[
  {
    "label": "2h",
    "experiment": "EXP-14",
    "runId": "aacb01a0-f4b6-4c4a-8cd8-e5102fe6c44c",
    "checkpoint": "e7173335e1b606bfcfb0f445e3beff0f4e69faac8d6d9bbd1ce72935c5f9c1e1",
    "players": {},
    "summary": {
      "scheduled_games": 25,
      "audited_games": 25,
      "unattempted_or_unaudited": 0,
      "wins": 1,
      "draws": 0,
      "losses": 24,
      "unfinished": 0,
      "score_on_scored_games": 0.04,
      "fully_scored_pairs": 12,
      "pair_mean_score": 0.041666666666666664,
      "pair_standard_error": 0.041666666666666664,
      "uncertainty_limit": "paired-opening standard error; small selected suite, not an Elo/generalization estimate",
      "model_queries": 544,
      "helper_overrides": 1,
      "autoplay_excluded_from_model_metrics": true
    },
    "gameCount": 25,
    "suiteSha256": "24cb7109c784c564c5f76e7eaada12df27305aa8e6f0768d62329fab64faf6df",
    "sourceManifestSha256": "3ba71c676e2a2ec4f2bbb42bc5e379806b25eb7dc6a505c3a86e49b1854d973f",
    "sourceResultsSha256": "78d4e5e68b687a4d64ff54ddc8c6b659e04948d3d8c4462e1dd6316c6f429f57",
    "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": "22acfb6276e27e54bbadc2161a086bbf31028a48c551ee0e12717f4f236b1e3f"
  },
  {
    "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"
  },
  {
    "label": "6h",
    "experiment": "EXP-14",
    "runId": "c4a6338d-d195-48f1-adda-94cb1b707911",
    "checkpoint": "7f0d163903c626c379feec6eb70b657a8559ddce48fd9a4b7aa9063f7069dd81",
    "players": {},
    "summary": {
      "scheduled_games": 25,
      "audited_games": 25,
      "unattempted_or_unaudited": 0,
      "wins": 4,
      "draws": 0,
      "losses": 21,
      "unfinished": 0,
      "score_on_scored_games": 0.16,
      "fully_scored_pairs": 12,
      "pair_mean_score": 0.16666666666666666,
      "pair_standard_error": 0.07106690545187015,
      "uncertainty_limit": "paired-opening standard error; small selected suite, not an Elo/generalization estimate",
      "model_queries": 798,
      "helper_overrides": 6,
      "autoplay_excluded_from_model_metrics": true
    },
    "gameCount": 25,
    "suiteSha256": "24cb7109c784c564c5f76e7eaada12df27305aa8e6f0768d62329fab64faf6df",
    "sourceManifestSha256": "c0b1f53cbce89a638e4e433214a530bf3b0314914652850052d53ac893089e9e",
    "sourceResultsSha256": "fac26f935149f824f38cfdb9e94c8e46138b3d0229bd16441cd2af66ecb8a3a7",
    "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": "d0035328ee5bbda7b448259a864fd9daf580fa324cc40f1748a715128ea15382"
  },
  {
    "label": "final",
    "experiment": "EXP-14",
    "runId": "c70798a1-1404-4a58-9f0f-8fde3df83109",
    "checkpoint": "57c714120d47a06d30386f63c135a3afd141c61d012b46774cd2485b8c78fa2e",
    "players": {},
    "summary": {
      "scheduled_games": 100,
      "audited_games": 100,
      "unattempted_or_unaudited": 0,
      "wins": 6,
      "draws": 2,
      "losses": 92,
      "unfinished": 0,
      "score_on_scored_games": 0.07,
      "fully_scored_pairs": 50,
      "pair_mean_score": 0.07,
      "pair_standard_error": 0.023733211036908786,
      "uncertainty_limit": "paired-opening standard error; small selected suite, not an Elo/generalization estimate",
      "model_queries": 2846,
      "helper_overrides": 12,
      "autoplay_excluded_from_model_metrics": true
    },
    "gameCount": 100,
    "suiteSha256": "7e71cc17c8da9e55cdceafa951cb58ab7f1e0d75594b44d7622bae679abada98",
    "sourceManifestSha256": "7012f7dd5849bb1454cd148a47b29891f068c0f8730080d099b392b731b6857a",
    "sourceResultsSha256": "13a615493025afc329accebe6f6cc2ea80bb692ea8ce8bdec0dfcc512dac7810",
    "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": [
        "final-validation-selected"
      ],
      "comparison": "Final full100 with exact prior25 subset derived without duplicate games; no promotion",
      "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": 100,
      "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": 7200,
      "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": "c220bc7476dc3040204225aa6c584c1ea4c05180f1909fdd29df8ccb4cbac68d"
  },
  {
    "label": "cm7-cm8",
    "experiment": "EXP-15",
    "runId": "3b5c0c65-e4f7-4791-bb59-952c1d0549d8",
    "checkpoint": null,
    "players": {
      "CM-7": {
        "sha256": "29e68a8e85336cbb1f56efdad36bd76217728b6a376d1e5f74e89887d21383da",
        "representation": "state-history16-v2",
        "helper": {
          "version": "adaptive-rust-1-3-5-7",
          "backend": "Rust",
          "maxPlies": 7,
          "nodeBudget": 50000,
          "timeBudgetMs": 250,
          "workers": 4
        }
      },
      "CM-8": {
        "sha256": "57c714120d47a06d30386f63c135a3afd141c61d012b46774cd2485b8c78fa2e",
        "representation": "state-history16-v2",
        "helper": {
          "version": "adaptive-rust-1-3-5-7",
          "backend": "Rust",
          "maxPlies": 7,
          "nodeBudget": 50000,
          "timeBudgetMs": 250,
          "workers": 4
        }
      }
    },
    "summary": {
      "scheduled_games": 100,
      "audited_games": 100,
      "unattempted_or_unaudited": 0,
      "total": {
        "CM8_wins": 10,
        "draws": 21,
        "CM7_wins": 69,
        "unfinished": 0,
        "CM8_points": 20.5,
        "CM7_points": 79.5,
        "CM8_score_bounds_including_unfinished": [
          20.5,
          20.5
        ]
      },
      "CM8_white": {
        "CM8_wins": 5,
        "draws": 9,
        "CM7_wins": 36,
        "unfinished": 0,
        "CM8_points": 9.5,
        "CM7_points": 40.5,
        "CM8_score_bounds_including_unfinished": [
          9.5,
          9.5
        ]
      },
      "CM8_black": {
        "CM8_wins": 5,
        "draws": 12,
        "CM7_wins": 33,
        "unfinished": 0,
        "CM8_points": 11.0,
        "CM7_points": 39.0,
        "CM8_score_bounds_including_unfinished": [
          11.0,
          11.0
        ]
      },
      "interpretation": "direct system comparison on fixed paired openings; no universal Elo estimate"
    },
    "gameCount": 100,
    "suiteSha256": "3dbb9810c39b1dec85b9ce58eb5142dd9481907d5f05faf64b9717bf0c8e9820",
    "sourceManifestSha256": "e741f0dc35b4cbddf5ad6efbcf02973391b17734fecd176ac5838a8c28405850",
    "sourceResultsSha256": "ef920b4b80bdb71e91248c1a977f0c66dec34872d5716d39531d88c250c25931",
    "protocol": {
      "players": [
        "CM-7",
        "CM-8"
      ],
      "schedule_model_color_means": "CM-8 color; CM-7 opposite",
      "representation": "state-v2 FEN-derived state plus up to16 genuine plies; 120 prompt /122 maximum input; fixed position IDs and masked left history padding",
      "history_plies": 16,
      "vocab_size": 512,
      "decoder": "exact legal joint argmax; raw unmasked512-class metrics separately",
      "weights": "FP32",
      "autocast": "BF16",
      "attention": "SDPA",
      "helper": "Rust adaptive iterative1/3/5/7",
      "helper_nodes": 50000,
      "helper_seconds": 0.25,
      "helper_workers": 4,
      "ply_cap_including_autoplay": 256,
      "game_walltime_seconds": 180,
      "model_load_and_query_timeout_seconds": 60,
      "idle_pacing_seconds_per_player_turn": 0.25,
      "physical_gpu": 1,
      "both_models_resident_during_game": true,
      "models_reloaded_each_game": true,
      "draw_claims": "symmetric immediate claims including a claim by declaring a move",
      "no_sunfish_player": true,
      "unfinished": "explicit *, never relabeled draw/loss",
      "total_budget_seconds": 14400,
      "no_promotion": true
    },
    "gamesPgnSha256": "122039607783fe9fc374a58bcad153d93992083f3f95dae29b2562a82c8d29e1"
  },
  {
    "label": "cm7-cm5",
    "experiment": "EXP-16",
    "runId": "f59419e3-0921-4ab7-91ec-73f70b4a7c4d",
    "checkpoint": null,
    "players": {
      "CM-7": {
        "sha256": "29e68a8e85336cbb1f56efdad36bd76217728b6a376d1e5f74e89887d21383da",
        "representation": "state-history16-v2",
        "helper": {
          "version": "adaptive-rust-1-3-5-7",
          "backend": "Rust",
          "maxPlies": 7,
          "nodeBudget": 50000,
          "timeBudgetMs": 250,
          "workers": 4
        }
      },
      "CM-5": {
        "sha256": "c0a3f67ac4b364276028950b2010c5644e03efb3ca0f1228b32c36b8b5340238",
        "representation": "full-history-v1",
        "helper": {
          "version": "adaptive-rust-1-3-5-7",
          "backend": "Rust",
          "maxPlies": 7,
          "nodeBudget": 50000,
          "timeBudgetMs": 250,
          "workers": 4
        }
      }
    },
    "summary": {
      "scheduled_games": 100,
      "audited_games": 59,
      "unattempted_or_unaudited": 41,
      "total": {
        "CM5_wins": 24,
        "draws": 17,
        "CM7_wins": 17,
        "unfinished": 1,
        "CM5_points": 32.5,
        "CM7_points": 25.5,
        "CM5_score_bounds_including_unfinished": [
          32.5,
          33.5
        ]
      },
      "CM5_white": {
        "CM5_wins": 13,
        "draws": 10,
        "CM7_wins": 6,
        "unfinished": 1,
        "CM5_points": 18.0,
        "CM7_points": 11.0,
        "CM5_score_bounds_including_unfinished": [
          18.0,
          19.0
        ]
      },
      "CM5_black": {
        "CM5_wins": 11,
        "draws": 7,
        "CM7_wins": 11,
        "unfinished": 0,
        "CM5_points": 14.5,
        "CM7_points": 14.5,
        "CM5_score_bounds_including_unfinished": [
          14.5,
          14.5
        ]
      },
      "interpretation": "direct system comparison on fixed paired openings; no universal Elo estimate"
    },
    "gameCount": 59,
    "suiteSha256": "030b100a84edc609b5bef71edd3af8035725af8e55186f2ab6b4ed80c13a8777",
    "sourceManifestSha256": "c2c2bb2e0398b13f17b2b612657372ac0db1ebc7f2f57270bfcc5b05495f83a4",
    "sourceResultsSha256": "42cea4e6b39cee53ded5f05b617b673c559caacb3bf611b5dda8d12f09f48eef",
    "protocol": {
      "players": [
        "CM-7",
        "CM-5"
      ],
      "schedule_model_color_means": "CM-5 color; CM-7 opposite",
      "representations": {
        "CM-7": "state-v2 FEN-derived state plus up to16 genuine plies; 120 prompt/122 maximum input; fixed position IDs and masked left history padding",
        "CM-5": "legacy BOS + every move triple from standard start;1024 context,no crop"
      },
      "vocab_sizes": {
        "CM-7": 512,
        "CM-5": 128
      },
      "decoder": "same exact legal joint argmax; raw unmasked native-vocabulary metrics separately",
      "weights": "FP32",
      "autocast": "BF16",
      "attention": "SDPA",
      "helper": "Rust adaptive iterative1/3/5/7",
      "helper_nodes": 50000,
      "helper_seconds": 0.25,
      "helper_workers": 4,
      "ply_cap_including_autoplay": 256,
      "game_walltime_seconds": 180,
      "model_load_and_query_timeout_seconds": 60,
      "idle_pacing_seconds_per_player_turn": 0.25,
      "physical_gpu": 1,
      "both_models_resident_during_game": true,
      "models_reloaded_each_game": true,
      "draw_claims": "symmetric immediate claims including a claim by declaring a move",
      "no_sunfish_player": true,
      "unfinished": "explicit *, never relabeled draw/loss",
      "total_budget_seconds": 14400,
      "no_promotion": true
    },
    "gamesPgnSha256": "5f5279c1d2dcaa1f80f260a874893ae9db23a62d48e7074cbb3308f1723dbc34",
    "status": "stopped_by_user",
    "completed": 58,
    "interruptedUnscored": 1,
    "unplayed": 41
  }
]