Execution-based bug identity

Text similarity does not decide whether two reports describe the same bug. The dedup study revalidates each sealed VERIFIED Case on its own target and fix, then runs test A on fix B and test B on fix A. Only deterministic mutual passes count as an equivalent pair. Infrastructure failures and flaky cross-results stay inconclusive.

The local bug-identity-corpus/v1 manifest contains path-contained Case files and checkouts:

{
  "schema_version": "bug-identity-corpus/v1",
  "cases": [
    {"case": "cases/a.json", "target": "repos/a-bug", "fixed": "repos/a-fix"},
    {"case": "cases/b.json", "target": "repos/b-bug", "fixed": "repos/b-fix"}
  ]
}

cd engine
python3 -m exhibit_a.cli dedup ../corpus/manifest.json --docker

Reports use bug-identity/v1. The exact pair matrix is authoritative. Reported clusters use conservative complete-link grouping: a Case joins a cluster only when it is mutually equivalent with every member. Inconclusive pairs therefore never merge clusters. This is a research view and cannot alter any underlying Case verdict.