Human-in-the-loop teach-back
AI output on each card is a proposal, not a decision.
Grounding / severity / actions still follow TRUST_GROUNDING.md and ADVISOR_WHY.md. Teach-back never auto-overrides trust rules.
Per card
| Control | Effect |
|---|---|
| Accept | Keep proposed action / summary / why; mark human_status=accepted |
| Adjust | Edit fields; store before→after; mark adjusted |
| Teach back | Free-text instruction for future models; mark taught |
| Reject | Mark wrong without replacement yet |
Capture (for later summarizers / prompt updates)
data/feedback.jsonl— append-only event logdata/teachback_index.json— deduped instruction candidatesdata/teachback_instructions.md— human-readable review file
Nothing is auto-injected into grounding.py or cited_facts.py. Promote candidates deliberately (Admin → Teach-backs), then wire into the next model/methodology pass — still subject to citation rules.