Model version
Provider, model, and configuration behind the output
Audit Grill-Me methodology
The Six-Month Reconstruction Test
A practical way to test whether one AI-assisted decision can be reconstructed, reviewed, and defended after the fact.
Reconstruction question
Six months from now, can you reconstruct the evidence behind a specific AI-assisted decision?
The method is deliberately narrow: one workflow, one decision, and the evidence trail that should exist when someone asks how the AI-assisted result was produced.
The eight elements
For one real AI decision, can you produce all eight?
The test is evidence-first. Policies and principles may explain intent, but these records determine whether the decision can be reconstructed under examination.
Prompt / system instruction
The exact system and user prompt, fully rendered
Input data
The input as it was at decision time - a snapshot
Retrieved sources
Documents, chunks, and ranking the model retrieved
Original output
The output as first generated, before any edit
Human reviewer
Who reviewed it - and what they actually saw
Approval / override record
Who approved or changed it, when, on what basis
Retained evidence
Stored, immutable, and time-bounded
How to run it
A reconstruction test, not a generic AI maturity survey.
Choose one decision
Start with one AI-assisted decision that matters enough to defend: a recommendation, routing, score, answer, memo, or approval support artifact.
Freeze the question
Ask whether the organization can reconstruct the decision six months later from retained evidence, not interviews, memory, or recreated context.
Produce the eight elements
Look for the model version, prompt, input snapshot, retrieved sources, original output, human reviewer, approval record, and retained evidence.
Score the trail
Treat missing evidence as a reconstruction gap. The goal is not a policy score; it is whether the actual decision can be reviewed after the fact.
Remediate in order
Close the evidence gaps that most affect defensibility first: provenance, retention, review binding, override trails, and source linkage.
Try the public simulator
Run the reconstruction test on fictional workflows before touching any internal system.
Open simulatorTake the risk check
Answer eight questions and get an instant posture read for one workflow.
Take risk checkView the 10-day sprint
See how the method becomes a fixed-scope evidence-readiness packet.
View sprintPublic pages and the simulator are synthetic-data safe. Do not submit PHI, CUI, customer secrets, privileged material, or regulated production data through public forms or examples.