Learning loop

Outcome: Understand the opt-in path from outcomes to better local routing (and optional fine-tunes).

Flow

1. Capture — implicit outcomes + optional accept/reject feedback
2. Tune — per-category confidence thresholds (daari learn recommend)
3. Examples — opt-in capture for datasets
4. Fine-tune — MLX LoRA (daari learn finetune) → daari learn deploy
5. Collective — anonymized stats export only (daari learn export-stats); never prompts
6. Propose defaults — daari learn propose-defaults writes YAML proposals (never auto-promotes)

Principles

• On-device by default
• Review-gated promotion
• Metadata-only for any shared export

Next

→ Config overview · PRD deep-dive (internal): docs/prd/learning.md