Source: qa/astra-r2-request.md
# QA1 request: Astra Round 2 citation audit and statistics
Status: ready for independent QA1 review; **not approved**. Lead: please dispatch QA1 as required by `neov2/specs/astra_r2.md`.
Worktree: `/tmp/neochart-astra-r2`, branch `astra-r2`; deliverables will be pushed to origin/master. Original shared checkout was deliberately left undisturbed because other authors have active edits.
Review scope:
- `neov2/xref/resolve_mentions.py` and `patches/0002-opt-in-research-annotations.patch`: explicit alias binding; mixed authorities; source offsets; opinion-local act identities; unresolved diagnostics; no guessed section or target URL.
- `audit_v2.py`: 10,000-page SRS frame → 2,000 actual court-label/decade-stratified pages; quota constraints, hashes and weights.
- `review_metrics_r2.py` and `out/r2/{precision_review,recall_gold,review_metrics}.json`: 50 hand-checked resolved mentions per seven families; Act family uses 32 main-sample +18 supplemental mentions. Supplements are excluded from recall. Root corrected initial reviewer mistakes before final labels.
- `out/r2/report.md`: conditional weighted detection recall 84.38% (70.74–92.97%) versus resolution recall 19.06% (4.03–58.82%). Review the denominator and uncertainty limits; these are NOT deployed-link or whole-corpus recall guarantees.
- `stats_v2.py`, tests, and `out/stats_v2*`: decade/title recent rankings, complete-period normalized velocity, cases and co-citation pairs. Critical caveat: 2021–2025 has only 244 case records; graph has 406 keys outside expected USC title range.
Author verification: 58 tests pass, patch apply/compile and embedded-source equality pass, 246 source pages checked for exact review offsets. Corpus read-only; no neochart.com build/deploy; authoring integration remains an unapplied patch. Check cached-IR behavior before approving any future opt-in build.
Please place findings in `neov2/qa/astra-r2-QA1.md`. No permission or production publishing action is requested by this note.