Source: assist/REVIEW_REQUEST2.md
# Lead: dispatch QA1 on frozen Round 4 cases 21–40 The redesigned proposer is frozen at SHA256: `8460b86e79fbea0787351c3e4332736c39534c4d03000df01401c0db35db0c1b` The commit introducing `r4/freeze.json` is the publication record. The manifest was first committed locally as `e1f1ebbd7bbdbfe1d0e4ec510de7e8cb8b4ffd35` before generation; the published commit rebases the identical freeze onto current master. Files and model identity were revalidated after rebase. `r4/generation.json` pins all packet hashes/times. No held-out quality metric has been calculated. Please dispatch **QA1 as an independent annotator** for `assist-021` through `assist-040` only. Do not use development cases1–20 as held-out evidence. 1. **Blind Phase A:** read canonical full sources from `r4/sample.json`, without opening either old `proposals/assist-021..040.json` or new `r4/proposals/`. Use the LAW catalogue to inventory applied factual evidence, recording polarity/role qualifiers and negative/partial cases honestly. The developer has not labeled these cases and must not supply ground truth. 2. Save independent inventories in `r4/annotations/<sample_id>.json`. Retain `reviewer: QA1`, `source_sha256`, `blind_before_proposals`, `review_complete`, and exact Unicode `[start,end)` source spans. Use fully qualified element IDs such as `42usc1983:color-of-law`, with explicit `law_id` on each gold item. The original QA1 development files remain unchanged. 3. **Phase B:** only after Phase A is saved, inspect new `r4/proposals/*.json`. There are20 packets and75 proposals, with empty packets allowed. Judge the element/passage relevance independently. Copy proposal IDs exactly and verify every judgment references an existing proposal. Null variants are explicit abstentions, not claimed exact matches: set `variant_accept:false` and report variant coverage zero / precision undefined instead of inventing labels. 4. Distinguish quote containment, partial overlap and substantive acceptance. Report precision over recognized independent judgments, recall against the complete blind inventory, and case coverage. State the matching convention; do not silently conflate QA1's earlier2/19 recall with the strict-containment development diagnostic. Empty-gold cases need no invented evidence. 5. Record actually measured review times, or null if not measured. Sequential Phase B times are confounded by prior reading; no causal effort-saving claim. The packet quotations, source hashes, target-law identifiers and freeze hashes have been mechanically verified. That is not a legal quality sign-off. Review findings may be written to `neov2/qa/assist-QA1-r4.md`. The existing independent scorer can be invoked as: ```bash python3 neov2/tools/score_assist.py --root neov2/assist/r4 ``` Do not mix these frozen packets with the original Round 3 packet IDs or score files. No independent annotations or held-out quality values have yet been supplied. Root tools are available in the published implementation; this request is the concrete handoff to Lead, not a claim that a reviewer has been dispatched.