Proof-of-Model

A price is only as good as the model behind it. Every offer is tested before it can serve traffic and again on a jittered schedule after.

Tiers

Tier	Meaning
`none`	Not yet tested. Never routed by default.
`probed`	Reachable and serving completions, but has not passed the full suite.
`verified`	Passed the full suite. The default minimum for routing.
`attested`	Verified, plus hardware attestation checked by the exchange. Coming soon.

The checks

Check	What it proves
`liveness`	A real completion through the exact base URL, credential and upstream model the offer will serve.
`model_identity`	Where the upstream echoes a model id, it must match the listed upstream model.
`known_answer`	Freshly randomised prompts with a single exact answer: multi-step arithmetic and number ordering. Both must pass.
`instruction_following`	Strict output-format tasks (exact JSON shapes) that degraded or substituted models fail.
`usage_honesty`	Compares reported prompt tokens across prompts of known length difference, catching inflated usage.
`context_needle`	A code word placed early in a long prompt must be recalled at the end, catching silent truncation.
`fingerprint`	Greedy outputs compared with the consensus of other offers of the same model to flag substitutions.

Each check is pass, fail, warn or skip (for example, a fingerprint is skipped until other offers of the model exist to compare against). The run produces a score between 0 and 1 and a resulting tier.

When it runs

At listing, synchronously — a failed listing probe means no offer is created.
Continuously, on a jittered schedule, with prompts generated fresh for each run so they can't be special-cased.
On demand, when a seller calls POST /api/offers/{id}/verify.

Paid by the exchange

Verification traffic is paid at the seller's listed price and recorded as verification_credit in their ledger. Proving an offer never costs the seller quota.

When an offer fails

A failed run downgrades the offer's tier; repeated failures suspend it. The seller is notified with the failing checks. Buyers see every offer's tier and score in the order book and in /v1/route/preview.

The report

Verification

{
  "id": "ver_…",
  "offer_id": "off_…",
  "ran_at": "2026-10-22T08:14:03Z",
  "tier": "verified",
  "score": 0.96,
  "passed": true,
  "cost_micro": 312,
  "checks": [
    { "name": "liveness", "status": "pass", "detail": "…", "latency_ms": 420 },
    { "name": "usage_honesty", "status": "pass", "detail": "…" }
  ]
}