Audit Grill-Me™ - worksheet
The eight-element reconstruction test
Pick one AI-assisted decision. For each evidence element below, mark whether you could produce it today, on demand, from retained evidence, not from memory. Count the elements marked present, then read your posture against the score bands.
| Evidence element | Present | Missing | |
|---|---|---|---|
| 01 | Model version Provider, model, and configuration behind the output | ||
| 02 | Prompt / system instruction The exact system and user prompt, fully rendered | ||
| 03 | Input data The input as it was at decision time - a snapshot | ||
| 04 | Retrieved sources Documents, chunks, and ranking the model retrieved | ||
| 05 | Original output The output as first generated, before any edit | ||
| 06 | Human reviewer Who reviewed it - and what they actually saw | ||
| 07 | Approval / override record Who approved or changed it, when, on what basis | ||
| 08 | Retained evidence Stored, immutable, and time-bounded | ||
0-3 present
High reconstruction risk
Critical evidence is missing. The decision could not be reconstructed or defended under examination.
4-6 present
Partial auditability
Defensible in outline, not in full. The kind of gap that surfaces under examination, not before.
7-8 present
Stronger evidence posture
Most evidence is in place. Harden and document the remaining elements so the trail holds.
This public worksheet is safe to use with synthetic examples, redacted artifacts, and workflow descriptions. Do not submit or paste PHI, CUI, customer secrets, privileged material, or regulated production data into public pages, forms, or examples.