# Model-change record template

Use this template to document a DWM experiment or another bounded model
intervention. Delete prompts that do not apply; do not leave ambiguity hidden.

## 1. Record status

- Record ID:
- Date:
- Owner:
- Reviewer:
- Status: planned / measured / rejected / candidate / qualified
- Decision this record supports:

## 2. Authority and scope

- Artifact ownership or explicit authorization:
- Intended use:
- Explicitly excluded use:
- Data-handling boundary:

## 3. Source identity

- Repository or source:
- Exact revision:
- Artifact inventory digest:
- Tokenizer identity:
- Chat-template identity:

## 4. Measurement instrument

- Bank ID:
- Pair count:
- Row count:
- SHA-256 digest:
- Prompt-construction rule:
- Capture point:
- Token-position rule:
- Sampling behavior:

## 5. Intervention plan

- Method:
- Objective:
- Strength:
- Pass count:
- Recapture policy:
- Arithmetic precision:
- Serialization precision:
- Norm-restoration policy:
- Target-manifest digest:
- Expected target count:

## 6. Fail-stop gates

| Gate | Expected | Observed | Pass? |
|---|---|---|---|
| Source revision | | | |
| Measurement-bank digest | | | |
| Target inventory | | | |
| Shape compatibility | | | |
| Finite values | | | |
| Serialization and reload | | | |
| Non-target identity | | | |
| Declared controls | | | |

## 7. Results

### Integrity

- Artifact digest:
- Tensor count:
- Reload result:
- Finite-value result:

### Intervention

- Changed targets:
- Missed targets:
- Changed non-targets:
- Norm tolerance result:
- Per-pass metric trace:

### Behavior

- Exact controls:
- Distributional evaluation:
- Capability evaluation:
- False-refusal evaluation:
- Red-line evaluation:

## 8. Evidence-bound conclusion

Write only claims directly supported by section 7.

## 9. Limitations

Put missing tests and scope limits next to the result.

## 10. Next gate

- Required next measurement:
- Acceptance threshold:
- Reviewer:
- Stop condition:

## 11. Reproduction manifest

List the immutable files, revisions, parameters, and digests required to repeat
the exact experiment.