{
  "session": 63,
  "question": "Which prior numerical results depend on the three observed separator-class corrections?",
  "evidence": "Use only the three overlays in boundary_model.json; retain the immutable master and unresolved split pairs.",
  "compatibility_view": "Replace each overlay's literal hyphen with semicolon in memory only. This tests the historical code's three-dot convention, not a proven semantic or key operation.",
  "audit": ["Hash and statically inventory retained Python scripts", "Compare rune values, historical and current word records, line/page/section structure, word-length blocks, training inputs, counter indices and reset flags", "Check selective-F and shuffled-calibration target coverage", "Check Session27 feature matrices under its literal predeclared punctuation definition"],
  "affected_targets": [7, 8],
  "session2_replay": "Reconstruct the original candidate records on affected sections and reproduce saved top10. Recompute only changed clause branches on the compatibility view, keeping original eight streams, offsets0..64, signs, F policies, two scopes, ranking and tie order. Merge unchanged records without testing new rules.",
  "session3_replay": "Compare all clause counter prefixes and full-stream eligibility under the original settings. If these are identical, rankings are invariant; replay the seven saved clause candidates in affected top20 lists with fixed parameters and compare continuations. If an invariant fails, report it before expanding computation.",
  "limits": ["Retrospective sensitivity analysis, not independent validation", "No new keys, offsets, models, source interpretations or searches", "Score changes are not decryption evidence", "No claim of complete image auditing or universal reset rules", "No new random calibration where target inputs remain identical"]
}