{
  "format": "chromatic-continuity-audit-summary/1",
  "passed": true,
  "audit_identity": "A separate agent in the same authoring team; independent code review and arithmetic, not an external replication.",
  "tests_run": 38,
  "test_families": {
    "MetricFixtures": 10,
    "SourceContract": 16,
    "BootstrapChecks": 4,
    "RetainedResultAudit": 8
  },
  "reports": {
    "audit-pretest.json": "3a7fb20162dff8ed0e2279c459456d6b9e48651ef440c420dca08e79a05ea1d3",
    "audit-historical-baseline.json": "a3b2a16e95aa62abd7475bac6a3bfa16abd04fb470017a33ea97be9c7d623570",
    "audit-release-checks.json": "79221c32ff55eb4d15f482fb83274155414beff8587ff722ff921b351d18c8e3",
    "audit-independent-scores.json": "07e4b48bd86f49b142fe4af6ac9ead4a44bf065b6e492d55561db9aba8ffe675",
    "audit-bootstrap.json": "691cc5fba8d1eab989d1aa7f9bfcc1d3a249fd44cc95a990106afa45dd340bfc"
  },
  "checks": [
    "Pretest review recorded before final scores; exact final model, generator and dependency files frozen.",
    "Train, development and reserved scene IDs disjoint, including comparison with the earlier recorded train/test IDs.",
    "Final input bytes hashed before inference; archive source and target edge/count arrays match retained input files.",
    "Source and predicted geometry are features; target counts, target geometry and teacher calls are absent from inference.",
    "Actual postgate predicted cell counts score final links; raw proposals, ungated ablation and deterministic repairs remain separate.",
    "Independent set arithmetic matches all reported primary, boundary, competition and four-step local link results.",
    "New endpoint-pair links and birth links are distinguished from adding a new type to an existing endpoint pair.",
    "No-growth false-positive scenes, persistent links and complete graph equality are checked independently.",
    "Learned availability veto cannot invent an absent backbone slot; selected-parent colour copying is explicit engineered decoding.",
    "Three descriptive paired-scene bootstrap intervals use the preselected deployment seed without tuning on test results."
  ],
  "limits": [
    "This is distillation of known authored numeric rules, with explicit near-boundary and priority drills; it is not discovery of meanings.",
    "Supervision and architecture both changed relative to Job136; the same-input comparison is not an equal-budget causal study.",
    "New relationship support remains one appended-cell return; arbitrary new old-cell-to-old-cell relations are outside this model.",
    "The retained neural backbone still controls missing-cell availability and other existing field channels.",
    "Canonical trajectory divergence can compare reordered slots, so it is not a semantic-memory score.",
    "Bootstrap intervals describe resampling these procedural cases, not generalization to people, subjective experience, or unrestricted artistic tasks.",
    "No claim of colour qualia, philosophical comprehension, general intelligence or representation-specific superiority follows from these checks.",
    "This audit does not establish rendered-browser, physical-phone or live-site behavior."
  ],
  "post_test_lineage_audit": {
    "passed": true,
    "tests_run": 7,
    "report": "audit-lineage.json",
    "sha256": "5ec102d36c847a9e7214d676bd8069374955b8413be2d89d7338c06b272e3b4b"
  }
}