Proof / Relay

AI output is easy to generate and hard to verify. A response can look correct, sound confident and still skip a requirement, cite evidence that doesn't support it, or take an action it was never permitted to take.

Evaluation has failure modes of its own. Scorers have bugs. Reference answers drift. A test that passes because the evaluator is wrong is worse than no test at all, because it creates confidence that hasn't been earned.

As agents move from answering questions to taking actions, the cost of those failures rises. It needs to be possible to see what an agent did, show that it stayed within its limits, and check the result.

Proof treats evaluation as an engineering discipline. Fixtures define inputs and expected behaviour. Model responses are saved and can be replayed, so scoring is deterministic and repeatable. Claims must be backed by evidence, and catastrophic failures are detected on their own rather than averaged away.

Proof also checks itself. When a result looks wrong, it distinguishes a genuine model failure from an evaluator false positive or a problem in the gold corpus, and the scorers and corpus are verified for correctness in their own right.

Relay sits around AI actions at execution time. Tool use is bounded by explicit policy, every action is recorded with its provenance, and the evidence needed to verify the work later is captured as it happens.