Commit 5f4f62e
authored
feat(bench): postgres_ingest Cat 7 LoCoMo E2E QA 0.392 ≈ flat + adapter wiring (#168)
postgres_ingest LoCoMo-10 E2E QA = 0.392 (n_scored 250, n_error 0) vs flat anchor 0.384; Δ +0.8pp = noise. Swapping flat's in-memory store for postgres+pgvector does not move QA — confirms storage substrate ≈ flat on both retrieval (Cat 1/2c 0.833 identical) and QA. MemPalace's lift is the graph/reader, not the backend.
Caveat: 4/250 reader calls (conv-44 q40/q47/q56/q157) hit Azure content_filter (hate-medium, non-deterministic corpus×reader artifact); answer_generator degrades to empty→judged incorrect, so 0.392 is conservative (flat hit 0 filters on 2026-05-29). postgres ≥ flat despite the handicap.
Adds: postgres-ingest adapter CLI registration, postgres factory in cross_validate (requires SME_POSTGRES_DSN, never prod), ingest script, runner, + 2 baselines. Additive-only (0 deletions). Closes #168.1 parent 58f06b2 commit 5f4f62e
6 files changed
Lines changed: 10970 additions & 0 deletions
0 commit comments