{
 "suite": "dexhand",
 "task": "usb_plug_insert",
 "family": "usb_plug_insert",
 "source": "oracle",
 "renderer": "Blender 4.5.3 LTS Cycles path tracing, OptiX on an RTX 4090, adaptive sampling (threshold 0.01, 64-192 spp), OpenImageDenoise with albedo and normal passes, AgX Medium High Contrast; studio floor, key/top/rim area lights and a dim sky dome; render 2560x1440; every frame is the exact pose sample recorded from the task's own replay.py (MuJoCo 3.4.0, 120 Hz simulated-time grid, nearest sample, no interpolation)",
 "rendering": "path-tracing",
 "width": 2560,
 "height": 1440,
 "fps": 30,
 "frames": 666,
 "duration_s": 22.2,
 "speed": 1.5,
 "success_check": true,
 "success_check_how": "verdict of the task's own tests/replay.py (the verifier's replay: fresh Session, frozen instance, fail-fast checks every physics substep) run on the oracle archive written by solution/oracle.py in the pinned grader image while the poses were recorded: success=True, reason='success'; a second, unrecorded replay in a fresh process ended in the same state hash",
 "replay": {
  "hand": "sharpa",
  "n_actions": 1595,
  "sim_steps": 1475,
  "physics_steps": 14750,
  "sim_time_s": 29.5,
  "state_sha": "c18aa345626c7752fdbcea593086867c0cc2b06c7320996c1c709af4695d265c",
  "state_sha_unrecorded_replay": "c18aa345626c7752fdbcea593086867c0cc2b06c7320996c1c709af4695d265c",
  "trajectory_sha256": "b33af3052c7f3fa95fbe09c48b4a3e5888f1f961a201ae85b7c791cdff1273f4",
  "task_dir": "tasks/dexhand-usb-plug-insert-i00-privileged (PR #241)"
 },
 "holds": {
  "start_s": 1.0,
  "end_s": 1.5,
  "note": "first and last recorded states held; the camera keeps moving, no extra simulation"
 },
 "camera": {
  "az_deg": 68.7,
  "el_deg": 30,
  "lens_mm": 50,
  "orbit_deg": 10.0,
  "fill": 0.8,
  "target": [
   0.8568,
   -0.2283,
   0.8624
  ],
  "distance_m": 2.4359
 },
 "poster_frame": 665,
 "image": "eai-dexhand:0.2.0@sha256:0d42994b83d95ab5301c592b35586550c9ee426490115e4b43f3b8cf8cc43350",
 "created": "2026-10-17T13:25:03Z"
}