wa:evolve

Confidence scoring for heuristic candidates evaluates occurrences and applies caps, boosting scores for cross-session evidence and correction signals. This filtering ensures only high-confidence behavioral patterns are promoted into your active guidance system.

Human-in-the-loop deduplication presents clustered heuristic candidates for your approval, allowing optional edits before appending them to the heuristics store. This prevents duplicate rules and ensures quality control during promotion.

To prevent reprocessing duplicate observations in JSONL pipelines, this Skill prunes promoted entries after they are approved and appended. Lifecycle hygiene ensures consolidated observations are not re-evaluated in subsequent clustering runs.

To implement confidence-scored heuristics, you need JSONL logs of behavioral observations and configurable thresholds for scoring. The pipeline clusters events and applies cross-session boosts to validate candidate rules.