# Adapter experiment record

- Task and measured gap:
- Baseline model/prompt/context/tool configuration:
- Data rights, privacy review and provenance:
- Family/group-aware train/development/test split:
- Duplicate and evaluation-leakage checks:
- Base-model/tokenizer revisions and chat template:
- Pinned training environment and configuration hash:
- Loss mask and synthetic tokenisation inspection:
- Profiled throughput, memory, cost and failure budget:
- Target metric and required safety/non-target slices:
- Blind rubric and disagreement handling:
- Base versus adapter serving identity/load/rollback checks:
- Stop/retrain/retire owner and triggers:

Illustrative support fixture: policy excludes personalised goods. Warm wording must preserve the exclusion and must not invent a refund. Include an eligible-return counterpart to detect over-refusal. This is a test design, not a measured training result.
