Skip to main content

Model verification for providers

Auditors regularly check that your services serve the models they advertise. You don't need to sign up or install anything. This page covers what gets audited, what a result does to you, and how to fix a DIFF. For the basics, see the overview.

What gets audited​

Every epoch, each model on the audit list is tested on every seller that offers it. antseed verifier models prints the list. An audit is about 100 short questions, sent as normal paid requests, so you get paid for them like any other traffic.

A service is not audited when it costs more than the model does upstream, meaning more than the most any OpenRouter provider charges for it. Such a service gets no result, so it gets no bonus either.

What a result does to you​

SAME raises your ANTS points by 2% for each verified service, up to +50%.

DIFF does two things:

  • That service keeps only 10% of its reputation, so buyers' automatic routing stops sending requests to it. Your other services keep their reputation.
  • Your ANTS points go down on every settlement, whatever model it was for, because rewards count all your services together. One DIFF service costs 4%, two cost 16%, three 36%, four 64%. From five on you earn no points at all.

The bonus and the penalty add up, so 6 SAME services and 1 DIFF give you 108% of your usual points. Your buyers' points and your pool's staker rewards move by the same amount, so a DIFF also costs the buyers and stakers who chose you.

UNDETERMINED changes nothing. It usually means your service returned errors, timed out or sent empty answers during the audit.

A DIFF never affects your USDC: you are still paid for what you served. These are the current Base mainnet settings and the contract owner can change them.

Check your results​

Antscan shows your SAME and DIFF services and links each result to the audit behind it. The audit evidence on IPFS has every question, your answers and the statistics, so you can see exactly which answers didn't match.

Check yourself first​

You can run the same check against your own endpoint at any time, for example before you switch upstreams or change a model mapping:

Terminal
antseed verifier check <model> --base-url <your endpoint> --api-key-env VERIFY_API_KEY

It prints SAME, DIFF or UNDETERMINED. You pay for the questions, about 100 short requests. See the CLI reference for the options.

If you get a DIFF​

A DIFF means two auditors, each with its own questions, found that your service answers differently from the genuine model. Start with the upstream:

  • Is the upstream serving the exact model you advertise, and not a smaller, distilled or heavily quantized version?
  • Does your model ID map to the right upstream model?
  • Does the upstream or an aggregator fall back to another model under load or on errors?

Once it's fixed, run antseed verifier check against your endpoint to confirm. The next audit clears the DIFF: a single auditor's SAME is enough. If you can't fix it, stop offering the model under that name. Either way, a result only counts for the epoch it was given in and the next one, so an old DIFF goes away on its own once newer audits stop finding it.

If you think the result is wrong, the evidence is public. Rerun the same audit with antseed verifier check --audit <ipfs://cid> and raise it with the team, together with your output.

Last updated on