[
  {
    "id": 1,
    "title": "Before learning",
    "question": "What does an untrained reference establish?",
    "limit": "Only two games. No strength estimate or Elo claim.",
    "panels": [
      {
        "title": "Two recorded games",
        "columns": [
          "Cohort",
          "Wins",
          "Draws",
          "Losses",
          "Unfinished",
          "Score / bounds"
        ],
        "rows": [
          [
            "Both colors",
            0,
            0,
            2,
            0,
            "0%"
          ],
          [
            "White",
            0,
            0,
            1,
            0,
            "0%"
          ],
          [
            "Black",
            0,
            0,
            1,
            0,
            "0%"
          ]
        ],
        "note": "Random initialization; no mate helper. Use the preserved replay links below to inspect each game."
      }
    ],
    "sources": [
      {
        "url": "/downloads/studies/sources/experiments.json",
        "label": "experiments.json",
        "sha256": "d58640ffbb2ddd505a34a4e154bf30fe7db326343099ceea00fd37489d00c739"
      }
    ],
    "source_note": "Original narrative and source links are preserved in the full record below. A linked release inventory may document identity without supplying every study value."
  },
  {
    "id": 2,
    "title": "Learning, then deterioration",
    "question": "Which recorded checkpoint would you keep?",
    "limit": "Tiny validation set: 17 games and 3,659 targets. No playing-strength claim.",
    "panels": [
      {
        "title": "Recorded checkpoints",
        "columns": [
          "Update",
          "Validation NLL"
        ],
        "rows": [
          [
            57,
            2.0629628263037203
          ],
          [
            114,
            1.9667897356048443
          ],
          [
            171,
            1.8375828636901306
          ],
          [
            228,
            1.707227583366784
          ],
          [
            285,
            1.6084580547857168
          ],
          [
            342,
            1.5552885102978102
          ],
          [
            399,
            1.52174292786195
          ],
          [
            456,
            1.5101759066898401
          ],
          [
            513,
            1.5198594286455487
          ],
          [
            570,
            1.5681352180091042
          ],
          [
            627,
            1.6614797403982475
          ],
          [
            684,
            1.7390809957040858
          ],
          [
            741,
            1.9005723292790466
          ]
        ],
        "note": "Recorded observations only. Selection and terminal checkpoints differ where stated. Loss is not playing strength; objectives and validation sets differ between experiments."
      },
      {
        "title": "Selection",
        "columns": [
          "Checkpoint",
          "Update",
          "NLL"
        ],
        "rows": [
          [
            "Initial",
            0,
            5.027132
          ],
          [
            "Selected",
            456,
            1.510176
          ],
          [
            "Terminal",
            741,
            1.900572
          ]
        ],
        "note": "The later terminal validation deteriorated. The selected checkpoint is not the terminal checkpoint."
      }
    ],
    "sources": [
      {
        "url": "/downloads/studies/sources/experiments.json",
        "label": "experiments.json",
        "sha256": "d58640ffbb2ddd505a34a4e154bf30fe7db326343099ceea00fd37489d00c739"
      }
    ],
    "source_note": "Original narrative and source links are preserved in the full record below. A linked release inventory may document identity without supplying every study value."
  },
  {
    "id": 3,
    "title": "A full month of games",
    "question": "What does teacher-forced validation measure?",
    "limit": "Legality diagnostics cover a 128-game subset, not autonomous rollout.",
    "panels": [
      {
        "title": "Recorded checkpoints",
        "columns": [
          "Update",
          "Validation NLL"
        ],
        "rows": [
          [
            5300,
            0.8899340847576866
          ],
          [
            10600,
            0.7769294259080382
          ],
          [
            15900,
            0.7174767785439452
          ],
          [
            21200,
            0.6790902393498442
          ],
          [
            21387,
            0.6762016430910819
          ],
          [
            26500,
            0.6517212626194913
          ],
          [
            31800,
            0.6303233843590866
          ],
          [
            37100,
            0.6128119806352297
          ],
          [
            41854,
            0.6058686751172994
          ]
        ],
        "note": "Recorded observations only. Selection and terminal checkpoints differ where stated. Loss is not playing strength; objectives and validation sets differ between experiments."
      },
      {
        "title": "Diagnostic boundaries",
        "columns": [
          "Measure",
          "Value"
        ],
        "rows": [
          [
            "Teacher-forced legality",
            0.6463698
          ],
          [
            "Illegal probability",
            0.009694815
          ]
        ],
        "note": "These statistics condition on recorded prefixes. Do not compare this NLL with the differently defined pilot objective."
      }
    ],
    "sources": [
      {
        "url": "/downloads/studies/sources/experiments.json",
        "label": "experiments.json",
        "sha256": "d58640ffbb2ddd505a34a4e154bf30fe7db326343099ceea00fd37489d00c739"
      }
    ],
    "source_note": "Original narrative and source links are preserved in the full record below. A linked release inventory may document identity without supplying every study value."
  },
  {
    "id": 4,
    "title": "Moves, filters and a helper",
    "question": "Where did the played moves come from?",
    "limit": "Every win included helper overrides; whether those overrides were necessary was not tested.",
    "panels": [
      {
        "title": "Color cohorts",
        "columns": [
          "Cohort",
          "Wins",
          "Draws",
          "Losses",
          "Unfinished",
          "Score / bounds"
        ],
        "rows": [
          [
            "All",
            3,
            0,
            17,
            0,
            "15%"
          ],
          [
            "White",
            2,
            0,
            8,
            0,
            "20%"
          ],
          [
            "Black",
            1,
            0,
            9,
            0,
            "10%"
          ]
        ],
        "note": "Color subsets partition the 20-game suite."
      },
      {
        "title": "Decision provenance",
        "columns": [
          "Stage",
          "Decisions"
        ],
        "rows": [
          [
            "All decisions",
            656
          ],
          [
            "Neural decisions",
            651
          ],
          [
            "Helper overrides",
            5
          ],
          [
            "Raw legal neural decisions",
            591
          ]
        ],
        "note": "591 of 651 raw neural decisions were legal. Legal filtering and helper moves are distinct from raw policy output."
      }
    ],
    "sources": [
      {
        "url": "/downloads/studies/sources/experiments.json",
        "label": "experiments.json",
        "sha256": "d58640ffbb2ddd505a34a4e154bf30fe7db326343099ceea00fd37489d00c739"
      },
      {
        "url": "/downloads/studies/sources/games.json",
        "label": "games.json",
        "sha256": "9e8c89f75a0563ffd255329527a402bd73a23c72be07aa2692cac46ab7f50aa4"
      }
    ],
    "source_note": "Original narrative and source links are preserved in the full record below. A linked release inventory may document identity without supplying every study value."
  },
  {
    "id": 5,
    "title": "A broader foundation",
    "question": "How much of the source pool was consumed?",
    "limit": "Partial first pass. Terminal update 75,558 was not selected; no terminal NLL is asserted.",
    "panels": [
      {
        "title": "Recorded checkpoints",
        "columns": [
          "Update",
          "Validation NLL"
        ],
        "rows": [
          [
            8022,
            0.8020281492112645
          ],
          [
            16044,
            0.7062413286860573
          ],
          [
            24066,
            0.6585811909788196
          ],
          [
            32088,
            0.6230701939552107
          ],
          [
            40110,
            0.5984513133298186
          ],
          [
            48132,
            0.5853722610531451
          ],
          [
            56154,
            0.5806068211344487
          ],
          [
            64176,
            0.5771529146043254
          ],
          [
            72198,
            0.5739030495971976
          ]
        ],
        "note": "Recorded observations only. Selection and terminal checkpoints differ where stated. Loss is not playing strength; objectives and validation sets differ between experiments."
      },
      {
        "title": "Inventory and exposure",
        "columns": [
          "Quantity",
          "Games",
          "Targets"
        ],
        "rows": [
          [
            "Source inventory",
            4137488,
            943591445
          ],
          [
            "Terminal consumed",
            2417856,
            551410959
          ],
          [
            "Selected consumed",
            "Not supplied",
            526834590
          ]
        ],
        "note": "Inventory is available data; consumed exposure is what training visited. Selected update 72,198 had NLL 0.5739030."
      }
    ],
    "sources": [
      {
        "url": "/downloads/studies/sources/experiments.json",
        "label": "experiments.json",
        "sha256": "d58640ffbb2ddd505a34a4e154bf30fe7db326343099ceea00fd37489d00c739"
      },
      {
        "url": "/downloads/studies/sources/foundation-20261029/training.json",
        "label": "foundation-20261029/training.json",
        "sha256": "6add612f5eea4cf5df5de393b4c9c77c472e1656b019283c4d61f3eaf5295b50"
      }
    ],
    "source_note": "Original narrative and source links are preserved in the full record below. A linked release inventory may document identity without supplying every study value."
  },
  {
    "id": 6,
    "title": "A matched but incomplete comparison",
    "question": "What changes when one game is unfinished?",
    "limit": "Bounds are possible outcomes of unfinished games, not confidence intervals. No established improvement.",
    "panels": [
      {
        "title": "Matched 50 opening pairs",
        "columns": [
          "Cohort",
          "Wins",
          "Draws",
          "Losses",
          "Unfinished",
          "Score / bounds"
        ],
        "rows": [
          [
            "CM-3",
            13,
            2,
            85,
            0,
            "14%"
          ],
          [
            "CM-4",
            21,
            9,
            69,
            1,
            "25.5\u201326.5%"
          ]
        ],
        "note": "An unfinished game is not a draw."
      },
      {
        "title": "Matched validation",
        "columns": [
          "Model",
          "NLL"
        ],
        "rows": [
          [
            "CM-3",
            0.6058687
          ],
          [
            "CM-4",
            0.5752402
          ]
        ],
        "note": "Visited-position raw legality is not a matched-position effect."
      }
    ],
    "sources": [
      {
        "url": "/downloads/studies/sources/foundation-20261029/comparison.json",
        "label": "foundation-20261029/comparison.json",
        "sha256": "9f2141246c14440f6396abda90d0e2b40e0ed518815c1ff4d8214cc216209b95"
      }
    ],
    "source_note": "Original narrative and source links are preserved in the full record below. A linked release inventory may document identity without supplying every study value."
  },
  {
    "id": 7,
    "title": "Continuing on Elite games",
    "question": "Do lower loss and better observed scores settle the comparison?",
    "limit": "Approximate paired intervals include zero; no equal-compute control.",
    "panels": [
      {
        "title": "Recorded checkpoints",
        "columns": [
          "Update",
          "Validation NLL"
        ],
        "rows": [
          [
            0,
            0.6061916479
          ],
          [
            103873,
            0.5300231569
          ]
        ],
        "note": "Recorded observations only. Selection and terminal checkpoints differ where stated. Loss is not playing strength; objectives and validation sets differ between experiments."
      },
      {
        "title": "Matched evaluation",
        "columns": [
          "Cohort",
          "Wins",
          "Draws",
          "Losses",
          "Unfinished",
          "Score / bounds"
        ],
        "rows": [
          [
            "CM-5",
            27,
            9,
            63,
            1,
            "31.5\u201332.5%"
          ],
          [
            "CM-4",
            21,
            9,
            69,
            1,
            "25.5\u201326.5%"
          ]
        ],
        "note": "Observed difference +5 to +7 points. Bounds reflect unfinished outcomes. Terminal update 109,393 was not freshly validated."
      }
    ],
    "sources": [
      {
        "url": "/downloads/studies/sources/experiments.json",
        "label": "experiments.json",
        "sha256": "d58640ffbb2ddd505a34a4e154bf30fe7db326343099ceea00fd37489d00c739"
      }
    ],
    "source_note": "Original narrative and source links are preserved in the full record below. A linked release inventory may document identity without supplying every study value."
  },
  {
    "id": 8,
    "title": "Changing the input",
    "question": "What information is present in a state-conditioned example?",
    "limit": "Research-only historical representation. Its objective differs from full-history training.",
    "panels": [
      {
        "title": "Recorded checkpoints",
        "columns": [
          "Update",
          "Validation NLL"
        ],
        "rows": [
          [
            0,
            10.821807
          ],
          [
            6500,
            0.6436094719295701
          ]
        ],
        "note": "Recorded observations only. Selection and terminal checkpoints differ where stated. Loss is not playing strength; objectives and validation sets differ between experiments."
      },
      {
        "title": "Representation and exposure",
        "columns": [
          "Property",
          "Recorded value"
        ],
        "rows": [
          [
            "Input",
            "Current board plus optional 16-ply history"
          ],
          [
            "Vocabulary",
            512
          ],
          [
            "Input slots",
            122
          ],
          [
            "Move targets",
            3
          ],
          [
            "Prepared pool",
            10000000
          ],
          [
            "Consumed decisions",
            3426304
          ]
        ],
        "note": "Pool and consumption were each half Main / half Elite. Selected update 6,500; terminal 6,692 had no fresh validation."
      }
    ],
    "sources": [
      {
        "url": "/downloads/studies/sources/experiments.json",
        "label": "experiments.json",
        "sha256": "d58640ffbb2ddd505a34a4e154bf30fe7db326343099ceea00fd37489d00c739"
      }
    ],
    "source_note": "Original narrative and source links are preserved in the full record below. A linked release inventory may document identity without supplying every study value."
  },
  {
    "id": 9,
    "title": "The system around the model",
    "question": "What does a bounded helper know when it stops?",
    "limit": "Different protocol from historical CM-5: state/BF16/adaptive seven-ply versus full-history/FP32/frozen three-ply.",
    "panels": [
      {
        "title": "CM-6 evaluation",
        "columns": [
          "Cohort",
          "Wins",
          "Draws",
          "Losses",
          "Unfinished",
          "Score / bounds"
        ],
        "rows": [
          [
            "CM-6",
            13,
            7,
            80,
            0,
            "16.5%"
          ]
        ],
        "note": "Historical CM-5 31.5\u201332.5% is not a matched comparison."
      },
      {
        "title": "Helper outcomes",
        "columns": [
          "Outcome",
          "Queries"
        ],
        "rows": [
          [
            "Proven",
            21
          ],
          [
            "Depth bound",
            1313
          ],
          [
            "Unknown: deadline",
            721
          ],
          [
            "Unknown: node bound",
            1140
          ]
        ],
        "note": "3,195 turns, 3,174 queries. Unknown is not evidence of no mate."
      }
    ],
    "sources": [
      {
        "url": "/downloads/studies/sources/catalog.json",
        "label": "catalog.json",
        "sha256": "ab9580dfbafbbc4da35cfbf9514f5fd9c56ea58b2000f0d8a413f6ae60378cde"
      }
    ],
    "source_note": "Original narrative and source links are preserved in the full record below. A linked release inventory may document identity without supplying every study value."
  },
  {
    "id": 10,
    "title": "Ordinary games and exact states",
    "question": "How do unique examples differ from repeat exposure?",
    "limit": "No book source in this stage. Exact accuracy change has an interval spanning zero.",
    "panels": [
      {
        "title": "Sources",
        "columns": [
          "Source",
          "Unique states/examples",
          "Consumed exposures"
        ],
        "rows": [
          [
            "Main",
            4878415,
            4878415
          ],
          [
            "Elite",
            4878415,
            4878415
          ],
          [
            "Exact",
            19659,
            58977
          ]
        ],
        "note": "Totals: 9,776,489 unique examples; 9,815,807 exposures; 29,447,421 targets."
      },
      {
        "title": "Recorded validation",
        "columns": [
          "Phase",
          "NLL"
        ],
        "rows": [
          [
            "Before",
            0.699053
          ],
          [
            "Selected",
            0.61841
          ],
          [
            "Terminal",
            0.620069
          ]
        ],
        "note": "Exact accuracy 86.8% to 87.0%. No continuous curve is inferred."
      }
    ],
    "sources": [
      {
        "url": "/downloads/studies/sources/cm7/training.json",
        "label": "cm7/training.json",
        "sha256": "ea8b8cc3c9d958274919fa6ad4ffc95c889aa9d23a3c7bbea43c8f0be6f01709"
      }
    ],
    "source_note": "Original narrative and source links are preserved in the full record below. A linked release inventory may document identity without supplying every study value."
  },
  {
    "id": 11,
    "title": "A theory-heavy follow-up",
    "question": "How is the scheduled mix different from source inventory?",
    "limit": "Do not sum unique counts across stages. Final-segment duration does not establish full-run duration.",
    "panels": [
      {
        "title": "Inventory to schedule",
        "columns": [
          "Source",
          "States / eligible rows",
          "Pairs",
          "Scheduled decisions"
        ],
        "rows": [
          [
            "Book",
            184657,
            228502,
            685506
          ],
          [
            "Exact",
            19659,
            160314,
            685506
          ],
          [
            "Elite",
            4864530,
            "Not supplied",
            342753
          ]
        ],
        "note": "Scheduled mixture: 40% book, 40% exact, 20% Elite."
      },
      {
        "title": "Validation snapshots",
        "columns": [
          "Phase",
          "NLL",
          "Exact accuracy"
        ],
        "rows": [
          [
            "Before",
            0.61841,
            "87%"
          ],
          [
            "Selected u2232",
            0.59579,
            "89.85%"
          ],
          [
            "Terminal u3348",
            0.596168,
            "90.2%"
          ]
        ],
        "note": "Selection uses the recorded validation criterion, not terminal exact accuracy alone."
      }
    ],
    "sources": [
      {
        "url": "/downloads/studies/sources/cm7/training.json",
        "label": "cm7/training.json",
        "sha256": "ea8b8cc3c9d958274919fa6ad4ffc95c889aa9d23a3c7bbea43c8f0be6f01709"
      }
    ],
    "source_note": "Original narrative and source links are preserved in the full record below. A linked release inventory may document identity without supplying every study value."
  },
  {
    "id": 12,
    "title": "Several suites, several denominators",
    "question": "Which games belong to each result?",
    "limit": "Suites overlap. Do not pool them. Outcome bounds are not confidence intervals.",
    "panels": [
      {
        "title": "Suite inspector",
        "columns": [
          "Cohort",
          "Wins",
          "Draws",
          "Losses",
          "Unfinished",
          "Score / bounds"
        ],
        "rows": [
          [
            "Pre-pass 10",
            4,
            3,
            3,
            0,
            "55%"
          ],
          [
            "Update 1500: 10",
            4,
            1,
            5,
            0,
            "45%"
          ],
          [
            "Selected: 100",
            21,
            19,
            58,
            2,
            "30.5\u201332.5%"
          ],
          [
            "Selected matching 10",
            3,
            4,
            2,
            1,
            "50\u201360%"
          ],
          [
            "Historical CM-6: 100",
            13,
            7,
            80,
            0,
            "16.5%"
          ]
        ],
        "note": "The selected ten-game subset is already inside the selected 100-game suite."
      }
    ],
    "sources": [
      {
        "url": "/downloads/studies/sources/cm7/evaluation.json",
        "label": "cm7/evaluation.json",
        "sha256": "572ed05b3b30241338cfe72b2f05a777707ccea050d1ffa403c3c1371f5412de"
      }
    ],
    "source_note": "Original narrative and source links are preserved in the full record below. A linked release inventory may document identity without supplying every study value."
  },
  {
    "id": 13,
    "title": "A stronger observed CM-9",
    "question": "What do two different opponents tell us?",
    "limit": "Historical promotion record, not the current live model. No general Elo estimate.",
    "panels": [
      {
        "title": "Recorded suites",
        "columns": [
          "Cohort",
          "Wins",
          "Draws",
          "Losses",
          "Unfinished",
          "Score / bounds"
        ],
        "rows": [
          [
            "Sunfish",
            23,
            18,
            59,
            0,
            "32%"
          ],
          [
            "CM-5 head-to-head",
            20,
            23,
            7,
            0,
            "63%"
          ]
        ],
        "note": "Matched CM-5 Sunfish score 16% (counts not supplied); gain interval +5.5 to +26.5 points. Head-to-head score interval 55\u201371%."
      },
      {
        "title": "Training sources",
        "columns": [
          "Source",
          "Games"
        ],
        "rows": [
          [
            "Engine",
            172656
          ],
          [
            "Elite",
            7446998
          ]
        ],
        "note": "7,619,654 games total. Final horizon mixture 70/20/10. No synthetic game replay is provided."
      }
    ],
    "sources": [
      {
        "url": "/downloads/studies/sources/cm9/release.json",
        "label": "cm9/release.json",
        "sha256": "c6e3fa0148f051e01d51cabc1c47660d9f38745f4550d91f25a2480c4eedeb8f"
      }
    ],
    "source_note": "Original narrative and source links are preserved in the full record below. A linked release inventory may document identity without supplying every study value."
  },
  {
    "id": 14,
    "title": "CM-8: the failed experiment matters",
    "question": "Can a much lower loss hide weak playing results?",
    "limit": "Initialization, inputs, objective and data changed together. These results do not isolate a cause.",
    "panels": [
      {
        "title": "Loss and exposure",
        "columns": [
          "Observation",
          "Value"
        ],
        "rows": [
          [
            "Initial NLL",
            6.338264
          ],
          [
            "Selected NLL u46875",
            0.677189
          ],
          [
            "Terminal update",
            55595
          ],
          [
            "Consumed decisions",
            28464640
          ]
        ],
        "note": "Do not interpret these endpoints as a full learning curve or compare this objective directly with full-history NLL."
      },
      {
        "title": "Playing results",
        "columns": [
          "Cohort",
          "Wins",
          "Draws",
          "Losses",
          "Unfinished",
          "Score / bounds"
        ],
        "rows": [
          [
            "Sunfish",
            6,
            2,
            92,
            0,
            "7%"
          ],
          [
            "CM-7 head-to-head",
            10,
            21,
            69,
            0,
            "20.5%"
          ]
        ],
        "note": "100 games per suite. CM-8 was not promoted."
      },
      {
        "title": "Scheduled sources",
        "columns": [
          "Source",
          "Percent"
        ],
        "rows": [
          [
            "Elite",
            81.640625
          ],
          [
            "Book",
            12.5
          ],
          [
            "Exact",
            3.125
          ],
          [
            "Puzzle",
            2.734375
          ]
        ],
        "note": "Random-initialized mixed curriculum with state plus recent history and a selected next-move objective."
      }
    ],
    "sources": [
      {
        "url": "/downloads/studies/sources/cm9/release.json",
        "label": "cm9/release.json",
        "sha256": "c6e3fa0148f051e01d51cabc1c47660d9f38745f4550d91f25a2480c4eedeb8f"
      }
    ],
    "source_note": "Original narrative and source links are preserved in the full record below. A linked release inventory may document identity without supplying every study value."
  },
  {
    "id": 15,
    "title": "CM-10: mixed evidence",
    "question": "Why can a benchmark and a head-to-head disagree?",
    "limit": "Experimental promotion by user choice, not proven superiority. This is historical; CM-14 is current.",
    "panels": [
      {
        "title": "Two comparisons",
        "columns": [
          "Cohort",
          "Wins",
          "Draws",
          "Losses",
          "Unfinished",
          "Score / bounds"
        ],
        "rows": [
          [
            "Sunfish",
            25,
            25,
            50,
            0,
            "37.5%"
          ],
          [
            "CM-9 head-to-head",
            10,
            22,
            18,
            0,
            "42%"
          ]
        ],
        "note": "Matched CM-9 Sunfish score 27.5% (counts not supplied), gain interval 0 to +20 points. Head-to-head interval 31\u201353%."
      },
      {
        "title": "Continuation",
        "columns": [
          "Setting",
          "Value"
        ],
        "rows": [
          [
            "Starting update",
            324384
          ],
          [
            "Final update",
            450173
          ],
          [
            "Learning rate",
            5e-06
          ],
          [
            "Mask horizons",
            "7 / 9 / 11"
          ],
          [
            "Masked example cap",
            "30%"
          ]
        ],
        "note": "Four-hour continuation. The cap is a fraction of examples, not a fraction of tokens."
      }
    ],
    "sources": [
      {
        "url": "/downloads/studies/sources/cm10/release.json",
        "label": "cm10/release.json",
        "sha256": "b2851bab85f3250e67161e0e31f0d22de83d83f7d584d505eedbe3e4550ddbed"
      }
    ],
    "source_note": "Original narrative and source links are preserved in the full record below. A linked release inventory may document identity without supplying every study value."
  },
  {
    "id": 16,
    "title": "CM-13: moves and outcomes",
    "question": "What does sharing a transformer imply about the objectives?",
    "limit": "Observed improvement was uncertain; the automatic promotion criterion was not met.",
    "panels": [
      {
        "title": "Targets",
        "columns": [
          "Distribution",
          "Training targets"
        ],
        "rows": [
          [
            "Move",
            2113162939
          ],
          [
            "Fixed-White outcome",
            316974608
          ]
        ],
        "note": "Shared transformer/output matrix, separate move and W/D/L distributions. Clean full-move history; no outcome labels in inputs."
      },
      {
        "title": "Matched eight-hour comparison",
        "columns": [
          "Cohort",
          "Wins",
          "Draws",
          "Losses",
          "Unfinished",
          "Score / bounds"
        ],
        "rows": [
          [
            "CM-13 u148481",
            26,
            2,
            22,
            0,
            "54%"
          ],
          [
            "CM-10",
            18,
            6,
            26,
            0,
            "42%"
          ]
        ],
        "note": "50 games per model; paired difference interval \u22126 to +31 points. Mate-only assistance. This is not EXP-17\u2019s freshly rerun reference."
      }
    ],
    "sources": [
      {
        "url": "/downloads/studies/sources/cm13/release.json",
        "label": "cm13/release.json",
        "sha256": "729bde0051eba941121badf94fb5bda7e709117d6bae6c260eecedb0519afc22"
      }
    ],
    "source_note": "Original narrative and source links are preserved in the full record below. A linked release inventory may document identity without supplying every study value."
  },
  {
    "id": 17,
    "title": "CM-13: the final snapshot",
    "question": "How did the later snapshot fare on matched openings?",
    "limit": "Sunfish evidence, not general Elo or proof that duration alone caused improvement.",
    "panels": [
      {
        "title": "Matched 50 opening pairs",
        "columns": [
          "Cohort",
          "Wins",
          "Draws",
          "Losses",
          "Unfinished",
          "Score / bounds"
        ],
        "rows": [
          [
            "Final u279393",
            63,
            9,
            28,
            0,
            "67.5%"
          ],
          [
            "Fresh eight-hour reference",
            47,
            12,
            41,
            0,
            "53%"
          ]
        ],
        "note": "Gain +14.5 points; paired opening-cluster bootstrap 95% interval +1.5 to +27 points, 20,000 resamples. All 200 games audited."
      },
      {
        "title": "Protocol boundaries",
        "columns": [
          "Interface",
          "Assistance"
        ],
        "rows": [
          [
            "Evaluated system",
            "Stockfish 19 mate-only"
          ],
          [
            "Browser",
            "Unassisted"
          ],
          [
            "Excluded from evaluation",
            "Value search and 97% takeover"
          ]
        ],
        "note": "Identical helper and timing for both evaluated snapshots."
      }
    ],
    "sources": [
      {
        "url": "/downloads/studies/sources/cm13-final/release.json",
        "label": "cm13-final/release.json",
        "sha256": "5591e07dd6d393a822b051b8ba323b283fdb3cc478f8c16cc2bca228358bf5a3"
      }
    ],
    "source_note": "Original narrative and source links are preserved in the full record below. A linked release inventory may document identity without supplying every study value."
  },
  {
    "id": 18,
    "title": "CM-14: current, with clear boundaries",
    "question": "Which system produced the published 72%?",
    "limit": "Single cohort; no matched comparator or confidence interval supplied for this older benchmark.",
    "panels": [
      {
        "title": "Published benchmark",
        "columns": [
          "Cohort",
          "Wins",
          "Draws",
          "Losses",
          "Unfinished",
          "Score / bounds"
        ],
        "rows": [
          [
            "CM-14 u210395",
            68,
            8,
            24,
            0,
            "72%"
          ]
        ],
        "note": "100 games; mate-only assistance. Do not substitute this result for the newer 72.5% matched reference in CM-15 evaluation."
      },
      {
        "title": "Model and interfaces",
        "columns": [
          "Interface",
          "Recorded configuration"
        ],
        "rows": [
          [
            "Initialization",
            "Fresh Muon/AdamW lineage; not CM-13 continuation"
          ],
          [
            "Benchmark",
            "Mate-only"
          ],
          [
            "Live bot",
            "Separate 97% endgame conversion enabled"
          ],
          [
            "Browser",
            "Unassisted"
          ]
        ],
        "note": "Current release identity is preserved; this presentation does not modify weights or serving configuration."
      }
    ],
    "sources": [
      {
        "url": "/downloads/studies/sources/cm14/release.json",
        "label": "cm14/release.json",
        "sha256": "d3dfc2815162b842dc6ec9e39988825dc8665526c5825316025377162758fc0a"
      }
    ],
    "source_note": "Original narrative and source links are preserved in the full record below. A linked release inventory may document identity without supplying every study value."
  },
  {
    "id": 19,
    "title": "CM-15: inspect the forecast",
    "question": "Does predicting the eventual result mean stronger play?",
    "limit": "Initial completed run only, not today\u2019s stopped continuation. Not promoted. CM-16 remains a proposal.",
    "panels": [
      {
        "title": "Playing comparisons",
        "columns": [
          "Cohort",
          "Wins",
          "Draws",
          "Losses",
          "Unfinished",
          "Score / bounds"
        ],
        "rows": [
          [
            "Sunfish",
            64,
            9,
            27,
            0,
            "68.5%"
          ],
          [
            "CM-14 head-to-head",
            7,
            7,
            11,
            0,
            "42%"
          ]
        ],
        "note": "New matched CM-14 Sunfish reference 72.5%; difference \u22124 points, paired interval \u221215 to +7. Both use frozen mate checks AND 97% endgame conversion. Head-to-head: 24 paired games score 41.7%, interval 29.2\u201354.2%; extra drawn Black game excluded from interval."
      },
      {
        "title": "Outcome prediction by retrospective third",
        "columns": [
          "Third",
          "Brier \u2193",
          "Log loss \u2193",
          "Actual-result probability"
        ],
        "rows": [
          [
            "Early",
            0.606,
            0.98,
            "41.1%"
          ],
          [
            "Middle",
            0.488,
            0.821,
            "53.0%"
          ],
          [
            "Late",
            0.305,
            0.543,
            "68.3%"
          ]
        ],
        "note": "7,945 nonterminal positions / 100 games, game-equal weighting and opening-pair bootstrap. Thirds use completed game length, not online-known labels. Overall Brier .468 vs .518 cross-fitted baseline; difference interval \u2212.120 to +.020. Nine draws limit calibration."
      }
    ],
    "sources": [],
    "source_note": "Original narrative and source links are preserved in the full record below. A linked release inventory may document identity without supplying every study value."
  }
]
