AI Assistant Refinements & Uncertainty Engine

The Uncertainty Engine saw Phases 2, 3, and 4 complete, adding ContradictionClassifierService, ReconcileAction, and refining source reliability checks in ActDispatcherService.

The ContradictionClassifierService now acts as an LLM pair classifier with a 600ms time-box, screening temporal versus true contradictions.

ReconcileAction autonomously handles drift by sampling traits and concepts to create uncertainty records or supersede temporal changes.

Surfacing and Resolution features include the placeholder and a resolution feedback loop in UserTraitService when traits are corrected.

Self-Tuning mechanisms were added, including downgrade_overexposed() to prevent anti-nag and resolve_by_reinforcement() for auto-resolution.

In document handling, embedding is now eagerly inserted into documents_vec synchronously after creation to ensure duplicate detection works before async processing.

The privacy export endpoint was optimized to stream results incrementally, preventing timeouts on large datasets by fetching rows in batches of 500.

Several database and tool handling fixes improved stability, including correcting migration issues on fresh installs and handling tool registration race conditions on startup.

• Uncertainty Engine: Phases 2, 3, & 4 are complete.

• Implemented streaming privacy export to prevent timeouts on large datasets.

• Embeddings are now inserted synchronously after document creation for duplicate detection.

• ContradictionClassifierService uses an LLM pair classifier with a 600ms time-box.

• Fixed database migrations (002/003/005) that crashed on fresh installation.

• Tool registration now handles race conditions by registering trusted tools before dependency installation.