A8 (Touchstone) is an evaluation model with a chat-model interface: the same request shape, the same SDKs, one changed base URL. The verdict is computed against a trained model rather than generated by a language model, so the same input returns the same verdict.
A8 is a DLM — a discriminative language model. It is encoder-only: it scores text against a criterion and cannot generate text.
A request carries two things. The system message is the criterion — what is being judged. The last user message is the subject — the thing under judgment. An optional json_schema names the fields the verdict should come back in.
Every field, the response shape, and the status codes are on the reference.
Set min_accuracy to the accuracy required — an integer 0–99, default 90. It is a promise about the answers the model does give: 90 means at least 90% of answered verdicts are correct, with the cases the evidence cannot carry returned as abstentions.
It is a floor the request sets, not a belief the model reports. Nothing in the response says a particular answer is 90% likely to be right; the number is the accuracy the evidence had to demonstrate before the model returned a verdict.
What a criterion has earned is a smaller, uneven subset of the range, so an ask can be served at a higher level than requested and the response reports the level it was served under. Declining is the mechanism the promise rests on. See abstention for the four reasons a criterion declines and what clears each.
No language model runs on the response path. A verdict is a function of the text and the model that read it, so the same request against the same snapshot returns a byte-identical response. In a regression suite, a changed verdict means the input changed or the model moved.
system_fingerprint in the response tells you something moved.
min_accuracy promises, and when the model declines.expected.