{
  "schema": "lewm-federated-probe/1",
  "role": "offline-math-cross-check",
  "seed": 20260612,
  "protocol": "offline math cross-check: disjoint per-participant expert episodes -> shipping JS federation reimplemented inline (shared init + offset, clipped deltas, deterministic mean) -> held-out validation pairs; before = identity adapter (frozen predictor), after = final global revision. The system-composed headline is docs/evidence/lewm_tworooms_system_probe.json (scripts/lewm_system_probe.py).",
  "checkpoint": {
    "repoId": "quentinll/lewm-tworooms",
    "revision": "77adaae0bc31deab21c93740d1f8bb947cd0bdec",
    "weightsSha256": "566f223624ea4bfb39dbfe6ae731198dd6ea73b7b8919fed6b1ecafca810f7dd"
  },
  "trainPairsPerParticipant": [
    40,
    40
  ],
  "result": {
    "rounds": 3,
    "participants": 2,
    "validationPairs": 20,
    "baselineMse": 0.06037891671884794,
    "adaptedMse": 0.05296706380249468,
    "relativeImprovement": 0.12275564583025336,
    "verdict": "improved",
    "displayVerdict": "improved",
    "collapseRisk": false,
    "collapseFlags": [],
    "diagnostics": {
      "baseline": {
        "latentStdMean": 0.9040295390801578,
        "effectiveRank": 9.860522985438644,
        "effectiveRankRatio": 0.0513568905491596,
        "sigregStatistic": 0.016371862411410806
      },
      "adapted": {
        "latentStdMean": 0.9065777602546818,
        "effectiveRank": 9.803875214332578,
        "effectiveRankRatio": 0.05106185007464884,
        "sigregStatistic": 0.01591906086748467
      },
      "collapseRisk": false,
      "collapseFlags": []
    },
    "roundTelemetry": [
      {
        "round": 1,
        "losses": [
          {
            "first": 0.10105719101924959,
            "last": 0.05379377836457425
          },
          {
            "first": 0.06371931703222192,
            "last": 0.03211384157185338
          }
        ]
      },
      {
        "round": 2,
        "losses": [
          {
            "first": 0.05429387585165301,
            "last": 0.03896894401556648
          },
          {
            "first": 0.05440326614240865,
            "last": 0.02523766954495268
          }
        ]
      },
      {
        "round": 3,
        "losses": [
          {
            "first": 0.04709869341838316,
            "last": 0.023008084976269137
          },
          {
            "first": 0.03596551025278097,
            "last": 0.015681571761421487
          }
        ]
      }
    ]
  },
  "passes": true,
  "nonClaims": [
    "Offline math cross-check for the Tapestry-like demo's federated adapter path; not the system-composed artifact, not paper-scale TwoRooms benchmark parity, and not evidence of production browser training."
  ]
}