Appendix A — The Drafter-to-Evaluator Loop
Use this optional method when an AI draft helps. The Evidence Locker can also be completed without a model. The companion book develops the conversation method further.
A.1 The loop, in five moves
- Define the task and permitted inputs. State the criterion, population, period and intended output. Use fictional exercise records unless the actual data and tool have been authorised.
- Obtain a draft. Keep the prompt, model/version if available and the supplied evidence. Label the output as a draft, not a verified result.
- Check the reasoning. Recompute numbers and trace material claims to their source. Ask where a missing input affects the conclusion.
- Resolve discrepancies. Correct unsupported claims, retain sound work and preserve unknowns. Do not require disagreement if the draft is supported.
- Approve and retain. Record your final judgement, changes and limitations. The responsible person reviews the evidence; the tool does not assume that accountability.
A.2 Failure modes to check
| Possible error | Check |
|---|---|
| Generic objective | Name the actual boundary, criterion and observable outcome |
| Incorrect mapping | Verify the source category and distinguish partial overlap from equivalence |
| Unsupported sample | Record selection method and assumptions; distinguish a census |
| Operation inferred from policy | Request the record of performance |
| Wrong legal shortcut | Establish jurisdiction, facts, exceptions and qualified review |
| Plausible but unsupported cause | Corroborate the mechanism or mark it provisional |
| Default recovery target | Use the approved service criteria and BIA |
| Invented closure | Require the actual retest population, results and limitations |
The chapter Field Notes are illustrative teaching drafts, not measured transcripts of a particular model. They show errors worth checking for without asserting that every model makes them.
A.3 A two-sided check
Check the draft for excessive reassurance and excessive suspicion. “All six leavers passed” is unsupported when R06 lacks an event. “R06 is definitely still active” is also unsupported. A careful result preserves the distinction between a missing record and a demonstrated failure.