",
+ "agent": "codex + GPT-6 Luna (xhigh)",
+ "run": "codex-gpt6_luna-xhigh",
+ "mode": "privileged",
+ "trial": "codex-gpt6_luna-xhigh-abc-sim-chess-i00-privileged/abc-sim-chess-i00-privileged__C6wgrLQ",
+ "trial_choice": "GPT-6 Luna's counted success; privileged before standard (a privileged trial is the submitted trajectory, replayed after the run)",
+ "renderer": "Blender 4.5.3 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); 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": 699,
+ "seconds": 23.3,
+ "speed": 12.0,
+ "success_check": true,
+ "success_check_how": "privileged trial: the submitted trajectory (trajectory.npz) replayed from the task's frozen instance with the task's own tooling (rcb_abc.Session, as tests' grade.py replays it): success by the task's rule (upstream's evaluator held 30 control steps), the end-state hash equals both grading replays' hash (4efa1720e99e91e1), and a second plain replay in a fresh session (Session.replay) ends in the same hash",
+ "end_check": {
+  "success": true,
+  "rule": "event",
+  "reason": "success",
+  "steps": 7495,
+  "success_run": 30,
+  "hold_steps": 30
+ },
+ "replay": {
+  "control_hz": 29.4118,
+  "control_dt_s": 0.034,
+  "n_actions": 7496,
+  "control_steps_replayed": 7495,
+  "sim_time_s": 254.83,
+  "frames_every_control_steps": 11.7647,
+  "final_state_sha": "4efa1720e99e91e1",
+  "plain_replay_state_sha": "4efa1720e99e91e1",
+  "replays_agree": true,
+  "instance_sha": "65f14022364f054f",
+  "init_sha": "65f14022364f054f",
+  "instance_ok": true,
+  "state_hashes_every_control_steps": 1000,
+  "trajectory_sha256": "4f75c780aee4c406e4e0372fbdb4ee7097bf2b46010afe6020042c84cb563510",
+  "mujoco": "3.6.0",
+  "recorded_grader_state_sha": "4efa1720e99e91e1"
+ },
+ "visual_changes": [
+  "studio floor and cove instead of MuJoCo's floor and the booth's three walls",
+  "the overhead camera gate (bar and camera) not drawn",
+  "the table top drawn as a solid light-wood block down to the floor, extended under the arms' base mounts (top face unchanged; the scene's table is a thin white slab)",
+  "unused objects the scene parks far below or above the table (they fall freely and touch nothing) not drawn: tin_box"
+ ],
+ "camera": {
+  "mode": "fixed",
+  "frames_on": "both arms and every object on the table",
+  "az_deg": 35.0,
+  "el_deg": 32.0,
+  "fovy_deg": 34.0,
+  "fill": 0.8,
+  "target": [
+   0.5429,
+   -0.0119,
+   0.8681
+  ],
+  "distance_m": 1.8784,
+  "orbit_deg": 10.0
+ },
+ "holds": {
+  "start_s": 0.5,
+  "end_s": 1.5,
+  "note": "first / last frame held, no extra simulation"
+ },
+ "image": "efe:abc-sim-0.1.0",
+ "encode": "libx264 preset slow CRF 16, yuv420p, bt709, faststart",
+ "poster": "last frame (solved state)",
+ "seconds_per_frame": 4.8,
+ "created": "2026-10-31T04:41:08-0500"
+}

masters/abc-sim/chess/video.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c5be7160e8e7b79400aca103b19286cbb0641fe604ce5c1d2954f728351f0baf
+size 12097385

masters/abc-sim/conveyor_pick/meta.json

@@ -0,0 +1,78 @@
+{
+ "suite": "abc-sim",
+ "task": "conveyor_pick",
+ "task_dir": "abc-sim-conveyor-pick-i00-privileged",
+ "source": "agent-success:codex-gpt6_luna-xhigh-abc-sim-conveyor-pick-i00-privileged/abc-sim-conveyor-pick-i00-privil__gMbxpq7",
+ "agent": "codex + GPT-6 Luna (xhigh)",
+ "run": "codex-gpt6_luna-xhigh",
+ "mode": "privileged",
+ "trial": "codex-gpt6_luna-xhigh-abc-sim-conveyor-pick-i00-privileged/abc-sim-conveyor-pick-i00-privil__gMbxpq7",
+ "trial_choice": "GPT-6 Luna's counted success; privileged before standard (a privileged trial is the submitted trajectory, replayed after the run)",
+ "renderer": "Blender 4.5.3 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); 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": 391,
+ "seconds": 13.03,
+ "speed": 1.0,
+ "success_check": true,
+ "success_check_how": "privileged trial: the submitted trajectory (trajectory.npz) replayed from the task's frozen instance with the task's own tooling (rcb_abc.Session, as tests' grade.py replays it): success by the task's rule (upstream's evaluator held 30 control steps), the end-state hash equals both grading replays' hash (58f73560a19e6d6f), and a second plain replay in a fresh session (Session.replay) ends in the same hash",
+ "end_check": {
+  "success": true,
+  "rule": "event",
+  "reason": "success",
+  "steps": 324,
+  "success_run": 30,
+  "hold_steps": 30
+ },
+ "replay": {
+  "control_hz": 29.4118,
+  "control_dt_s": 0.034,
+  "n_actions": 324,
+  "control_steps_replayed": 324,
+  "sim_time_s": 11.016,
+  "frames_every_control_steps": 0.9804,
+  "final_state_sha": "58f73560a19e6d6f",
+  "plain_replay_state_sha": "58f73560a19e6d6f",
+  "replays_agree": true,
+  "instance_sha": "73fd8e08bd4000ee",
+  "init_sha": "73fd8e08bd4000ee",
+  "instance_ok": true,
+  "state_hashes_every_control_steps": 1000,
+  "trajectory_sha256": "e6cd92bde92818e7540e6885adcfc9fad598115f51c1852a0b4e768f7e549774",
+  "mujoco": "3.6.0",
+  "recorded_grader_state_sha": "58f73560a19e6d6f"
+ },
+ "visual_changes": [
+  "studio floor and cove instead of MuJoCo's floor and the booth's three walls",
+  "the overhead camera gate (bar and camera) not drawn",
+  "the table top drawn as a solid light-wood block down to the floor, extended under the arms' base mounts (top face unchanged; the scene's table is a thin white slab)"
+ ],
+ "camera": {
+  "mode": "fixed",
+  "frames_on": "both arms and every object on the table",
+  "az_deg": 35.0,
+  "el_deg": 32.0,
+  "fovy_deg": 34.0,
+  "fill": 0.8,
+  "target": [
+   0.414,
+   0.0774,
+   0.8928
+  ],
+  "distance_m": 1.6922,
+  "orbit_deg": 10.0
+ },
+ "holds": {
+  "start_s": 0.5,
+  "end_s": 1.5,
+  "note": "first / last frame held, no extra simulation"
+ },
+ "image": "efe:abc-sim-0.1.0",
+ "encode": "libx264 preset slow CRF 16, yuv420p, bt709, faststart",
+ "poster": "last frame (solved state)",
+ "seconds_per_frame": 2.87,
+ "created": "2026-10-31T05:25:29-0500"
+}

masters/abc-sim/conveyor_pick/video.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:66a6ba5c549ccf870b04f2d20918a6372729252e07a0a38150cb420f58c84356
+size 8474381

masters/abc-sim/count/meta.json

@@ -0,0 +1,78 @@
+{
+ "suite": "abc-sim",
+ "task": "count",
+ "task_dir": "abc-sim-count-i00-privileged",
+ "source": "agent-success:codex-gpt6_luna-xhigh-abc-sim-count-i00-privileged/abc-sim-count-i00-privileged__AM5P4XK",
+ "agent": "codex + GPT-6 Luna (xhigh)",
+ "run": "codex-gpt6_luna-xhigh",
+ "mode": "privileged",
+ "trial": "codex-gpt6_luna-xhigh-abc-sim-count-i00-privileged/abc-sim-count-i00-privileged__AM5P4XK",
+ "trial_choice": "GPT-6 Luna's counted success; privileged before standard (a privileged trial is the submitted trajectory, replayed after the run)",
+ "renderer": "Blender 4.5.3 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); 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": 564,
+ "seconds": 18.8,
+ "speed": 4.0,
+ "success_check": true,
+ "success_check_how": "privileged trial: the submitted trajectory (trajectory.npz) replayed from the task's frozen instance with the task's own tooling (rcb_abc.Session, as tests' grade.py replays it): success by the task's rule (upstream's evaluator held 30 control steps), the end-state hash equals both grading replays' hash (04b33d53e5808f8a), and a second plain replay in a fresh session (Session.replay) ends in the same hash",
+ "end_check": {
+  "success": true,
+  "rule": "event",
+  "reason": "success",
+  "steps": 1972,
+  "success_run": 30,
+  "hold_steps": 30
+ },
+ "replay": {
+  "control_hz": 29.4118,
+  "control_dt_s": 0.034,
+  "n_actions": 1972,
+  "control_steps_replayed": 1972,
+  "sim_time_s": 67.048,
+  "frames_every_control_steps": 3.9216,
+  "final_state_sha": "04b33d53e5808f8a",
+  "plain_replay_state_sha": "04b33d53e5808f8a",
+  "replays_agree": true,
+  "instance_sha": "080ddb7334cfb5d4",
+  "init_sha": "080ddb7334cfb5d4",
+  "instance_ok": true,
+  "state_hashes_every_control_steps": 1000,
+  "trajectory_sha256": "6bde4f097a82747a243b1141ab2d631d1b0e1161e54748f6fda5ba1f8edba17c",
+  "mujoco": "3.6.0",
+  "recorded_grader_state_sha": "04b33d53e5808f8a"
+ },
+ "visual_changes": [
+  "studio floor and cove instead of MuJoCo's floor and the booth's three walls",
+  "the overhead camera gate (bar and camera) not drawn",
+  "the table top drawn as a solid light-wood block down to the floor, extended under the arms' base mounts (top face unchanged; the scene's table is a thin white slab)"
+ ],
+ "camera": {
+  "mode": "fixed",
+  "frames_on": "both arms and every object on the table",
+  "az_deg": 35.0,
+  "el_deg": 32.0,
+  "fovy_deg": 34.0,
+  "fill": 0.8,
+  "target": [
+   0.4595,
+   0.0087,
+   0.9239
+  ],
+  "distance_m": 1.6569,
+  "orbit_deg": 10.0
+ },
+ "holds": {
+  "start_s": 0.5,
+  "end_s": 1.5,
+  "note": "first / last frame held, no extra simulation"
+ },
+ "image": "efe:abc-sim-0.1.0",
+ "encode": "libx264 preset slow CRF 16, yuv420p, bt709, faststart",
+ "poster": "last frame (solved state)",
+ "seconds_per_frame": 7.76,
+ "created": "2026-10-31T06:36:22-0500"
+}

masters/abc-sim/count/video.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a4ad490bd40cc30dcf9d41553320d23ceafcc1e36230262ec8297ed1c4367246
+size 8719049

masters/abc-sim/count_four/meta.json

@@ -0,0 +1,78 @@
+{
+ "suite": "abc-sim",
+ "task": "count_four",
+ "task_dir": "abc-sim-count-four-i00-privileged",
+ "source": "agent-success:codex-gpt6_luna-xhigh-abc-sim-count-four-i00-privileged/abc-sim-count-four-i00-privilege__oMffyCC",
+ "agent": "codex + GPT-6 Luna (xhigh)",
+ "run": "codex-gpt6_luna-xhigh",
+ "mode": "privileged",
+ "trial": "codex-gpt6_luna-xhigh-abc-sim-count-four-i00-privileged/abc-sim-count-four-i00-privilege__oMffyCC",
+ "trial_choice": "GPT-6 Luna's counted success; privileged before standard (a privileged trial is the submitted trajectory, replayed after the run)",
+ "renderer": "Blender 4.5.3 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); 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": 648,
+ "seconds": 21.6,
+ "speed": 8.0,
+ "success_check": true,
+ "success_check_how": "privileged trial: the submitted trajectory (trajectory.npz) replayed from the task's frozen instance with the task's own tooling (rcb_abc.Session, as tests' grade.py replays it): success by the task's rule (upstream's evaluator held 30 control steps), the end-state hash equals both grading replays' hash (5a16271694057122), and a second plain replay in a fresh session (Session.replay) ends in the same hash",
+ "end_check": {
+  "success": true,
+  "rule": "event",
+  "reason": "success",
+  "steps": 4604,
+  "success_run": 30,
+  "hold_steps": 30
+ },
+ "replay": {
+  "control_hz": 29.4118,
+  "control_dt_s": 0.034,
+  "n_actions": 4604,
+  "control_steps_replayed": 4604,
+  "sim_time_s": 156.536,
+  "frames_every_control_steps": 7.8431,
+  "final_state_sha": "5a16271694057122",
+  "plain_replay_state_sha": "5a16271694057122",
+  "replays_agree": true,
+  "instance_sha": "c8a0c7b3eb8622b4",
+  "init_sha": "c8a0c7b3eb8622b4",
+  "instance_ok": true,
+  "state_hashes_every_control_steps": 1000,
+  "trajectory_sha256": "62358acbf400ad4be5b0324a7ac799c94b97d07f123ecd121ecfa47dc9ad77a3",
+  "mujoco": "3.6.0",
+  "recorded_grader_state_sha": "5a16271694057122"
+ },
+ "visual_changes": [
+  "studio floor and cove instead of MuJoCo's floor and the booth's three walls",
+  "the overhead camera gate (bar and camera) not drawn",
+  "the table top drawn as a solid light-wood block down to the floor, extended under the arms' base mounts (top face unchanged; the scene's table is a thin white slab)"
+ ],
+ "camera": {
+  "mode": "fixed",
+  "frames_on": "both arms and every object on the table",
+  "az_deg": 35.0,
+  "el_deg": 32.0,
+  "fovy_deg": 34.0,
+  "fill": 0.8,
+  "target": [
+   0.4584,
+   0.0256,
+   1.02
+  ],
+  "distance_m": 1.9123,
+  "orbit_deg": 10.0
+ },
+ "holds": {
+  "start_s": 0.5,
+  "end_s": 1.5,
+  "note": "first / last frame held, no extra simulation"
+ },
+ "image": "efe:abc-sim-0.1.0",
+ "encode": "libx264 preset slow CRF 16, yuv420p, bt709, faststart",
+ "poster": "last frame (solved state)",
+ "seconds_per_frame": 10.18,
+ "created": "2026-10-31T06:47:48-0500"
+}

masters/abc-sim/count_four/video.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:95f5c55778143c05d609d8f0d6c300ec779ae5ee9aacc46674a040f0b18beb4c
+size 9420711

masters/abc-sim/count_one/meta.json

@@ -0,0 +1,78 @@
+{
+ "suite": "abc-sim",
+ "task": "count_one",
+ "task_dir": "abc-sim-count-one-i00-privileged",
+ "source": "agent-success:codex-gpt6_luna-xhigh-abc-sim-count-one-i00-privileged/abc-sim-count-one-i00-privileged__4hBS3vt",
+ "agent": "codex + GPT-6 Luna (xhigh)",
+ "run": "codex-gpt6_luna-xhigh",
+ "mode": "privileged",
+ "trial": "codex-gpt6_luna-xhigh-abc-sim-count-one-i00-privileged/abc-sim-count-one-i00-privileged__4hBS3vt",
+ "trial_choice": "GPT-6 Luna's counted success; privileged before standard (a privileged trial is the submitted trajectory, replayed after the run)",
+ "renderer": "Blender 4.5.3 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); 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": 212,
+ "seconds": 7.07,
+ "speed": 0.5,
+ "success_check": true,
+ "success_check_how": "privileged trial: the submitted trajectory (trajectory.npz) replayed from the task's frozen instance with the task's own tooling (rcb_abc.Session, as tests' grade.py replays it): success by the task's rule (upstream's evaluator held 30 control steps), the end-state hash equals both grading replays' hash (21a0e0be11ba1e73), and a second plain replay in a fresh session (Session.replay) ends in the same hash",
+ "end_check": {
+  "success": true,
+  "rule": "event",
+  "reason": "success",
+  "steps": 74,
+  "success_run": 30,
+  "hold_steps": 30
+ },
+ "replay": {
+  "control_hz": 29.4118,
+  "control_dt_s": 0.034,
+  "n_actions": 74,
+  "control_steps_replayed": 74,
+  "sim_time_s": 2.516,
+  "frames_every_control_steps": 0.4902,
+  "final_state_sha": "21a0e0be11ba1e73",
+  "plain_replay_state_sha": "21a0e0be11ba1e73",
+  "replays_agree": true,
+  "instance_sha": "39fcd07ad171ba75",
+  "init_sha": "39fcd07ad171ba75",
+  "instance_ok": true,
+  "state_hashes_every_control_steps": 1000,
+  "trajectory_sha256": "9dae15aa7935b47d5fe0d2097df5f201f78c979179f7a9e20059df6aef95970c",
+  "mujoco": "3.6.0",
+  "recorded_grader_state_sha": "21a0e0be11ba1e73"
+ },
+ "visual_changes": [
+  "studio floor and cove instead of MuJoCo's floor and the booth's three walls",
+  "the overhead camera gate (bar and camera) not drawn",
+  "the table top drawn as a solid light-wood block down to the floor, extended under the arms' base mounts (top face unchanged; the scene's table is a thin white slab)"
+ ],
+ "camera": {
+  "mode": "fixed",
+  "frames_on": "both arms and every object on the table",
+  "az_deg": 35.0,
+  "el_deg": 32.0,
+  "fovy_deg": 34.0,
+  "fill": 0.8,
+  "target": [
+   0.458,
+   0.0434,
+   0.9225
+  ],
+  "distance_m": 1.5963,
+  "orbit_deg": 10.0
+ },
+ "holds": {
+  "start_s": 0.5,
+  "end_s": 1.5,
+  "note": "first / last frame held, no extra simulation"
+ },
+ "image": "efe:abc-sim-0.1.0",
+ "encode": "libx264 preset slow CRF 16, yuv420p, bt709, faststart",
+ "poster": "last frame (solved state)",
+ "seconds_per_frame": 7.52,
+ "created": "2026-10-31T06:02:03-0500"
+}

masters/abc-sim/count_one/video.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:81df79ac63293b45ed2bbd1a598061cc619dc22ff2a087f24a9e6c6086d79e18
+size 3469956

masters/abc-sim/count_three/meta.json

@@ -0,0 +1,78 @@
+{
+ "suite": "abc-sim",
+ "task": "count_three",
+ "task_dir": "abc-sim-count-three-i00-privileged",
+ "source": "agent-success:codex-gpt6_luna-xhigh-abc-sim-count-three-i00-privileged/abc-sim-count-three-i00-privileg__LgzsQAJ",
+ "agent": "codex + GPT-6 Luna (xhigh)",
+ "run": "codex-gpt6_luna-xhigh",
+ "mode": "privileged",
+ "trial": "codex-gpt6_luna-xhigh-abc-sim-count-three-i00-privileged/abc-sim-count-three-i00-privileg__LgzsQAJ",
+ "trial_choice": "GPT-6 Luna's counted success; privileged before standard (a privileged trial is the submitted trajectory, replayed after the run)",
+ "renderer": "Blender 4.5.3 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); 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": 647,
+ "seconds": 21.57,
+ "speed": 1.5,
+ "success_check": true,
+ "success_check_how": "privileged trial: the submitted trajectory (trajectory.npz) replayed from the task's frozen instance with the task's own tooling (rcb_abc.Session, as tests' grade.py replays it): success by the task's rule (upstream's evaluator held 30 control steps), the end-state hash equals both grading replays' hash (16cde7e8423c0ce3), and a second plain replay in a fresh session (Session.replay) ends in the same hash",
+ "end_check": {
+  "success": true,
+  "rule": "event",
+  "reason": "success",
+  "steps": 862,
+  "success_run": 30,
+  "hold_steps": 30
+ },
+ "replay": {
+  "control_hz": 29.4118,
+  "control_dt_s": 0.034,
+  "n_actions": 862,
+  "control_steps_replayed": 862,
+  "sim_time_s": 29.308,
+  "frames_every_control_steps": 1.4706,
+  "final_state_sha": "16cde7e8423c0ce3",
+  "plain_replay_state_sha": "16cde7e8423c0ce3",
+  "replays_agree": true,
+  "instance_sha": "4a649f7e73ba4df1",
+  "init_sha": "4a649f7e73ba4df1",
+  "instance_ok": true,
+  "state_hashes_every_control_steps": 1000,
+  "trajectory_sha256": "d3648bd3fef062b205af60be9812dfe5d3dae3662cb73c308216ee91421acba5",
+  "mujoco": "3.6.0",
+  "recorded_grader_state_sha": "16cde7e8423c0ce3"
+ },
+ "visual_changes": [
+  "studio floor and cove instead of MuJoCo's floor and the booth's three walls",
+  "the overhead camera gate (bar and camera) not drawn",
+  "the table top drawn as a solid light-wood block down to the floor, extended under the arms' base mounts (top face unchanged; the scene's table is a thin white slab)"
+ ],
+ "camera": {
+  "mode": "fixed",
+  "frames_on": "both arms and every object on the table",
+  "az_deg": 35.0,
+  "el_deg": 32.0,
+  "fovy_deg": 34.0,
+  "fill": 0.8,
+  "target": [
+   0.4264,
+   0.0576,
+   0.8823
+  ],
+  "distance_m": 1.5317,
+  "orbit_deg": 10.0
+ },
+ "holds": {
+  "start_s": 0.5,
+  "end_s": 1.5,
+  "note": "first / last frame held, no extra simulation"
+ },
+ "image": "efe:abc-sim-0.1.0",
+ "encode": "libx264 preset slow CRF 16, yuv420p, bt709, faststart",
+ "poster": "last frame (solved state)",
+ "seconds_per_frame": 2.84,
+ "created": "2026-10-31T04:41:13-0500"
+}

masters/abc-sim/count_three/video.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:abdb7e6c2007d0cafc85c4977ee6973d83d2634a55c7c415a8d4dd63e328e683
+size 10685828

masters/abc-sim/count_two/meta.json

@@ -0,0 +1,78 @@
+{
+ "suite": "abc-sim",
+ "task": "count_two",
+ "task_dir": "abc-sim-count-two-i00-privileged",
+ "source": "agent-success:codex-gpt6_luna-xhigh-abc-sim-count-two-i00-privileged/abc-sim-count-two-i00-privileged__hjJa4h2",
+ "agent": "codex + GPT-6 Luna (xhigh)",
+ "run": "codex-gpt6_luna-xhigh",
+ "mode": "privileged",
+ "trial": "codex-gpt6_luna-xhigh-abc-sim-count-two-i00-privileged/abc-sim-count-two-i00-privileged__hjJa4h2",
+ "trial_choice": "GPT-6 Luna's counted success; privileged before standard (a privileged trial is the submitted trajectory, replayed after the run)",
+ "renderer": "Blender 4.5.3 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); 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": 448,
+ "seconds": 14.93,
+ "speed": 1.0,
+ "success_check": true,
+ "success_check_how": "privileged trial: the submitted trajectory (trajectory.npz) replayed from the task's frozen instance with the task's own tooling (rcb_abc.Session, as tests' grade.py replays it): success by the task's rule (upstream's evaluator held 30 control steps), the end-state hash equals both grading replays' hash (1d0fc8f13afd69fa), and a second plain replay in a fresh session (Session.replay) ends in the same hash",
+ "end_check": {
+  "success": true,
+  "rule": "event",
+  "reason": "success",
+  "steps": 379,
+  "success_run": 30,
+  "hold_steps": 30
+ },
+ "replay": {
+  "control_hz": 29.4118,
+  "control_dt_s": 0.034,
+  "n_actions": 379,
+  "control_steps_replayed": 379,
+  "sim_time_s": 12.886,
+  "frames_every_control_steps": 0.9804,
+  "final_state_sha": "1d0fc8f13afd69fa",
+  "plain_replay_state_sha": "1d0fc8f13afd69fa",
+  "replays_agree": true,
+  "instance_sha": "5099066e29874d64",
+  "init_sha": "5099066e29874d64",
+  "instance_ok": true,
+  "state_hashes_every_control_steps": 1000,
+  "trajectory_sha256": "535dcc951ef08cddca38241fedd73d9634446ddb98c461207c0f59aab4532cfe",
+  "mujoco": "3.6.0",
+  "recorded_grader_state_sha": "1d0fc8f13afd69fa"
+ },
+ "visual_changes": [
+  "studio floor and cove instead of MuJoCo's floor and the booth's three walls",
+  "the overhead camera gate (bar and camera) not drawn",
+  "the table top drawn as a solid light-wood block down to the floor, extended under the arms' base mounts (top face unchanged; the scene's table is a thin white slab)"
+ ],
+ "camera": {
+  "mode": "fixed",
+  "frames_on": "both arms and every object on the table",
+  "az_deg": 35.0,
+  "el_deg": 32.0,
+  "fovy_deg": 34.0,
+  "fill": 0.8,
+  "target": [
+   0.452,
+   0.0334,
+   0.9174
+  ],
+  "distance_m": 1.6136,
+  "orbit_deg": 10.0
+ },
+ "holds": {
+  "start_s": 0.5,
+  "end_s": 1.5,
+  "note": "first / last frame held, no extra simulation"
+ },
+ "image": "efe:abc-sim-0.1.0",
+ "encode": "libx264 preset slow CRF 16, yuv420p, bt709, faststart",
+ "poster": "last frame (solved state)",
+ "seconds_per_frame": 3.04,
+ "created": "2026-10-31T05:38:11-0500"
+}

masters/abc-sim/count_two/video.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:90d9100bf161041edd2e3b6c6450e53358849405b04711dcc8967f30c6ef7055
+size 7076838

masters/abc-sim/dish_rack/meta.json

@@ -0,0 +1,78 @@
+{
+ "suite": "abc-sim",
+ "task": "dish_rack",
+ "task_dir": "abc-sim-dish-rack-i00-privileged",
+ "source": "agent-success:codex-gpt6_luna-xhigh-abc-sim-dish-rack-i00-privileged/abc-sim-dish-rack-i00-privileged__aKL9b6V",
+ "agent": "codex + GPT-6 Luna (xhigh)",
+ "run": "codex-gpt6_luna-xhigh",
+ "mode": "privileged",
+ "trial": "codex-gpt6_luna-xhigh-abc-sim-dish-rack-i00-privileged/abc-sim-dish-rack-i00-privileged__aKL9b6V",
+ "trial_choice": "GPT-6 Luna's counted success; privileged before standard (a privileged trial is the submitted trajectory, replayed after the run)",
+ "renderer": "Blender 4.5.3 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); 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": 648,
+ "seconds": 21.6,
+ "speed": 1.5,
+ "success_check": true,
+ "success_check_how": "privileged trial: the submitted trajectory (trajectory.npz) replayed from the task's frozen instance with the task's own tooling (rcb_abc.Session, as tests' grade.py replays it): success by the task's rule (upstream's evaluator held 30 control steps), the end-state hash equals both grading replays' hash (e94b0ef819675914), and a second plain replay in a fresh session (Session.replay) ends in the same hash",
+ "end_check": {
+  "success": true,
+  "rule": "event",
+  "reason": "success",
+  "steps": 863,
+  "success_run": 30,
+  "hold_steps": 30
+ },
+ "replay": {
+  "control_hz": 29.4118,
+  "control_dt_s": 0.034,
+  "n_actions": 863,
+  "control_steps_replayed": 863,
+  "sim_time_s": 29.342,
+  "frames_every_control_steps": 1.4706,
+  "final_state_sha": "e94b0ef819675914",
+  "plain_replay_state_sha": "e94b0ef819675914",
+  "replays_agree": true,
+  "instance_sha": "f008c6c643f75ca1",
+  "init_sha": "f008c6c643f75ca1",
+  "instance_ok": true,
+  "state_hashes_every_control_steps": 1000,
+  "trajectory_sha256": "b48fabf316c93acd20b4479261f27409f45a01e52db6ded514450c8015454406",
+  "mujoco": "3.6.0",
+  "recorded_grader_state_sha": "e94b0ef819675914"
+ },
+ "visual_changes": [
+  "studio floor and cove instead of MuJoCo's floor and the booth's three walls",
+  "the overhead camera gate (bar and camera) not drawn",
+  "the table top drawn as a solid light-wood block down to the floor, extended under the arms' base mounts (top face unchanged; the scene's table is a thin white slab)"
+ ],
+ "camera": {
+  "mode": "fixed",
+  "frames_on": "both arms and every object on the table",
+  "az_deg": 35.0,
+  "el_deg": 32.0,
+  "fovy_deg": 34.0,
+  "fill": 0.8,
+  "target": [
+   0.4782,
+   0.0163,
+   0.9157
+  ],
+  "distance_m": 1.51,
+  "orbit_deg": 10.0
+ },
+ "holds": {
+  "start_s": 0.5,
+  "end_s": 1.5,
+  "note": "first / last frame held, no extra simulation"
+ },
+ "image": "efe:abc-sim-0.1.0",
+ "encode": "libx264 preset slow CRF 16, yuv420p, bt709, faststart",
+ "poster": "last frame (solved state)",
+ "seconds_per_frame": 8.41,
+ "created": "2026-10-31T05:38:15-0500"
+}

masters/abc-sim/dish_rack/video.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5c904215b916baea14c610f8c0e0fbb4521f006f13aaed03dca5213f8282e536
+size 12167666

masters/abc-sim/drawer_search/meta.json

@@ -0,0 +1,80 @@
+{
+ "suite": "abc-sim",
+ "task": "drawer_search",
+ "task_dir": "abc-sim-drawer-search-i00-privileged",
+ "source": "agent-success:codex-gpt6_luna-xhigh-abc-sim-drawer-search-i00-privileged/abc-sim-drawer-search-i00-privil__anJwWjZ",
+ "agent": "codex + GPT-6 Luna (xhigh)",
+ "run": "codex-gpt6_luna-xhigh",
+ "mode": "privileged",
+ "trial": "codex-gpt6_luna-xhigh-abc-sim-drawer-search-i00-privileged/abc-sim-drawer-search-i00-privil__anJwWjZ",
+ "trial_choice": "GPT-6 Luna's counted success; privileged before standard (a privileged trial is the submitted trajectory, replayed after the run)",
+ "renderer": "Blender 4.5.3 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); 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": 682,
+ "seconds": 22.73,
+ "speed": 6.0,
+ "success_check": true,
+ "success_check_how": "privileged trial: the submitted trajectory (trajectory.npz) replayed from the task's frozen instance with the task's own tooling (rcb_abc.Session, as tests' grade.py replays it): success by the task's rule (upstream's evaluator held 30 control steps), the end-state hash equals both grading replays' hash (c774fb4ae8abfd70), and a second plain replay in a fresh session (Session.replay) ends in the same hash",
+ "end_check": {
+  "success": true,
+  "rule": "event",
+  "reason": "success",
+  "steps": 3653,
+  "success_run": 30,
+  "hold_steps": 30
+ },
+ "replay": {
+  "control_hz": 29.4118,
+  "control_dt_s": 0.034,
+  "n_actions": 3653,
+  "control_steps_replayed": 3653,
+  "sim_time_s": 124.202,
+  "frames_every_control_steps": 5.8824,
+  "final_state_sha": "c774fb4ae8abfd70",
+  "plain_replay_state_sha": "c774fb4ae8abfd70",
+  "replays_agree": true,
+  "instance_sha": "c2f19b21de544dd2",
+  "init_sha": "c2f19b21de544dd2",
+  "instance_ok": true,
+  "state_hashes_every_control_steps": 1000,
+  "trajectory_sha256": "b29470b6af4515c5b3a50bea07d69d118a5b5122d0966373950987fb09a03c85",
+  "mujoco": "3.6.0",
+  "recorded_grader_state_sha": "c774fb4ae8abfd70"
+ },
+ "visual_changes": [
+  "studio floor and cove instead of MuJoCo's floor and the booth's three walls",
+  "the overhead camera gate (bar and camera) not drawn",
+  "the table top drawn as a solid light-wood block down to the floor, extended under the arms' base mounts (top face unchanged; the scene's table is a thin white slab)",
+  "unused objects the scene parks far below or above the table (they fall freely and touch nothing) not drawn: drawer_search_distractor_01, drawer_search_distractor_02, drawer_search_distractor_03, drawer_search_distractor_04, drawer_search_distractor_05, drawer_search_distractor_06, drawer_search_distractor_07, drawer_search_distractor_08, drawer_search_distractor_09, drawer_search_distractor_10, drawer_search_distractor_11, drawer_search_distractor_13, drawer_search_distractor_16, drawer_search_distractor_18, drawer_search_distractor_19, drawer_search_distractor_20, drawer_search_distractor_21, drawer_search_distractor_22, drawer_search_distractor_23, drawer_search_distractor_24, drawer_search_distractor_25, drawer_search_target, drawer_stack_1, drawer_stack_2"
+ ],
+ "camera": {
+  "mode": "fixed",
+  "frames_on": "both arms and every object on the table",
+  "az_deg": 155.0,
+  "el_deg": 40.0,
+  "fovy_deg": 34.0,
+  "fill": 0.8,
+  "target": [
+   0.4354,
+   -0.0612,
+   1.0015
+  ],
+  "distance_m": 1.9168,
+  "orbit_deg": 10.0,
+  "note": "viewed from behind the arms: from the usual 3/4 front view the drawer stack hides the arms' work"
+ },
+ "holds": {
+  "start_s": 0.5,
+  "end_s": 1.5,
+  "note": "first / last frame held, no extra simulation"
+ },
+ "image": "efe:abc-sim-0.1.0",
+ "encode": "libx264 preset slow CRF 16, yuv420p, bt709, faststart",
+ "poster": "last frame (solved state)",
+ "seconds_per_frame": 9.77,
+ "created": "2026-10-31T06:49:04-0500"
+}

masters/abc-sim/drawer_search/video.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c403e25f68a5f07eec53878dc90f5eac60baa3698272be3a1c1719d81098e2a4
+size 9123286

masters/abc-sim/grab_clutter/meta.json

@@ -0,0 +1,78 @@
+{
+ "suite": "abc-sim",
+ "task": "grab_clutter",
+ "task_dir": "abc-sim-grab-clutter-i00-privileged",
+ "source": "agent-success:codex-gpt6_luna-xhigh-abc-sim-grab-clutter-i00-privileged/abc-sim-grab-clutter-i00-privile__gcbkAVs",
+ "agent": "codex + GPT-6 Luna (xhigh)",
+ "run": "codex-gpt6_luna-xhigh",
+ "mode": "privileged",
+ "trial": "codex-gpt6_luna-xhigh-abc-sim-grab-clutter-i00-privileged/abc-sim-grab-clutter-i00-privile__gcbkAVs",
+ "trial_choice": "GPT-6 Luna's counted success; privileged before standard (a privileged trial is the submitted trajectory, replayed after the run)",
+ "renderer": "Blender 4.5.3 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); 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": 425,
+ "seconds": 14.17,
+ "speed": 1.0,
+ "success_check": true,
+ "success_check_how": "privileged trial: the submitted trajectory (trajectory.npz) replayed from the task's frozen instance with the task's own tooling (rcb_abc.Session, as tests' grade.py replays it): success by the task's rule (upstream's evaluator held 30 control steps), the end-state hash equals both grading replays' hash (a51263df9f4d3c28), and a second plain replay in a fresh session (Session.replay) ends in the same hash",
+ "end_check": {
+  "success": true,
+  "rule": "event",
+  "reason": "success",
+  "steps": 357,
+  "success_run": 30,
+  "hold_steps": 30
+ },
+ "replay": {
+  "control_hz": 29.4118,
+  "control_dt_s": 0.034,
+  "n_actions": 357,
+  "control_steps_replayed": 357,
+  "sim_time_s": 12.138,
+  "frames_every_control_steps": 0.9804,
+  "final_state_sha": "a51263df9f4d3c28",
+  "plain_replay_state_sha": "a51263df9f4d3c28",
+  "replays_agree": true,
+  "instance_sha": "a2e0c4bffcd832d7",
+  "init_sha": "a2e0c4bffcd832d7",
+  "instance_ok": true,
+  "state_hashes_every_control_steps": 1000,
+  "trajectory_sha256": "bab6ad70e1b3ac7a85ccb8570fcee579f58858749cd843408e5491d15f7c3b65",
+  "mujoco": "3.6.0",
+  "recorded_grader_state_sha": "a51263df9f4d3c28"
+ },
+ "visual_changes": [
+  "studio floor and cove instead of MuJoCo's floor and the booth's three walls",
+  "the overhead camera gate (bar and camera) not drawn",
+  "the table top drawn as a solid light-wood block down to the floor, extended under the arms' base mounts (top face unchanged; the scene's table is a thin white slab)"
+ ],
+ "camera": {
+  "mode": "fixed",
+  "frames_on": "both arms and every object on the table",
+  "az_deg": 35.0,
+  "el_deg": 32.0,
+  "fovy_deg": 34.0,
+  "fill": 0.8,
+  "target": [
+   0.5113,
+   -0.0571,
+   0.8954
+  ],
+  "distance_m": 1.7274,
+  "orbit_deg": 10.0
+ },
+ "holds": {
+  "start_s": 0.5,
+  "end_s": 1.5,
+  "note": "first / last frame held, no extra simulation"
+ },
+ "image": "efe:abc-sim-0.1.0",
+ "encode": "libx264 preset slow CRF 16, yuv420p, bt709, faststart",
+ "poster": "last frame (solved state)",
+ "seconds_per_frame": 8.63,
+ "created": "2026-10-31T06:47:44-0500"
+}

masters/abc-sim/grab_clutter/video.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f87aa2e881d702e6e395cbde182f36bb76e61bef7537fed732b168bbe8c91019
+size 6726898

masters/abc-sim/hang_mug/meta.json

@@ -0,0 +1,78 @@
+{
+ "suite": "abc-sim",
+ "task": "hang_mug",
+ "task_dir": "abc-sim-hang-mug-i00-privileged",
+ "source": "agent-success:codex-gpt6_luna-xhigh-abc-sim-hang-mug-i00-privileged/abc-sim-hang-mug-i00-privileged__eEKwPb7",
+ "agent": "codex + GPT-6 Luna (xhigh)",
+ "run": "codex-gpt6_luna-xhigh",
+ "mode": "privileged",
+ "trial": "codex-gpt6_luna-xhigh-abc-sim-hang-mug-i00-privileged/abc-sim-hang-mug-i00-privileged__eEKwPb7",
+ "trial_choice": "GPT-6 Luna's counted success; privileged before standard (a privileged trial is the submitted trajectory, replayed after the run)",
+ "renderer": "Blender 4.5.3 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); 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": 609,
+ "seconds": 20.3,
+ "speed": 1.5,
+ "success_check": true,
+ "success_check_how": "privileged trial: the submitted trajectory (trajectory.npz) replayed from the task's frozen instance with the task's own tooling (rcb_abc.Session, as tests' grade.py replays it): success by the task's rule (upstream's evaluator held 30 control steps), the end-state hash equals both grading replays' hash (59e68561c05dcde6), and a second plain replay in a fresh session (Session.replay) ends in the same hash",
+ "end_check": {
+  "success": true,
+  "rule": "event",
+  "reason": "success",
+  "steps": 806,
+  "success_run": 30,
+  "hold_steps": 30
+ },
+ "replay": {
+  "control_hz": 29.4118,
+  "control_dt_s": 0.034,
+  "n_actions": 806,
+  "control_steps_replayed": 806,
+  "sim_time_s": 27.404,
+  "frames_every_control_steps": 1.4706,
+  "final_state_sha": "59e68561c05dcde6",
+  "plain_replay_state_sha": "59e68561c05dcde6",
+  "replays_agree": true,
+  "instance_sha": "f971caba9bc2c81c",
+  "init_sha": "f971caba9bc2c81c",
+  "instance_ok": true,
+  "state_hashes_every_control_steps": 1000,
+  "trajectory_sha256": "13854835a1ace62a05cdd7fe130b118676d185a2491d4d898cc17f7519facded",
+  "mujoco": "3.6.0",
+  "recorded_grader_state_sha": "59e68561c05dcde6"
+ },
+ "visual_changes": [
+  "studio floor and cove instead of MuJoCo's floor and the booth's three walls",
+  "the overhead camera gate (bar and camera) not drawn",
+  "the table top drawn as a solid light-wood block down to the floor, extended under the arms' base mounts (top face unchanged; the scene's table is a thin white slab)"
+ ],
+ "camera": {
+  "mode": "fixed",
+  "frames_on": "both arms and every object on the table",
+  "az_deg": 35.0,
+  "el_deg": 32.0,
+  "fovy_deg": 34.0,
+  "fill": 0.8,
+  "target": [
+   0.4312,
+   0.0217,
+   0.9232
+  ],
+  "distance_m": 1.465,
+  "orbit_deg": 10.0
+ },
+ "holds": {
+  "start_s": 0.5,
+  "end_s": 1.5,
+  "note": "first / last frame held, no extra simulation"
+ },
+ "image": "efe:abc-sim-0.1.0",
+ "encode": "libx264 preset slow CRF 16, yuv420p, bt709, faststart",
+ "poster": "last frame (solved state)",
+ "seconds_per_frame": 4.52,
+ "created": "2026-10-31T04:29:10-0500"
+}

masters/abc-sim/hang_mug/video.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3fa4180e2e79bf2a6ddc22559fc7ae05f064ba783021c4cdb929265011a18c3b
+size 10073096

masters/abc-sim/highlights.json

@@ -0,0 +1,20 @@
+{
+ "teaser": "lego",
+ "highlights": [
+  "lego",
+  "nuts_bolts",
+  "hang_mug",
+  "chess",
+  "throw_bottles",
+  "pouring"
+ ],
+ "why": {
+  "lego": "both arms sort colour bricks into two bins, then one bin is tipped into the tote",
+  "nuts_bolts": "the hardest task: no GPT-6 Luna success in any mode; GPT-6 Astra's standard success, both arms sorting nuts and bolts",
+  "hang_mug": "a mug handed from one arm to the other and hung on the tree",
+  "chess": "pieces picked off the board and placed with both arms",
+  "throw_bottles": "bottles picked up and dropped into the bin",
+  "pouring": "pouring from one cup into the other"
+ },
+ "why_fewer": ""
+}

masters/abc-sim/inhand_transfer/meta.json

@@ -0,0 +1,78 @@
+{
+ "suite": "abc-sim",
+ "task": "inhand_transfer",
+ "task_dir": "abc-sim-inhand-transfer-i00-privileged",
+ "source": "agent-success:codex-gpt6_luna-xhigh-abc-sim-inhand-transfer-i00-privileged/abc-sim-inhand-transfer-i00-priv__bB2rjgH",
+ "agent": "codex + GPT-6 Luna (xhigh)",
+ "run": "codex-gpt6_luna-xhigh",
+ "mode": "privileged",
+ "trial": "codex-gpt6_luna-xhigh-abc-sim-inhand-transfer-i00-privileged/abc-sim-inhand-transfer-i00-priv__bB2rjgH",
+ "trial_choice": "GPT-6 Luna's counted success; privileged before standard (a privileged trial is the submitted trajectory, replayed after the run)",
+ "renderer": "Blender 4.5.3 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); 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": 285,
+ "seconds": 9.5,
+ "speed": 1.0,
+ "success_check": true,
+ "success_check_how": "privileged trial: the submitted trajectory (trajectory.npz) replayed from the task's frozen instance with the task's own tooling (rcb_abc.Session, as tests' grade.py replays it): success by the task's rule (upstream's evaluator held 30 control steps), the end-state hash equals both grading replays' hash (3ac4b107f71cbcb8), and a second plain replay in a fresh session (Session.replay) ends in the same hash",
+ "end_check": {
+  "success": true,
+  "rule": "event",
+  "reason": "success",
+  "steps": 220,
+  "success_run": 30,
+  "hold_steps": 30
+ },
+ "replay": {
+  "control_hz": 29.4118,
+  "control_dt_s": 0.034,
+  "n_actions": 220,
+  "control_steps_replayed": 220,
+  "sim_time_s": 7.48,
+  "frames_every_control_steps": 0.9804,
+  "final_state_sha": "3ac4b107f71cbcb8",
+  "plain_replay_state_sha": "3ac4b107f71cbcb8",
+  "replays_agree": true,
+  "instance_sha": "2fee35e343f1e9d5",
+  "init_sha": "2fee35e343f1e9d5",
+  "instance_ok": true,
+  "state_hashes_every_control_steps": 1000,
+  "trajectory_sha256": "fd5659868b6c420f89dbb0d3454dd6da3b10a1edcf8895b0bfbc016c31772043",
+  "mujoco": "3.6.0",
+  "recorded_grader_state_sha": "3ac4b107f71cbcb8"
+ },
+ "visual_changes": [
+  "studio floor and cove instead of MuJoCo's floor and the booth's three walls",
+  "the overhead camera gate (bar and camera) not drawn",
+  "the table top drawn as a solid light-wood block down to the floor, extended under the arms' base mounts (top face unchanged; the scene's table is a thin white slab)"
+ ],
+ "camera": {
+  "mode": "fixed",
+  "frames_on": "both arms and every object on the table",
+  "az_deg": 35.0,
+  "el_deg": 32.0,
+  "fovy_deg": 34.0,
+  "fill": 0.8,
+  "target": [
+   0.4653,
+   0.0452,
+   0.9364
+  ],
+  "distance_m": 1.4135,
+  "orbit_deg": 10.0
+ },
+ "holds": {
+  "start_s": 0.5,
+  "end_s": 1.5,
+  "note": "first / last frame held, no extra simulation"
+ },
+ "image": "efe:abc-sim-0.1.0",
+ "encode": "libx264 preset slow CRF 16, yuv420p, bt709, faststart",
+ "poster": "last frame (solved state)",
+ "seconds_per_frame": 8.4,
+ "created": "2026-10-31T06:15:20-0500"
+}

masters/abc-sim/inhand_transfer/video.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b97d9b9772f8ad840b744dff686c68005f18e54e2124c0ba09f35b84e5f1c5c6
+size 5183804

masters/abc-sim/lego/meta.json

@@ -0,0 +1,78 @@
+{
+ "suite": "abc-sim",
+ "task": "lego",
+ "task_dir": "abc-sim-lego-i00-privileged",
+ "source": "agent-success:codex-gpt6_luna-xhigh-abc-sim-lego-i00-privileged/abc-sim-lego-i00-privileged__ku7HjL5",
+ "agent": "codex + GPT-6 Luna (xhigh)",
+ "run": "codex-gpt6_luna-xhigh",
+ "mode": "privileged",
+ "trial": "codex-gpt6_luna-xhigh-abc-sim-lego-i00-privileged/abc-sim-lego-i00-privileged__ku7HjL5",
+ "trial_choice": "GPT-6 Luna's counted success; privileged before standard (a privileged trial is the submitted trajectory, replayed after the run)",
+ "renderer": "Blender 4.5.3 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); 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": 589,
+ "seconds": 19.63,
+ "speed": 20.0,
+ "success_check": true,
+ "success_check_how": "privileged trial: the submitted trajectory (trajectory.npz) replayed from the task's frozen instance with the task's own tooling (rcb_abc.Session, as tests' grade.py replays it): success by the task's rule (upstream's evaluator held 30 control steps), the end-state hash equals both grading replays' hash (96b18189a5afadc2), and a second plain replay in a fresh session (Session.replay) ends in the same hash",
+ "end_check": {
+  "success": true,
+  "rule": "event",
+  "reason": "success",
+  "steps": 10334,
+  "success_run": 30,
+  "hold_steps": 30
+ },
+ "replay": {
+  "control_hz": 29.4118,
+  "control_dt_s": 0.034,
+  "n_actions": 10334,
+  "control_steps_replayed": 10334,
+  "sim_time_s": 351.356,
+  "frames_every_control_steps": 19.6078,
+  "final_state_sha": "96b18189a5afadc2",
+  "plain_replay_state_sha": "96b18189a5afadc2",
+  "replays_agree": true,
+  "instance_sha": "2ebc48bf64974f98",
+  "init_sha": "2ebc48bf64974f98",
+  "instance_ok": true,
+  "state_hashes_every_control_steps": 1000,
+  "trajectory_sha256": "3974f80dec008fb9855d149fd58b32bba9ae9629b1fe41e702c4cb966998267f",
+  "mujoco": "3.6.0",
+  "recorded_grader_state_sha": "96b18189a5afadc2"
+ },
+ "visual_changes": [
+  "studio floor and cove instead of MuJoCo's floor and the booth's three walls",
+  "the overhead camera gate (bar and camera) not drawn",
+  "the table top drawn as a solid light-wood block down to the floor, extended under the arms' base mounts (top face unchanged; the scene's table is a thin white slab)"
+ ],
+ "camera": {
+  "mode": "fixed",
+  "frames_on": "both arms and every object on the table",
+  "az_deg": 35.0,
+  "el_deg": 32.0,
+  "fovy_deg": 34.0,
+  "fill": 0.8,
+  "target": [
+   0.4871,
+   0.0341,
+   0.9031
+  ],
+  "distance_m": 1.7469,
+  "orbit_deg": 10.0
+ },
+ "holds": {
+  "start_s": 0.5,
+  "end_s": 1.5,
+  "note": "first / last frame held, no extra simulation"
+ },
+ "image": "efe:abc-sim-0.1.0",
+ "encode": "libx264 preset slow CRF 16, yuv420p, bt709, faststart",
+ "poster": "last frame (solved state)",
+ "seconds_per_frame": 2.93,
+ "created": "2026-10-31T04:08:17-0500"
+}

masters/abc-sim/lego/video.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dd0a43fb19da550869df6db66138c5bfc3c132e8babf8d845989a660948c6f10
+size 9892185

masters/abc-sim/nuts_bolts/meta.json

@@ -0,0 +1,79 @@
+{
+ "suite": "abc-sim",
+ "task": "nuts_bolts",
+ "task_dir": "abc-sim-nuts-bolts-i00-standard",
+ "source": "agent-success:codex-gpt6_astra-high-abc-sim-nuts-bolts-i00-standard/abc-sim-nuts-bolts-i00-standard__KAo3uKg",
+ "agent": "codex + GPT-6 Astra (high)",
+ "run": "codex-gpt6_astra-high",
+ "mode": "standard",
+ "trial": "codex-gpt6_astra-high-abc-sim-nuts-bolts-i00-standard/abc-sim-nuts-bolts-i00-standard__KAo3uKg",
+ "trial_choice": "no privileged success on this task; GPT-6 Luna failed it in every mode, and a counted success by another agent is shown before the reference solution: GPT-6 Astra's standard trial, re-rendered because its replay is bit-exact",
+ "renderer": "Blender 4.5.3 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); 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": 713,
+ "seconds": 23.77,
+ "speed": 30.0,
+ "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 (the sim service's trajectory.npz) were replayed from the task's frozen instance (byte-identical in the privileged and standard directories) with the task's own tooling (rcb_abc.Session, the session the sim service runs): success by the task's rule (upstream's evaluator held 30 control steps), the end-state hash equals the live episode's recorded hash (539fdec941c59936), and a second plain replay in a fresh session (Session.replay) ends in the same hash",
+ "end_check": {
+  "success": true,
+  "rule": "event",
+  "reason": "success",
+  "steps": 19156,
+  "success_run": 30,
+  "hold_steps": 30
+ },
+ "replay": {
+  "control_hz": 29.4118,
+  "control_dt_s": 0.034,
+  "n_actions": 19156,
+  "control_steps_replayed": 19156,
+  "sim_time_s": 651.304,
+  "frames_every_control_steps": 29.4118,
+  "final_state_sha": "539fdec941c59936",
+  "plain_replay_state_sha": "539fdec941c59936",
+  "replays_agree": true,
+  "instance_sha": "36a10f65fd90f73b",
+  "init_sha": "36a10f65fd90f73b",
+  "instance_ok": true,
+  "state_hashes_every_control_steps": 1000,
+  "trajectory_sha256": "231517cc43bb082a8a0f32dc7654f5f03fa99d004a54d3dfc5c1ba4a37dae716",
+  "mujoco": "3.6.0",
+  "recorded_live_state_sha": "539fdec941c59936",
+  "matches_live_episode": true
+ },
+ "visual_changes": [
+  "studio floor and cove instead of MuJoCo's floor and the booth's three walls",
+  "the overhead camera gate (bar and camera) not drawn",
+  "the table top drawn as a solid light-wood block down to the floor, extended under the arms' base mounts (top face unchanged; the scene's table is a thin white slab)"
+ ],
+ "camera": {
+  "mode": "fixed",
+  "frames_on": "both arms and every object on the table",
+  "az_deg": 35.0,
+  "el_deg": 32.0,
+  "fovy_deg": 34.0,
+  "fill": 0.8,
+  "target": [
+   0.4552,
+   -0.0151,
+   0.9698
+  ],
+  "distance_m": 2.0078,
+  "orbit_deg": 10.0
+ },
+ "holds": {
+  "start_s": 0.5,
+  "end_s": 1.5,
+  "note": "first / last frame held, no extra simulation"
+ },
+ "image": "efe:abc-sim-0.2.0",
+ "encode": "libx264 preset slow CRF 16, yuv420p, bt709, faststart",
+ "poster": "last frame (solved state)",
+ "seconds_per_frame": 8.55,
+ "created": "2026-10-31T07:22:47-0500"
+}

masters/abc-sim/nuts_bolts/video.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c384a6ded59bc4e632fd86ea8e15c557aa31e09bfd63d281a3d8da2bfc8ab2d8
+size 10853677

masters/abc-sim/pouring/meta.json

@@ -0,0 +1,78 @@
+{
+ "suite": "abc-sim",
+ "task": "pouring",
+ "task_dir": "abc-sim-pouring-i00-privileged",
+ "source": "agent-success:codex-gpt6_luna-xhigh-abc-sim-pouring-i00-privileged/abc-sim-pouring-i00-privileged__oa56369",
+ "agent": "codex + GPT-6 Luna (xhigh)",
+ "run": "codex-gpt6_luna-xhigh",
+ "mode": "privileged",
+ "trial": "codex-gpt6_luna-xhigh-abc-sim-pouring-i00-privileged/abc-sim-pouring-i00-privileged__oa56369",
+ "trial_choice": "GPT-6 Luna's counted success; privileged before standard (a privileged trial is the submitted trajectory, replayed after the run)",
+ "renderer": "Blender 4.5.3 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); 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": 632,
+ "seconds": 21.07,
+ "speed": 5.0,
+ "success_check": true,
+ "success_check_how": "privileged trial: the submitted trajectory (trajectory.npz) replayed from the task's frozen instance with the task's own tooling (rcb_abc.Session, as tests' grade.py replays it): success by the task's rule (upstream's evaluator held 30 control steps), the end-state hash equals both grading replays' hash (ebbb534731924b41), and a second plain replay in a fresh session (Session.replay) ends in the same hash",
+ "end_check": {
+  "success": true,
+  "rule": "event",
+  "reason": "success",
+  "steps": 2798,
+  "success_run": 30,
+  "hold_steps": 30
+ },
+ "replay": {
+  "control_hz": 29.4118,
+  "control_dt_s": 0.034,
+  "n_actions": 2798,
+  "control_steps_replayed": 2798,
+  "sim_time_s": 95.132,
+  "frames_every_control_steps": 4.902,
+  "final_state_sha": "ebbb534731924b41",
+  "plain_replay_state_sha": "ebbb534731924b41",
+  "replays_agree": true,
+  "instance_sha": "b54e0a7dafa2a18d",
+  "init_sha": "b54e0a7dafa2a18d",
+  "instance_ok": true,
+  "state_hashes_every_control_steps": 1000,
+  "trajectory_sha256": "16e54616c9e2a467da108ffd390297070353e526bc8bd3136daa8a26eaea4719",
+  "mujoco": "3.6.0",
+  "recorded_grader_state_sha": "ebbb534731924b41"
+ },
+ "visual_changes": [
+  "studio floor and cove instead of MuJoCo's floor and the booth's three walls",
+  "the overhead camera gate (bar and camera) not drawn",
+  "the table top drawn as a solid light-wood block down to the floor, extended under the arms' base mounts (top face unchanged; the scene's table is a thin white slab)"
+ ],
+ "camera": {
+  "mode": "fixed",
+  "frames_on": "both arms and every object on the table",
+  "az_deg": 35.0,
+  "el_deg": 32.0,
+  "fovy_deg": 34.0,
+  "fill": 0.8,
+  "target": [
+   0.4651,
+   0.0504,
+   0.9745
+  ],
+  "distance_m": 1.3681,
+  "orbit_deg": 10.0
+ },
+ "holds": {
+  "start_s": 0.5,
+  "end_s": 1.5,
+  "note": "first / last frame held, no extra simulation"
+ },
+ "image": "efe:abc-sim-0.1.0",
+ "encode": "libx264 preset slow CRF 16, yuv420p, bt709, faststart",
+ "poster": "last frame (solved state)",
+ "seconds_per_frame": 2.78,
+ "created": "2026-10-31T05:14:26-0500"
+}

masters/abc-sim/pouring/video.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:808095474c92128cd8c86a3ec0b4f91c443d62c8e990796c03dd9eeeaf862c81
+size 10052972

masters/abc-sim/put_bottles/meta.json

@@ -0,0 +1,78 @@
+{
+ "suite": "abc-sim",
+ "task": "put_bottles",
+ "task_dir": "abc-sim-put-bottles-i00-privileged",
+ "source": "agent-success:codex-gpt6_luna-xhigh-abc-sim-put-bottles-i00-privileged/abc-sim-put-bottles-i00-privileg__HHK8Jsa",
+ "agent": "codex + GPT-6 Luna (xhigh)",
+ "run": "codex-gpt6_luna-xhigh",
+ "mode": "privileged",
+ "trial": "codex-gpt6_luna-xhigh-abc-sim-put-bottles-i00-privileged/abc-sim-put-bottles-i00-privileg__HHK8Jsa",
+ "trial_choice": "GPT-6 Luna's counted success; privileged before standard (a privileged trial is the submitted trajectory, replayed after the run)",
+ "renderer": "Blender 4.5.3 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); 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": 621,
+ "seconds": 20.7,
+ "speed": 3.0,
+ "success_check": true,
+ "success_check_how": "privileged trial: the submitted trajectory (trajectory.npz) replayed from the task's frozen instance with the task's own tooling (rcb_abc.Session, as tests' grade.py replays it): success by the task's rule (upstream's evaluator held 30 control steps), the end-state hash equals both grading replays' hash (f6857b8e253c4485), and a second plain replay in a fresh session (Session.replay) ends in the same hash",
+ "end_check": {
+  "success": true,
+  "rule": "event",
+  "reason": "success",
+  "steps": 1647,
+  "success_run": 30,
+  "hold_steps": 30
+ },
+ "replay": {
+  "control_hz": 29.4118,
+  "control_dt_s": 0.034,
+  "n_actions": 1647,
+  "control_steps_replayed": 1647,
+  "sim_time_s": 55.998,
+  "frames_every_control_steps": 2.9412,
+  "final_state_sha": "f6857b8e253c4485",
+  "plain_replay_state_sha": "f6857b8e253c4485",
+  "replays_agree": true,
+  "instance_sha": "9caab7a4776d5af8",
+  "init_sha": "9caab7a4776d5af8",
+  "instance_ok": true,
+  "state_hashes_every_control_steps": 1000,
+  "trajectory_sha256": "2bfeac5067246d77a1d2cb4eaf56de87aac1d06ba6611626240af22fc889f0e2",
+  "mujoco": "3.6.0",
+  "recorded_grader_state_sha": "f6857b8e253c4485"
+ },
+ "visual_changes": [
+  "studio floor and cove instead of MuJoCo's floor and the booth's three walls",
+  "the overhead camera gate (bar and camera) not drawn",
+  "the table top drawn as a solid light-wood block down to the floor, extended under the arms' base mounts (top face unchanged; the scene's table is a thin white slab)"
+ ],
+ "camera": {
+  "mode": "fixed",
+  "frames_on": "both arms and every object on the table",
+  "az_deg": 35.0,
+  "el_deg": 32.0,
+  "fovy_deg": 34.0,
+  "fill": 0.8,
+  "target": [
+   0.4334,
+   0.0605,
+   0.9094
+  ],
+  "distance_m": 1.7387,
+  "orbit_deg": 10.0
+ },
+ "holds": {
+  "start_s": 0.5,
+  "end_s": 1.5,
+  "note": "first / last frame held, no extra simulation"
+ },
+ "image": "efe:abc-sim-0.1.0",
+ "encode": "libx264 preset slow CRF 16, yuv420p, bt709, faststart",
+ "poster": "last frame (solved state)",
+ "seconds_per_frame": 4.63,
+ "created": "2026-10-31T05:14:30-0500"
+}