Skip to content

v3.9.5

Choose a tag to compare

View all tags
@anulum anulum released this 22 Mar 23:44
· 2399 commits to main since this release
v3.9.5
0f97ac7

What's Changed

Features

  • Coalesced review_batch(): batches logical and factual NLI pairs through NLIScorer.score_batch() — 2 GPU forward passes total instead of 2*N. Dialogue items fall back to sequential review().
  • Per-task-type judge escalation: _should_escalate() now receives the detected task type at all 4 call sites. Per-task thresholds: dialogue=0.35, summarization=0.25, qa=0.30, fact_check=0.20.
  • 17 new tests (no-KB calibration, cross-turn blending, judge threshold differentiation).

Fixes

  • Scorer: _finalise_review() log message uses effective threshold override instead of base threshold.
  • Scorer: meta-classifier NLI-to-coherence threshold conversion uses actual weights instead of hardcoded 0.4/0.6.
  • 50+ documentation corrections across docs-site, notebooks, README, ROADMAP, CONTRIBUTING, CITATION.cff — verified by dual audit (Claude + Gemini).

Full Changelog: v3.9.4...v3.9.5

Assets 5
Loading