-
An answer from the evaluator, typed to match the question that produced it.
A rating against the question's ordered levels.
The answers to one fan-out of questions, keyed by the names the caller gave them, together
with what the evaluation cost.
An error the API itself reported, carrying the HTTP status and what the API said, so a caller
can tell a rejected token from a rejected question without reading prose.
Everything the client needs that is not a question: where the evaluator lives, which model
answers, how long to wait, and how hard to retry.
Every failure this library raises, unchecked, carrying the operation that failed and the
kind of failure so a caller can branch on the kind without parsing a message.
The single seam between this library's records and the evaluator's HTTP surface: paths, JSON
field names and the headers that carry meaning.
One entry of the evaluator's model list, as GET /v1/models reports it.
A question put to the evaluator about a state.
How hard, and how politely, to retry.
What the evaluator reports it spent answering.