{
  "tasks": [
    "stool",
    "drawer"
  ],
  "rounds": 8,
  "steps": 5000,
  "qos": "extra",
  "model": "current native Codex session",
  "backend": "native_session",
  "paid_api_allowed": false,
  "dev_seeds": [
    3101,
    3102
  ],
  "test_seeds": [
    9101,
    9102,
    9103,
    9104,
    9105,
    9106,
    9107,
    9108,
    9109,
    9110
  ],
  "benchmark_sha256": "bb239bdba6da79a8e609bdd9b42a6caba7ce530e45ee0b54394ae61eaf127c12",
  "harness_sources": {
    "check_fixture.py": "7caa075c28c6164626a27a78036e8c494b705796585cba8c9bb5818101850972",
    "evaluate.py": "d5ea74efc9dd1c7b0e8fd4803636e5f104bf9cc717e09ab7118fcf454ddd8b7f",
    "policy_host.py": "0e505285b2d14b93970cd48fd31c3c222b65925a61b00065cba7a6ed88d3d120",
    "probe_policy.py": "ad30ab24fd6800ed1987349d796d69169f4280a50747c5a3661c96a6ea47e426",
    "render_report.py": "ce481393701981cfe61740120d0183897296605615dd34b7a92a8bd0ef541b69",
    "resume.py": "caa21c10fe9df1cd2e4418889ef962b1f0b3047aae59e4dbee05c1fff29d8bb6",
    "run.py": "13cb305bf961cf75c5c39080c8cb363abbf5b2a580bb9a3232d802904ee26da6",
    "servo.py": "a602aa85a93d8c3c03f87043f92343392312632e2aefb09808475c7910acd58c",
    "summarize.py": "4aaaf0b3f0771abfbe009438e182d757e3bf7c25e1a943d2f49ca3d034cf25db",
    "check_fixture.sh": "d6867ee1006eb8c4b226c9b811d6aae6e865d7cc2596dd9ca328075cf338ce79",
    "evaluate.sh": "50cd70680aecc0095a2d5042ae6827a33be752525778b66e7236182da934a41d"
  },
  "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"
}