Verify · the transactional product
Is that endpoint the model it claims to be?
An identity audit statistically compares a suspect endpoint against the claimed reference model, paired prompt sets, distribution tests, refusal patterns, formatting fingerprints, latency and token profiles, and returns one of three verdicts in 72 hours. Pricing is by request during the pre-series pilot.
The verdict format
frozen wording, behavior statistics, never accusationsOne endpoint, one claim
- Endpoint vs one claimed reference model
- Report in 72h: PDF + JSON + private permalink
- n disclosed per test; false-positive characterization referenced
We scope and quote by email
Longitudinal recheck
- Full battery + canaries + sampling-parameter behavior
- Rechecks at day 14 and day 30, catches time-of-day and load-based substitution
- Continuous monitoring upsells into the Feed as a watched custom endpoint
72h first report, then scheduled rechecks
The rails
Authorization first: you attest that you are authorized to test the endpoint, we never probe endpoints without it. Statistics, not accusations: reports state behavioral (in)consistency with a claimed model; what you do with that is your decision. Held-out prompts: audit prompt sets are never disclosed, so a gateway cannot special-case them. Buyers: procurement vetting a vendor, diligence teams, journalists verifying a story, and anyone whose gateway bill looks too good to be true.
Want this continuously instead of once? Watched custom endpoints live in the Feed.
Drift & identity feed →