Forward one completion to the model provider under the server-held key

Carries the device header. The answer is priced and the call's counts are added to the member's meter for the UTC day; over the daily cap the answer is 429 with Retry-After in seconds until midnight UTC. Nothing of the call is kept but its counts. With upstream: jev the call goes to Jev: a request over Jev's context limit is a 400 of type urn:six-minute:problem:jev-capacity, and a Jev failure a 502 of type urn:six-minute:problem:provider. With upstream: decisions the call is one closed question to OpenAI's Decisions API: the request carries decision (the property the answer fills and the choices, each with its description) and names its model, and a request without either is a 400. The answer is { [property]: choice } under output, with the probabilities and the confidence under decision. A request the Decisions API refuses for what it holds, and an answer outside the choices, is a 502 of type urn:six-minute:problem:provider-answer, which the caller answers by asking the same question on the openai upstream; with the upstream switched off by the operator the answer is the 503 of type urn:six-minute:problem:provider-unconfigured. sampling: pinned on an openai call asks for the least sampling randomness the model serves (reasoning effort none and temperature 0 on the fixed-temperature models). A provider that fails is a 502 of type urn:six-minute:problem:provider; one that answers with nothing the request can use (a refusal, an answer cut at the output cap, no content, not JSON) is a 502 of type urn:six-minute:problem:provider-answer, which asking again will not change. A purpose only the cloud pipeline asks under (the calendar's: judge_calendar_event, judge_calendar_matter, judge_calendar_attendance; matter learning's: learn_terms) is a 422 and meters nothing.