Stochastic fixation designates the subcase where a non-deterministic model output is fixed by a delivery layer, associated with a semantic neighbourhood, and re-served as a reference answer.
The mechanism comes from Melanie Maquet’s work on semantic caching and reconstruction fidelity. Pagup uses it as an audit guardrail: a stable answer may be reliable, but it may also be a frozen probabilistic output.
Why the term matters
The visible stability of an AI answer can be misleading. If an imperfect first generation is cached, later users may receive the same reconstruction without the model being truly queried again.
In that case, the audit does not measure only what the model knows or reconstructs. It also measures what the application chose to keep.
How Pagup uses it
In an audit, stochastic fixation is not assumed as a fact by default. It forces a question: is the observed answer fresh, native, delivered, cached, routed, or impossible to qualify?
For the primary doctrinal definition, see gautierdorval.com. For normative qualification, see interpretive-governance.org.