dom da34bdc8d7 feat(pmsi): add DP candidate pool + pool rank LLM + benchmark ...

- DPPoolCandidate model (terme, section, preuve, score_initial)
- build_dp_candidate_pool() with filters (_is_pool_excluded, _dedup_pool)
- Pool exclusion: admin noise, bio values, vague symptoms, place names
- DP_POOL_RANK template for LLM-based ranking among pool candidates
- llm_dp_pool_rank() with guardrails (GF-1 evidence, GF-3 confidence)
- benchmark_quality.py: --dp-candidates, --use-dp-pool-rank flags
- 41 new tests (pool, exclusion, dedup, pool rank, synthese)

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

2026-02-24 00:06:44 +01:00

62 KiB

Raw Blame History

View Raw