{
  "format": "carrier-inheritance-independent-audit-plan/1",
  "written_before_final_data_and_scores": true,
  "audit_identity": "A separate agent within the same authoring team; independent metric implementation and artifact checks, not external or blinded replication.",
  "historical_problem": "Public Job137 outcomes are development knowledge: missing or spurious appended cells and original-cell colour divergence over four canonicalized turns. The earlier canonical-slot comparison confounded permutations with identity; stable external labels are required from the outset here.",
  "acceptance_checks": [
    "Hash the actual teacher, generator, train/development data, all three fitted checkpoints, runtime, evaluator, independent metrics and predeclared deployment seed before generating final inputs.",
    "Retain disjoint fresh base IDs for train, development, primary, stress and trajectory groups. Compare them with every available earlier data manifest and retained data split.",
    "Retain all final source/target/predicted numerical field channels, actual emitted cells and edges, all training seeds, COPY and same-input frozen Job137 baseline.",
    "Inference must not call the teacher, access targets or mutate source data; poisoning target-looking extra fields and inactive padding must leave predictions unchanged.",
    "Document teacher-informed structural choices separately from fitted coefficients, classifiers and pointers. State that extra intermediate supervision and structural knowledge prevent an equal-budget causal comparison.",
    "Score complete field equality and each count, colour, coordinate, layer, emphasis, typed link and region channel. Score source-to-target colour changes separately from unchanged colour retention; absent target cells count wrong.",
    "Distinguish actual appended-cell birth links and birth count decisions from retyped triples and new old-cell endpoint pairs. Preserve integer TP/FP/FN and false positives on no-birth scenes.",
    "For eight-turn trajectories carry persistent original labels through every actual canonical map in each independent trajectory, starting on turn one. Report turns four and eight, surviving-original coverage and unmatched originals, with all-original graph counts that penalize omissions.",
    "Birth labels matched by scene, turn and append offset form a secondary cohort convention. They are neither discovered persistent identity nor semantic memory.",
    "Report local agreement against teacher(current model state) separately from accumulated divergence against the independently evolving teacher state.",
    "Retain errors, failed trials and development revisions. No model selection or repair may follow reading final results without explicitly reclassifying those inputs as development.",
    "Use paired scene resampling for four descriptive intervals on preselected deployment-versus-Job137 primary and hard birth F1, and turns four and eight original-colour agreement, retaining integer numerator/denominator. Do not present them as population or meaning uncertainty."
  ],
  "scope": "Learned numerical emulation of known authored response rules. The supplied philosophical syntax guides interpretation and design, not semantic ground truth. No score establishes experienced colour, semantic comprehension, consciousness, general intelligence or superiority to an otherwise matched symbolic representation."
}