{
  "schema_version": 1,
  "generated_at": "2026-06-05T13:07:24.934179Z",
  "checkpoint": {
    "repo_id": "abdelstark/lensemble-phase2-so100-checkpoint",
    "repo_type": "model",
    "revision": "da52ef380ac87317c89e87f048d65bae65c16b9e",
    "artifact_path": "artifacts/round-00003",
    "checkpoint_hash": "8f1494fd9e57b7496daf96e379a3de1457a435080b81b9e0ea1d20a52f4827c4",
    "training_job_id": "6a22ba68e6aa50b87b9ebef7",
    "training_job_url": "https://huggingface.co/jobs/abdelstark/6a22ba68e6aa50b87b9ebef7",
    "code_sha": "4b446a558882f25e47ee6410a4c32982bbf33477",
    "train_config_hash": "82296109ed452a1aaf494306b100bd4f2b7e3b968a8d4219bc89a3461b0294a3"
  },
  "eval_report": {
    "schema_version": 1,
    "checkpoint_hash": "8f1494fd9e57b7496daf96e379a3de1457a435080b81b9e0ea1d20a52f4827c4",
    "env_id": "synthetic://toy",
    "planner": "icem",
    "success_rate": 0.5,
    "planning_samples": 1,
    "time_per_action_ms": 91.15450602257624,
    "effective_dim": 1.0000066342911489,
    "probe_accuracy": null,
    "run_manifest_hash": "7092bdda71bb1c01510ce8486e2e99cd3065dfed7834faa3235e6d7a2c7c17fb"
  },
  "task": {
    "env_id": "synthetic://toy",
    "task_scale": "synthetic downstream harness over a published Phase 2 checkpoint",
    "held_out_policy": "synthetic://toy is deterministic and not drawn from the SO-100 training silos; the harness evaluates the configured --num-episodes seed-pinned episodes derived from root_seed.",
    "goal_policy": "fixed synthetic goal clip generated by the built-in synthetic://toy world with seed 7919; no raw goal observation is serialized.",
    "n_episodes": 2,
    "action_kind": "continuous",
    "action_units": [
      "u",
      "u"
    ],
    "raw_data_in_report": false
  },
  "planner_budget": {
    "planner": "icem",
    "horizon": 1,
    "planning_samples": 1,
    "planner_iterations": 1,
    "action_dim": 2,
    "action_low": [
      -1.0,
      -1.0
    ],
    "action_high": [
      1.0,
      1.0
    ],
    "action_clipping": "none in the current Planner path; sampled Gaussian actions are passed to the fresh local ActionHead, while ActionSpec bounds are recorded for audit."
  },
  "eval_config_hash": "f926ca5d1f230c960b6c10810a7f42620e99a05f144ac2a45991541940d014d4",
  "eval_command": "uv run --extra dev python scripts/phase2_eval_checkpoint.py --checkpoint-repo abdelstark/lensemble-phase2-so100-checkpoint --checkpoint-revision da52ef380ac87317c89e87f048d65bae65c16b9e --checkpoint-artifact-path artifacts/round-00003 --download-dir /tmp/lensemble-phase2-checkpoint --output /tmp/phase2_downstream_eval_report.json --expected-checkpoint-hash 8f1494fd9e57b7496daf96e379a3de1457a435080b81b9e0ea1d20a52f4827c4 --env-id synthetic://toy --planner icem --planning-samples 1 --horizon 1 --planner-iters 1 --num-episodes 2 --root-seed 0 --push --out-repo abdelstark/lensemble-phase2-so100-checkpoint --out-repo-type model --out-path reports/phase2_downstream_eval_report.json",
  "published_report_uri": "hf://models/abdelstark/lensemble-phase2-so100-checkpoint/reports/phase2_downstream_eval_report.json",
  "source_report_uri": "hf://models/abdelstark/lensemble-phase2-so100-checkpoint/claim_mvp_report.json",
  "claim_boundary": "Compact engineering-scale downstream eval of a published Phase 2 checkpoint on the built-in synthetic://toy harness. It verifies checkpoint load, latent MPC execution, planner-cost reporting, and residency-safe EvalReport emission; it does not claim SO-100 task success or paper-scale LeWorldModel performance."
}