Glossary

The vocabulary the other pages use, each defined the way the library implements it. The reference for a term's fields is API; for where it is computed, Architecture.

Term	What it means here
State	The untrusted input being classified: a string, a JSON value, or a chat-message list. It is rendered last in the prompt, so every call about one rubric shares the cached prefix.
Question	One rubric entry — Noul, Choice or Score, with its instructions and criteria. The key it is passed under is the question id it answers under.
Answer	What comes back per question: NoulAnswer, ChoiceAnswer or ScoreAnswer, each carrying the distribution and, where it is defined, confidence.
Probability	One option's share of the distribution: the provider's own numbers on the JSON methods, a softmax of the label logprobs on the label methods.
Distribution	The probabilities for every option of a question, rescaled to sum to 1 when they are off by more than 1e-6. normalize_probabilities=False keeps the provider's numbers.
Confidence	A value derived from the distribution, never asked of the model.
Score	The probability-weighted level index of a Score answer, levels numbered from zero.
Surface	The wire dialect a request is sent in: chat_completions, responses or messages. api picks one; api="auto" reads what the client can do.
Method	How the answer is elicited: logprobs, grammar, structured, discrete, or auto to let the client choose.
Readout	The step that turns one reply into a distribution. It raises LabelReadoutError or MalformedAnswerError rather than guessing, and the client decides whether to spend another call.
Label	The token a model answers with: A–Z up to 26 options, two letters past that. Only the JSON methods may use the two-letter range.
Native reasoning	The provider's own reasoning field, when the plan is native: reasoning_effort on Chat Completions, reasoning on Responses, thinking on Messages.
Two-step	The other reasoning plan: an analysis call first, its trace quoted into the answer call, for a provider that has no native reasoning field.
Correction turn	The retry turn after an answer that could not be read, built from the failure reason alone. n_retry_malformed counts how many are allowed.
Capability ladder	The order in which a field jevper added for capability is given up when a server refuses it: include, the schema carrier, the reasoning parameters, the cache key, thinking.
Server limit	A refusal remembered per (model, surface), so a later call starts without the field. Reported in debug["server_limits"].
Prompt cache key	The prompt_cache_key a provider uses to route requests that can share a cache. jevper derives it from the rubric prefix, never the state; your own key replaces it.
extra_body	Request fields jevper does not own, merged into the call as the provider's client sends them. A key named there is the one that reaches the wire.
Debug record	response.debug: the method and surface actually used, one entry per provider attempt including failures, the retry reasons, and anything normalization had to repair.
Usage	response.usage: provider calls and token counts for the call. n_calls counts successful results, and a token count any response omitted stays None rather than becoming 0.