WIP: build ground truth corpus and calculate recall

[RFOxbury]

Test this code by:

1. collecting a sample of 15 systematic reviews with uv run --project backend python scripts/eval_ground_truth/fetch_review_corpus.py --limit 15

2. calculating recall (at both the search stage, and the screening stage) for those 15 reviews with uv run --project backend --env-file backend/.env python scripts/eval_ground_truth/run_and_score.py --corpus scripts/eval_ground_truth/results/corpus.json

...
...
...
...
...

What / why

Task: docs/tasks/<task-id>/

Proof it works

Evidence lives in docs/tasks/<task-id>/verification.md — summarise here, don't duplicate the table.

• make verify: pass / fail / stub
• Manual / end-to-end: what was exercised + the exact command (or link to verification.md)

Risk tier

Tier N — why.

Tier	Review
0–1	tests + AI review or human skim
2	contract verifier + tests + human review
3	+ security + adversarial review + human deep review
4	+ human-approved plan + ADR + rollback plan

AI role

Review focus

Correctness · missed requirements · security · provenance integrity · scope creep · over-abstraction.

Reviews run

Findings recorded in verification.md.

• Contract verifier
• /code-review
• /security-review
• Adversarial review (Tier 2+)
• /simplify

Known gaps & deferred seams

Public safety

• No secrets, credentials, or real/acquired source text in the diff or evidence.
• Logs / traces / screenshots are public-safe.
• No approval-gated change (schema · auth · runtime egress · deps · CI · prod config · public interface · scaffold) snuck in unapproved. (Runtime egress = the running product reaching search/model providers with project data; agent/dev-time lookups, MCP and installs are fine.)