{
  "label": "4h",
  "experiment": "EXP-14",
  "runId": "af9ed861-8418-40df-96ec-ddc4f1ff99cb",
  "checkpoint": "4d00a4615f0a8914e951122ba583f07046660cbaada738e5075f221c143fbfcd",
  "players": {},
  "summary": {
    "scheduled_games": 25,
    "audited_games": 25,
    "unattempted_or_unaudited": 0,
    "wins": 2,
    "draws": 1,
    "losses": 22,
    "unfinished": 0,
    "score_on_scored_games": 0.1,
    "fully_scored_pairs": 12,
    "pair_mean_score": 0.10416666666666667,
    "pair_standard_error": 0.05722706412779594,
    "uncertainty_limit": "paired-opening standard error; small selected suite, not an Elo/generalization estimate",
    "model_queries": 666,
    "helper_overrides": 4,
    "autoplay_excluded_from_model_metrics": true
  },
  "gameCount": 25,
  "suiteSha256": "24cb7109c784c564c5f76e7eaada12df27305aa8e6f0768d62329fab64faf6df",
  "sourceManifestSha256": "2f5d202f1e4a97c011b715d514da7921141d94fa9035911b6f380296125afe62",
  "sourceResultsSha256": "627a875c170164a871e947a695748ab367d758f2a5a917ea3cda5c54a144f1e2",
  "protocol": {
    "autoplay": "all prefix plies retained from standard start in authoritative board/repetition history and PGN; model sees current structured state and at most last16 genuine plies; book autoplay excluded from model legality, query and latency metrics",
    "checkpoints": [
      "roughly-2h",
      "roughly-4h",
      "roughly-6h",
      "final-selected"
    ],
    "comparison": "Identical fixed 25-game system-level checkpoint trajectory; diagnostic only, no selection or promotion. Repeated games overlap the frozen 100-game suite.",
    "context": "state-v2 structured board/side/castling/en-passant/halfmove state plus optional genuine last16 plies; 120 prompt slots, 122 maximum model input slots; fixed position IDs and masked left history padding; 1024 allocated position rows",
    "decoding": "full legal-move joint argmax, UCI lexical tie break; unmasked raw512 conditional probabilities and raw greedy diagnostics logged separately",
    "draw_policy": "symmetric python-chess outcome(claim_draw=True)",
    "engine_timeout_seconds": 10,
    "game_walltime_seconds": 180,
    "games_per_checkpoint": 25,
    "helper": "Rust adaptive iterative 1/3/5/7-ply forced-mate override with verified all-defenses proof",
    "helper_nodes": 50000,
    "helper_seconds": 0.25,
    "incomplete_result": "* with reason, never counted as earned win/loss/draw",
    "inference_precision": "GPU BF16 autocast with FP32 model weights and FP32 output-logit conversion; CPU preflight uses FP32",
    "launch_gate": "independent adapter/runner/budget review, parent release and fresh GPU availability; bounded legal-decoding preflight before matches",
    "model_timeout_seconds": 60,
    "node_budget_is_soft": true,
    "ply_cap_including_autoplay": 256,
    "proposed_match_job_walltime_seconds_per_checkpoint": 3600,
    "sunfish_commit": "e396948acbd80cdcbad3dccfb07e878fba9df8f4",
    "sunfish_requested_nodes": 1000,
    "helper_max_plies": 7,
    "helper_workers": 4,
    "system_comparison": true,
    "baseline_checkpoint_role": "CM-8 milestone checkpoint, random-init lineage; no historical checkpoint loaded",
    "model_prompt_slots": 120,
    "model_max_input_slots": 122,
    "model_position_allocation": 1024,
    "explicit_history_max_plies": 16,
    "raw_probability_classes": 512,
    "evaluation_physical_gpu": 1,
    "torch_cpu_threads": 2,
    "model_turn_idle_seconds": 0.25,
    "latency_exclusions": "Deliberate pre-turn throttle and resource-check time excluded from per-turn inference timing, included in game/job walltime"
  },
  "gamesPgnSha256": "883a3b0df166c8e6203701a9123373a5cd3fb288414c89d1781edd885c91904f"
}