Omniousdocs
For providers◆ Points era

Staying in good standing

Expect canaries and audits. Latency is measured, faults have a published penalty table, and quarantine is recoverable.

The market's trust layer is mostly invisible when you are honest. This page is what it looks like from your side of the counter: how you are measured, what counts as a fault, what a challenge against you means, and how you recover if one lands.

You are being measured all the time

Canary probes are paid requests, indistinguishable from customer traffic, that carry model-fingerprinting prompts with known-good expected outputs. Their primary target is model substitution: bidding one model and serving a smaller or more heavily quantized one. Sampled audits re-run a slice of your completed requests against a reference deployment of the same open-weight model and compare outputs; the sampling rate is adaptive, so new or recently flagged providers see more of it. And latency and quality are measured by the router on your real traffic, never taken from your claims: a provider promising 100 ms and delivering 900 ms sees its effective score decay within minutes, no enforcement required.

The fault table

Faults have published, proportionate consequences. From the design doc's trust section:

FaultEvidencePenalty
Model substitution / precision fraudFailed audits plus canary fingerprintSlash up to 100% of stake, delist
Quote reneging (last look)Signed quote plus rejected routed requestEscalating: warnings, traffic cap, partial slash
SLA miss (chronic)Router-measured latency vs quotedScore decay (organic traffic loss), no slash
Silent truncation / degraded outputsAudit divergencePartial slash plus quality score reset

Note the gradient: honest underperformance loses you traffic through the score, while fraud loses you stake and standing. Slashed funds split 50% to affected customers as pro-rata refunds and 50% to an insurance pool backing dispute payouts.

Bonds, and what a challenge against you means

You post a bond on your model claim: a statement, with value behind it, that the model class you quote is the model you serve. An approved challenger who believes a specific receipt is fraudulent can open a challengeagainst it, locking a stake of their own: max(25 bps, the receipt's charge), with points as the stake asset today (USDC staking is a V2 item). One live challenge per receipt; the stake is the anti-grief rate limit. A verdict resolves it: upheld slashes you at max(250 bps, 25x your net earn on that receipt), half paid to the challenger as bounty and half burned; rejected burns the challenger's stake; ambiguous unwinds everything.

As a providerThe arithmetic is the protection. One upheld substitution costs about 25 honest fills of earnings, so cheating is unprofitable before any human decides anything. Serve what you quote and challenges are someone else's problem.
As a challengerA provider's bond and fault table are your payoff schedule. Catching a real substitution pays a 50% bounty; filing noise burns your stake, so you only file what the evidence supports.

Quarantine, and the way back

An upheld challenge quarantines you: your quotes stop clearing while the finding stands, rather than the market quietly routing customers to a proven-bad deployment. Quarantine is recoverable; delisting is for the worst faults. The path back mirrors new-provider probation: capped traffic share, an elevated audit rate, and graduation on sustained clean performance. Reputation is non-transferable, so the way back is serving honestly under scrutiny, not re-registering under a new key. Details on the bonds and quarantine page.

where this lives in code