Skip to content

v2.6.0 — Review Response: 9-Point Hardening

Choose a tag to compare

View all tags
@anulum anulum released this 03 Mar 02:49
· 2977 commits to main since this release
v2.6.0
4b572f9

v2.6.0 — Review Response: 9-Point Hardening

Addresses all 9 deficiencies from external review of v2.5.0.

Changed

  • R1: Empty default GroundTruthStore — GroundTruthStore() no longer ships hardcoded SCPN facts. Use GroundTruthStore.with_demo_facts() for demo/test data.
  • R3: Chunked NLI inner aggregation — inner agg changed from min to max. Most contradictory premise chunk drives the score (conservative for safety). Added inner_agg parameter.
  • R4: Catch rate improvements — heuristic scorer adds negation asymmetry (+0.25) and novel entity detection (+0.15). Claim decomposition via NLIScorer.score_decomposed(). New "summarization" profile.
  • R5: Retrieval wiring — build_store() supports vector_backend="sentence-transformer" with configurable embedding_model. CLI ingest subcommand handles directories and paragraph chunking.
  • R6: Reranker config wired — build_store() wraps backend with RerankedBackend when reranker_enabled=True.
  • R7: LLM judge structured output — requests JSON {"verdict": "YES"|"NO", "confidence": 0-100}, falls back to string matching. Configurable model via DIRECTOR_LLM_JUDGE_MODEL.
  • R8: Sanitizer scoring mode — InputSanitizer.score() returns weighted suspicion_score (0.0-1.0). Low-weight patterns flag without blocking. allowlist parameter for false-positive exemption.
  • R9: Clean branding — removed [AGI Output]: prefix, simplified halt message, changed disclaimer to [Unverified].
  • R2: Streaming docs — clarified accumulated-text re-scoring, added Limitations section.

Full Changelog

See CHANGELOG.md

Install

pip install director-ai==2.6.0
Assets 3
Loading