FoxEIS is a receipt-backed evidence protocol for testing what local models actually did.
Most AI evaluation starts after the model speaks. Someone reads the output, scores whether it seems better, and argues about whether the change mattered. That is useful, but it misses a harder question: what happened inside the inference run?
FoxEIS treats that question as a court.
A seam is a scoped place where the inference process can be observed or perturbed. A witness checks whether the intervention physically applied. An effect is a clean token or text movement that survives control checks. A predictor is a candidate rule that may explain where effects appear. Authority is permission to use that evidence for influence.
Those are separate. FoxEIS is built to keep them separate.
The public pipeline looks like this:
- Model manifest
- Runtime doctor
- Seam passport
- Island atlas
- Prompt geometry
- Rule mine
- Cleanup screen
- Preregistration
- Court receipt
Each gate exists because easy wins are cheap. Router movement does not equal token movement. A perturbation that appears in one place does not automatically become portable. A same-prompt result does not become authority. Cleanup drift can make a result look meaningful when the system only changed formatting, decoding behavior, or some other surface artifact.
FoxEIS is not trying to make a model say anything on command.
It is trying to separate signal from theater.
That means a useful receipt can still end with no steering authority. A failed run can still become evidence. A model-porting workflow can learn from negative results without pretending the system earned influence.
This public explanation does not include exact coordinates, prompt packs, kernel details, commands, source code, or reproduction paths. Those are engine-room details.
The public claim is narrower and stronger: FoxEIS tests scoped inference interventions, records what happened, rejects fake wins, and keeps authority off by default.