Jevper

No, and the mechanism matters. method auto reads real logprobs where the provider exposes them, so you get the model's distribution, but where it cannot, it asks for JSON and the numbers are the model's self-report rescaled to sum 1. A reasoning model such as gpt-5.6-terra returns no logprobs at all, which is exactly that fallback case.

An independent implementation of the documented wire format, not affiliated with or supported by TypeSafe. It imports no provider SDK at all: any object exposing responses.create, chat.completions.create or messages.create works as the client. The same system_one call also reaches the hosted Jev endpoint and self-hosted llama.cpp servers through api="systemone".

It degrades instead. A server that refuses structured output, the reasoning parameters, the Responses include list, the cache key or the Messages output_config gets that field dropped and the call re-asked, so a partially implemented server answers rather than failing, and debug["server_limits"] reports what it refused.

Up to 255 options, which is the Jev API limit, but not with every method. logprobs and grammar read a label token and stop at 26, since the first token of "AA" is "A"; past 26 they raise InvalidQuestionError pointing at structured and discrete. The default method="auto" never hits that error, answering a wide Choice in JSON instead.

jevper is an independent implementation of OpenAI's documented Decisions API schema and the documented System One wire format. OpenAI serves that schema on POST /v1/decisions in public beta for gpt-6-luna alone, and some gateways expose it, but local llama.cpp, LM Studio, ollama, vLLM or SGLang servers and Anthropic-compatible endpoints do not. jevper puts the same questions to those over Chat Completions, Responses or the Messages API.

Four, all sharing the same label-to-option mapping: logprobs, grammar, structured and discrete, with auto the default. The two methods that read a label token, logprobs and grammar, stop at 26 options because the first token of a two-letter label is its first letter, and past that they raise InvalidQuestionError pointing at structured and discrete, which answer in JSON. The default auto never hits that error, answering a wide Choice in JSON.

Each question has its own provider call sequence and questions are independent, running concurrently at a default max_concurrency of 8. Answers come back keyed by your question ids in insertion order. The Decisions wrapper uses the same flow but hands back the wire's shape instead: an answers list in ask order and one usage summed over the calls, with the raw response kept on decision.raw and left out of every dump.