# Model adaptation decision

- Task, consequence and acceptable fallback:
- Frozen baseline: model, prompt, tools, data and configuration:
- Failed cases and expected source-grounded result:
- Candidate intervention and single changed factor:
- Development/test split and leakage check:
- Per-slice quality, permission/freshness, latency and full cost:
- Decision: retain baseline / change prompt / change evidence path / compare model or tool / tune / stop:
- Recheck trigger and rollback:

Illustrative policy case: current P2 says 14 days with a personalised-goods exclusion; historical P1 says 30 days. Supply P2 directly before assuming retrieval will help. A stale retrieval result and a model ignoring a supplied exclusion are different failures. No outcome is measured by this worksheet.
