Canaries & sampled audits

Shadow operator canary tooling, separated from Proposed sampled re-runs and a Proposed reference deployment.

Owner

Trust & Safety

Version

integrity-2026.07

Verified

2026-07-13

Canary requests and sampled re-runs are two different verification ideas, at two different stages. The repository contains operator canary tooling and evidence records. A continuously scheduled audit network against a reference deployment is not part of the opening Beta commitment.

Canary probes: Shadow operator tooling

A canary is a paid request carrying a fingerprinting prompt. The goal is to compare what actually answered with the model class a provider declared, while sending the request through the ordinary serving path. The probe runner and result records exist in the repository, but cadence, coverage, thresholds, and customer guarantees are not published as a live service level.

Canary output is evidence, not an automatic punishment. An operator can inspect a mismatch alongside the signed quote, request record, and measured service history. Quarantine and challenge records are also Shadow controls.

Sampled audits: Proposed

The broader design would re-run a sample of requests against a known reference deployment and compare outputs statistically. That requires an approved reference, sampling policy, privacy treatment, cost budget, thresholds, and appeal path. None is a customer term today. The proposed first-party reference provider is described at hyper; it has no committed launch, hardware, capacity, model, or price.

How evidence may feed decisions

Evidence can inform the verification-trust input used by the auction and can support an approved-wallet Shadow review. It does not create a public bounty, automatic verdict, escrowed provider bond, customer refund pool, or insurance promise.