[aisona-lab]

What

The fail-on gate now reads hunk recall instead of rule recall. Hunk recall is the share of in-scope defective hunks that draw at least one finding, under any rule. Rule recall is still printed, but it no longer gates anything.

Why

In run 2 (#13), the reviewer described two labelled defects correctly but under rules other than the label: R2/R13/R16 for an R7 label, and R7/R13 for an R4 label. The rules overlap, so rule recall partly measures which name the reviewer picked rather than whether it saw the defect.

Pre-registration

• This decision comes after runs 1–2, and it changes neither outcome. Re-scored with this code, hunk recall is 1/15 for run 1 and 3/15 for run 2, both under the 50% floor.
• It is committed now so that Stage 3a is scored against it, not fitted to it.
• The amendment is in docs/hardening-plan.md.

Known weakness: a finding about a different problem in the same hunk still counts (1 of 3 in run 2). unlabelled is printed so a human can check.

Proof

env -u ANTHROPIC_API_KEY uv run pytest -q          # 173 passed
env -u ANTHROPIC_API_KEY uv run python scripts/corpus_cli.py prove corpus/seed.jsonl   # OFFLINE_PROVE: PASS
ruff check . && black --check . && mypy src        # clean

New tests:

• A defect found under a different rule counts toward hunk recall but not rule recall, and the gate reads it.
• A defect whose labelled rule is outside the preset is excluded from hunk recall.

---

Generated by Claude Code