Sam Gilson

Hallucination autopsy procedure

Definition

The hallucination autopsy procedure is a teaching routine for checking a chatbot’s proof, locating an error, and examining a possible explanation for it.

Explanation

Start with the claim the problem actually asks you to prove. Then compare each step with that claim and the assumptions you are allowed to use. The aim is to find a specific divergence, rather than decide whether the whole answer sounds convincing.

LabelPurposeAction
A0Calibrate your relianceBefore reading, write down how much you expect to rely on the response for this kind of problem and why. Revisit that judgment after checking it.
A1Restate the claim as givenWrite the exact claim, including its assumptions, domain, and right-hand side.
A2Extract the rule the chatbot actually usedIdentify the formula or inference used in the response. Compare it with the rule the response says it used.
A3Locate the first divergenceMark the first line that no longer follows from the stated claim, assumptions, or earlier steps.
A4Test a small caseCheck an appropriate small case by hand. A counterexample can refute a general claim; matching cases do not prove it. Check the derivation as well.
A5Classify the failure modeDescribe the error. Use axiom reversion, invalid syntactic shortcut, or spatial or counting miscount when one fits; the list is not exhaustive.
A6Explain why the failure was likelyOffer a possible explanation and distinguish it from what you observed. The response alone does not reveal the model’s training data or internal cause.

After checking, return to your initial judgment about reliance. What evidence changed it? That reflection is useful even when the response turns out to be correct.

Scope and evidence

Procedure A was taught in the Summer 2026 classroom study and presented in the Game AI lecture. The preprint reports an ongoing analysis; the effectiveness of this procedure has not been established.