ModelRefs / Evidence API
Evidence API
A read API for governed AI model evidence: citations and benchmark runs with their methodology limits attached. Coverage is public and free to query.
What is actually in it
ModelRefs holds governed evidence for 52 of 477 catalogued models, which is 11 per cent. That evidence comprises 382 citations and 123 governed benchmark runs. These counts are computed at build time from the same registries the endpoint serves, so the number on this page is the number the API returns.
A model with no governed evidence answers 404 rather than an empty result. Absence is a statement about our coverage and not a judgement about the model, and 425 catalogued models are in that position today. The coverage endpoint is free and needs no key, so which models are held can be checked before a request is spent.
| Measure | Count |
|---|---|
| Models with governed evidence | 52 |
| Citations | 382 |
| Governed benchmark runs | 123 |
| Models in catalogue | 477 |
Two endpoints
GET /v1/coverage
No key required. Every model slug ModelRefs holds governed evidence for, plus the size of the catalogue. Call it before iterating a model list so requests are spent only on models that have evidence.
GET /v1/models/{slug}/evidence
Requires a key. A model's citations and governed benchmark runs. Each run carries its source, publisher, claim scope and methodology limitations, so a score never travels without the conditions under which it holds.
curl https://vmwgrcnqpavfrwrxsinx.supabase.co/functions/v1/evidence-api/v1/coverageAllowances
Allowances are per account and reset on the first of each month, UTC. Every response carries the remaining allowance in its headers, so a caller never has to guess where it stands. The free tier is 1,000 calls a month.
| Plan | Calls per month |
|---|---|
| Free | 1,000 API calls / month |
| Starter | 10,000 API calls / month |
| Pro | 50,000 API calls / month |
| Business | 500,000 API calls / month |
What this API deliberately will not do
It does not run inference. There are no completion, embedding or image endpoints. This is a reference layer, not a model gateway.
It does not return a score without its scope. A benchmark result describes a named model version on a named benchmark version under a stated claim scope, and does not transfer to other versions or configurations.
It does not pretend to cover the catalogue. A model with no governed evidence answers 404 rather than an empty payload.
Getting started
The full reference, covering authentication, the response shape, every status code and the rate-limit headers, is on the API reference page. Keys are created on the API keys page and shown once, because only a hash is stored.