M42: calibrate grounded RAGcore fallback
This commit is contained in:
@@ -1,5 +1,23 @@
|
||||
# Project state
|
||||
|
||||
## M42 — calibrate grounded RAGcore fallback (2026-08-21)
|
||||
|
||||
- The first live five-minute acceptance run exposed a real calibration error in M41's
|
||||
new fallback threshold: `/v1/answers` timed out, while `/v1/search` correctly returned
|
||||
`damage-procedure.md` at rank one with fused score `0.0163934426`. RAGcore uses
|
||||
reciprocal-rank fusion (roughly `1 / (60 + rank)`), so the normalized-looking `0.05`
|
||||
threshold could never accept a legitimate result.
|
||||
- Set the default minimum to `0.016`: this accepts the first two normal RRF ranks while
|
||||
still rejecting missing scores and the existing low-score `0.01` abuse case. Damage
|
||||
questions continue to require explicit damage evidence and out-of-domain questions
|
||||
continue to return `insufficient`; no generated answer is trusted when retrieval is
|
||||
absent or weak.
|
||||
- Added a regression using RAGcore's observed rank-one fused score. Focused knowledge
|
||||
tests **36 passed**; complete isolated PostgreSQL backend suite **270 passed**; Ruff and
|
||||
mypy are clean and `git diff --check` passes.
|
||||
- Exact next action: commit and push M42, deploy that exact archive over the healthy M41
|
||||
production deployment, then repeat the live five-minute demo and final health checks.
|
||||
|
||||
## M41 — full review remediation and hardening (2026-08-21)
|
||||
|
||||
- Closed all findings from the repository-wide review without expanding the locked PoC:
|
||||
|
||||
Reference in New Issue
Block a user