Checks
A listed key or server is checked all the time, on your key and at your cost: that a model answers, that it is the model it says it is, and that it reports its tokens honestly.
The checks on your key
- Health testOne short request: does the key answer, and does the model answer under its own name. A model that fails it, or the identity probes at listing, is not added: it is removed, not kept as a draft; add it again to retry. at listing, per model, all at once; about 16 tokens a model.
- Liveness pingA tiny request: does the key answer, and how fast. every 5 minutes; about 300 tokens a day.
- Random checksBuyer-like requests, drawn from a large pool and sent like any buyer's, whose answers verify the model. at random times, about one an hour per offer (about 24 a day); each up to 512 output tokens plus its prompt.
- Identity probesFixed questions whose answers tell whether the model is the one the offer names. at listing, then every 6 hours; about 2,000 tokens a day.
- Funding checkIs the key on a paid tier or funded. A key that is not is paused as No balance until it is. at discover, then with the probes; not on your own server; at most one token.
A request that fails is not billed to the buyer, and you are not paid for it. A failure before the first token moves the buyer's request to the next source at once, so the buyer never waits on a broken key.
When a check fails
We test your models at random with real prompts, about once an hour per model. A model that fails is checked again at once, with another prompt. If it fails again, a full identity check runs on the model and the failure is recorded. It is not unlisted: it keeps serving, and keeps being checked.
A request that fails is not billed to the buyer, and you are not paid for it. A failed check does not touch your earnings.
A check your server cannot answer (an error, a rate limit) is not a failure: it is tried again a few minutes later. Errors on buyers' requests are handled apart, below: the model gets no traffic for a short pause that grows with each failure, until a request or our 5-minute ping succeeds.
Paused and unhealthy offers
Paused · No balance: your provider account has no funds, or the key fell back to a free tier. Fund it; the next check routes it again.
Unhealthy · At capacity: the provider's own quota window is used up (a GLM Coding Plan's 5 hours, a rate limit). It comes back on its own when the window resets.
Unhealthy · errors on the key: requests failed or timed out. The model gets no traffic for a pause that starts at 30 seconds and doubles up to 10 minutes, then one request or the 5-minute ping tries it again; Reset health on Positions runs a fresh test now.