{
  "tasks": [
    "stool",
    "drawer"
  ],
  "rounds": 2,
  "steps": 2500,
  "qos": "own",
  "model": "current native Codex session",
  "backend": "native_session",
  "paid_api_allowed": false,
  "dev_seeds": [
    3101,
    3102,
    3103
  ],
  "test_seeds": [
    8101,
    8102,
    8103,
    8104,
    8105,
    8106,
    8107,
    8108,
    8109,
    8110
  ],
  "benchmark_sha256": "bb239bdba6da79a8e609bdd9b42a6caba7ce530e45ee0b54394ae61eaf127c12",
  "harness_sources": {
    "check_fixture.py": "ed5456d206f5e499c8b21e6e4bad1b4ea630f516c6f930e997c35114e3183c98",
    "evaluate.py": "68873e4d76710303467480132410a904449b6c01d48804524e8c648263aaa50c",
    "policy_host.py": "0e505285b2d14b93970cd48fd31c3c222b65925a61b00065cba7a6ed88d3d120",
    "probe_policy.py": "ad30ab24fd6800ed1987349d796d69169f4280a50747c5a3661c96a6ea47e426",
    "run.py": "da807b24f6934eb28f9667e64612ce0b6c55c4c047391d52c61b492b3f51c642",
    "servo.py": "a602aa85a93d8c3c03f87043f92343392312632e2aefb09808475c7910acd58c",
    "summarize.py": "4aaaf0b3f0771abfbe009438e182d757e3bf7c25e1a943d2f49ca3d034cf25db",
    "check_fixture.sh": "2e0a565a235d86066ce5f6546e99be37a8354db4fae9391af1a89ffe18c97bf7",
    "evaluate.sh": "2df23b45a19058ea478f5cf0cd690f99e10c9c39dc7042037a541ff860dc562c"
  },
  "observation_setting": "privileged ground-truth poses + scorer-derived assembly design",
  "initial_candidate_origin": "native session",
  "human_intervention": "user sets goals/constraints; active session supplies policy revisions via recorded responses"
}