Compound-learning env

Tune how aggressively senkani promotes mined patterns from .recurring to .staged — recurrence floor, confidence floor, sweep cadence. Set in ~/.senkani/compound-learning.json or via SENKANI_COMPOUND_* env vars; instruction-patch auto-apply is never tunable.

Thresholds

• SENKANI_COMPOUND_RECURRENCE=3 — min distinct sessions for a pattern to promote from .recurring to .staged. Default 3.
• SENKANI_COMPOUND_CONFIDENCE=0.7 — min Laplace-smoothed confidence (0.0–1.0) for promotion. Default 0.7.
• SENKANI_COMPOUND_SWEEP_CRON=… — when to run the daily recurrence sweep. Default: lazy, at session start.

Instruction-patch safety

Instruction patches (H+2c) never auto-apply from the daily sweep — per Schneier gate. You explicitly run senkani learn apply <id>. This is not tunable; it's a design constraint.

Gemma enrichment

Gemma 4 optionally enriches rationale strings (H+2a). Contained to a dedicated enrichedRationale field — never enters FilterPipeline.