Inference Audit

OpenAI-compatible endpoint evidence

What did the endpoint actually deliver?

Controlled requests across pinned providers reveal delivery errors, token-accounting differences, cache reporting, latency, and repeatability. These are observations—not verdicts about cause or intent.

Loading published observations…
Observations — campaigns
Providers pinned, no fallback
Successful responses — request errors
Median latency successful requests

Explore

Provider comparison

Provider n Error Expected Trunc. Tokens/modal Cache hit Latency

Tokens/modal compares reported prompt tokens for the same case and repeat. Cache hit uses provider-reported cached tokens. Neither identifies a cause by itself.

Inspect

Underlying evidence

Read carefully

Evidence, with boundaries

  • Consensus is not truth, and disagreement is not proof of misrepresentation.
  • Reported usage and cache counts are not independently attested.
  • One run describes one model, route, region, configuration, and time window.
  • Provider behavior can change after collection.

Observation

Evidence record

Prompt


      

Response or error