Evidence layer
Probative surfaces brought into scope by this page
This page does more than point to governance files. It is also anchored to surfaces that make observation, traceability, fidelity, and audit more reconstructible. Their order below makes the minimal evidence chain explicit.
- 01Evidence artifactconcept-registry.json
- 02Evidence artifactbridge-vocabulary.json
concept-registry.json
https://gautierdorval.com/concept-registry.json
Published surface that contributes to making an evidence chain more reconstructible.
- Makes provable
- Part of the observation, trace, audit, or fidelity chain.
- Does not prove
- Neither total proof, obedience guarantee, nor implicit certification.
- Use when
- When a page needs to make its evidence regime explicit.
bridge-vocabulary.json
https://gautierdorval.com/bridge-vocabulary.json
Published surface that contributes to making an evidence chain more reconstructible.
- Makes provable
- Part of the observation, trace, audit, or fidelity chain.
- Does not prove
- Neither total proof, obedience guarantee, nor implicit certification.
- Use when
- When a page needs to make its evidence regime explicit.
Stochastic fixation
Stochastic fixation designates the subcase in which a non-deterministic model realization is captured by a delivery layer, associated with a cluster of nearby queries, and re-served as if it represented the reference answer.
The term is used here in continuity with the mechanism described by Melanie Maquet: a semantic cache does not necessarily store a canonical answer. It may store one particular sample from a distribution of possible outputs, then industrialize that sample.
Short definition
Stochastic fixation appears when an output that could have varied at each call becomes stable because it was stored, not because it became more true.
It turns probabilistic variation into application-level stability. That stability may help cost and latency, but it may also freeze an incomplete, stale, miscategorized, or merely acceptable reconstruction at the moment of the first call.
Difference from interpretive variability
Interpretive variability observes response dispersion across systems, contexts, or moments. Stochastic fixation describes a mechanism that may artificially reduce that dispersion by freezing one realization.
Repeatability must therefore not be confused with fidelity. A stable answer may be stable because it is correctly governed. It may also be stable because a cache keeps serving the same erroneous reconstruction.
Audit implication
Stochastic fixation requires a distinction between the model’s native reconstruction and the application-delivered reconstruction.
Direct model interrogation may show good alignment. The final application may keep serving an older answer. Conversely, a highly stable delivered answer may hide native model uncertainty because the delivery layer removed visible variance.
What it is not
Stochastic fixation must not be used as an automatic accusation against a platform. Without cache logs, technical documentation, controlled repetition, or application-level access, it remains a mechanistic hypothesis rather than proof.
It also does not mean that semantic caching is bad. It indicates that an economic optimization layer may freeze a version of the reconstruction that is not the best, freshest, or most faithful.
Reading rule
Use stochastic fixation when observed stability may come from the freezing of a non-deterministic output at the delivery layer.
If that hypothesis cannot be demonstrated, use the more cautious term delivery-layer fixation or unattributed delivered stability.