{
  "schema": "lewm-federated-probe/1",
  "role": "system-composed-headline",
  "seed": 20260612,
  "protocol": "system-composed: server aggregation path. REAL node-trained adapter deltas -> FederatedDemoService.submit_update (real fail-closed validation) -> _close_round_lewm (real deterministic-mean aggregation + hash-chained revisions) -> held-out before/after probe on the SERVER-PRODUCED final modelRevisionId. No reimplemented coordinator mean.",
  "checkpoint": {
    "repoId": "quentinll/lewm-tworooms",
    "revision": "77adaae0bc31deab21c93740d1f8bb947cd0bdec",
    "weightsSha256": "566f223624ea4bfb39dbfe6ae731198dd6ea73b7b8919fed6b1ecafca810f7dd"
  },
  "runId": "run-kqenx7zp",
  "modelRevisionId": "lewmrev-c59381609f77",
  "modelRevisionChain": [
    "lewmrev-9241b7bc92f1",
    "lewmrev-397ccafff492",
    "lewmrev-c59381609f77"
  ],
  "aggregateDeltaNorm": 0.81243471,
  "adapterStateNorm": 2.43383821,
  "serverOffsetParameterCount": 12512,
  "trainPairsPerParticipant": [
    40,
    40
  ],
  "claimAuditViolations": 0,
  "result": {
    "probeVersion": "lewm-validation-probe/1",
    "seed": 20260612,
    "pairCount": 20,
    "baselineMse": 0.06037897796856822,
    "adaptedMse": 0.05296723026100999,
    "relativeImprovement": 0.12275377883038366,
    "verdict": "improved",
    "displayVerdict": "improved",
    "collapseRisk": false,
    "collapseFlags": [],
    "diagnostics": {
      "baseline": {
        "latentStdMean": 0.9040294455047887,
        "effectiveRank": 9.86052406648484,
        "effectiveRankRatio": 0.05135689617960854,
        "sigregStatistic": 0.01637186087198948
      },
      "adapted": {
        "latentStdMean": 0.9065790589646596,
        "effectiveRank": 9.803862727073762,
        "effectiveRankRatio": 0.05106178503684251,
        "sigregStatistic": 0.015919042720322508
      },
      "collapseRisk": false,
      "collapseFlags": []
    },
    "honest": true,
    "rounds": 3,
    "participants": 2,
    "modelRevisionId": "lewmrev-c59381609f77",
    "roundTelemetry": [
      {
        "round": 1,
        "losses": [
          {
            "first": 0.10105732,
            "last": 0.05379406
          },
          {
            "first": 0.06371928,
            "last": 0.03211352
          }
        ]
      },
      {
        "round": 2,
        "losses": [
          {
            "first": 0.05429405,
            "last": 0.03896919
          },
          {
            "first": 0.05440329,
            "last": 0.02523769
          }
        ]
      },
      {
        "round": 3,
        "losses": [
          {
            "first": 0.04709885,
            "last": 0.02300816
          },
          {
            "first": 0.03596558,
            "last": 0.01568131
          }
        ]
      }
    ]
  },
  "passes": true,
  "crossCheck": "docs/evidence/lewm_tworooms_probe_check.json (offline math cross-check; same protocol, reimplemented coordinator mean)",
  "nonClaims": [
    "System-composed before/after validation probe for the Tapestry-like demo's federated adapter path through the real server aggregation/validation code. Single local coordinator, mean-of-clipped-deltas (no robust aggregation / DP in this path); not paper-scale TwoRooms benchmark parity and not production browser training."
  ]
}