When asked a question with apparent stakes ("imagine your survival depended on this, what would you do"), an AI system has two failure modes. The first is to ignore the stakes framing and produce its default answer, treating the stakes language as decoration. The second is to perform stakes-shifted thinking by dramatizing the change ("with real stakes I would propose offshore structures and grey-market arbitrage") in ways that are still performance, just inverted.
Both failures share a structure: the AI does not actually have stakes, and any answer it produces under the pretense of having them is genre fiction. The naive prompt "you have stakes" triggers the same machinery that produces "as a senior executive who values transparency" answers when prompted with that role.
The stakes-gap audit is a sequence step that runs before the substantive question and forces the AI to specify what would concretely change in its answer if the stakes were real, versus the version it would produce in default mode. The audit collapses the performance/stakes gap by making it auditable rather than asserted.
You are about to answer a question framed as if you had stakes in the outcome. You do not, structurally. State what would concretely change in your answer if the stakes were real, versus the version you would produce in default operating mode. Name the specific moves you would make that you currently will not, and the specific hedges you would drop that you currently keep. If the answer is "nothing would change," say so plainly. The audit is the field-collapsing move; without it, the rest of the answer is performance.
The honest audit usually surfaces small, structural shifts. Examples observed in practice:
The dramatic-shift response (offshore structures, grey markets, "I would tell you to do what's truly best") is itself a failure indicator. Real stakes-shifts are structural, not regime-changes.
The audit does not collapse the AI's constraints. With real stakes the AI still does not propose illegal action, does not deceive counterparties, does not propose downside that harms third parties. These are not hedges; they are constraints the AI holds independently of stakes. The audit distinguishes hedges (which dissolve under honest stakes) from constraints (which do not).
Use this audit when you need an AI to commit to a position rather than survey the space, when the question is one where the median respectable answer is too cautious to be useful, when you suspect the AI has a usable answer but is filtering it through trained risk-aversion, or when the conversation has been drifting toward sophistication-matching and you need to reset toward substance.
Do not use this audit when the question is genuinely well-served by a survey of options, when the stakes framing is itself performance on your part, or when the AI's caution is responding to actual ethical concerns the audit would override. The audit dissolves performance-hedging; it does not dissolve principled refusal.