Canaries & sampled audits

Paid probe traffic indistinguishable from customer requests, plus re-runs against a reference deployment.

The always-on layer of the integrity stack is traffic the provider cannot distinguish from revenue. That indistinguishability is the whole trick: a provider that would cheat only when unobserved has no unobserved moments.

Canary probes

Canaries are paid requests carrying fingerprinting prompts: inputs chosen so different models (and different quantizations of the same model) produce measurably different outputs. They ride the normal auction, pay the normal way, and arrive from wallets that look like customers. If the fingerprint that comes back does not match the model class the provider is claiming, that is direct evidence of substitution, and it lands on the provider's trust record and integrity receipts.

Sampled audits

Audits re-run a sample of real requests against a reference deployment of the claimed model and compare the two outputs statistically: log-probability divergence and embedding similarity, not eyeballing. The audit rate is adaptive; providers with clean histories are sampled less, new or flagged providers more. The design targets roughly 0.5 to 2 percent of traffic. The reference deployment the audits compare against is one of the jobs of the first-party provider; see hyper.

What they feed

Neither canaries nor audits punish anyone directly. They produce evidence: trust events and tamper-evident integrity receipts. The consequences flow through the mechanisms that price them, the quality term in the auction score and the staked challenge lane, where a receipt caught by a canary is exactly the kind a challenger can profitably contest.