pandadyn s09_toss_recatch_delivery: master = Claude Opus 5.5 privileged success (replaces the oracle render)

masters/pandadyn/s09_toss_recatch_delivery/meta.json

@@ -1,14 +1,29 @@
 {
  "suite": "pandadyn",
  "task": "s09_toss_recatch_delivery",
- "source": "oracle",
- "renderer": "SAPIEN 3.0.3 ray tracer (ManiSkill 3.0.1 rt shader pack) spp=32 path_depth=8, OIDN 2.0.1 denoiser on CUDA",
+ "task_dir": "pandadyn-s09-toss-recatch-delivery-i00-privileged",
+ "source": "agent-success:claude_code-opus_5_5-high-pandadyn-s09-toss-recatch-delivery-i00-privileged/pandadyn-s09-toss-recatch-delive__P9UV6sR",
+ "agent": "Claude Code + Claude Opus 5.5 (high)",
+ "run": "claude_code-opus_5_5-high",
+ "mode": "privileged",
+ "trial": "P9UV6sR",
+ "trial_choice": "a counted privileged success of a ranked agent (Claude Opus 5.5), replacing the reference-solution (oracle) video",
+ "renderer": "SAPIEN 3.0.3 'rt' ray-tracing shader pack (ManiSkill 3.0.1), spp=32, path_depth=8, OIDN denoiser, rendered natively at 2560x1440 from one added static camera (eye [0.95,-1.10,0.85], look [0.03,-0.02,0.40], fovy 43 deg); the environment's own lights (ambient 0.45 + shadowed directional), as the suite's other masters",
  "rendering": "ray-tracing",
  "width": 2560,
  "height": 1440,
  "fps": 30,
+ "frames": 445,
+ "seconds": 14.83,
  "speed": 1.0,
+ "speed_note": "real time: 30 fps sampled on exact 1 kHz physics substeps round(k*1000/30); 415 episode frames + 30 frames (1.0 s) holding the end state",
+ "shows": "the whole graded episode, control steps 0-1381 (0-13.81 s sim): catch, self-toss, recatch, delivery into the moving box",
  "success_check": true,
- "image": "eai-pandadyn:0.1.0",
- "note": "the task's frozen oracle rows replayed exactly as the grader replays them (CPU PhysX 1 kHz); frames sampled on exact physics substeps; the replay ended in the frozen oracle end state (3f8dccf95457), success at control step 1241 (12.41 s); the last frame is held 1.0 s"
+ "success_check_how": "the trial's submitted trajectory (1381 rows) replayed from the frozen instance with the task's own tests/rcb_pandadyn.py, physics unchanged (camera added only); the replay ends in success with the recorded executed rows, phase 5 and end reason 'success' (reward.json / grade.log); the privileged grader records no state hash, so the rendered pass was checked against an unrendered replay: both end in the same final state",
+ "replay": {"instance_matches": true, "start_sha": "f5d85574581dd770", "final_sha": "1c32706e0fc948e2", "executed_rows": 1381, "sim_time_s": 13.81, "end_reason": "success"},
+ "task_version_note": "rendered in efe:pandadyn-0.1.0; the task moved to efe:pandadyn-0.2.0 on 2026-10-23 with the same frozen instance (initial_sha f5d85574581dd770) and the same scene and goal",
+ "image": "efe:pandadyn-0.1.0",
+ "encode": "libx264 High yuv420p CRF 16 preset slow, 30 fps constant, bt709, faststart, no audio",
+ "poster": "last frame (the ball delivered into the box), full resolution PNG",
+ "created": "2026-10-19"
 }

masters/pandadyn/s09_toss_recatch_delivery/video.mp4

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:02f97de0ca1395da1af570b80d23b6478c1b218ee55350c26c1b3aef2386f70c
-size 3631011
+oid sha256:0bf87edde7e8baa301ce16c076559d9d1c65bd817833308975f292b1224d5f6d
+size 4217009