Stop coaxing a chat model into "answer only yes or no". POST /v1/decide asks named questions about a
piece of state and returns, per question, a value and the probability behind it — in well under a second, for a fraction
of a cent.
A chat model asked for "a probability" produces a plausible-looking number. A decision model produces a calibrated one. That's the whole point, and why Fire refuses to fake this endpoint with a prompt.
Optional instructions, optional criteria. Answer: {"noul": 0.98}, the probability of "yes". Pick your
own threshold per use case instead of trusting a string.
Criteria is a {label: description} map, 1–255 labels. Answer: the chosen label, a confidence, and the
full probability distribution across labels.
Criteria is an ordered list of 2–10 level descriptions. Answer: a probability-weighted score, per-level probabilities, and the legend echoed back so the rubric stays visible.
state is a string or a JSON object/array of text — a ticket, an email, a form. Text only; no images or audio.
~32K tokens for state plus the longest question.
Fire checks every question shape before the provider is called, so a malformed request is a free 422, not a billed mistake. Unknown keys are dropped, never forwarded.
Jev is $0.042 per million input tokens at the provider; output is free. A three-question call on a paragraph is around
two hundred-thousandths of a dollar. meta.price.usd says exactly.
Three questions, one pass: what are they asking for, are they angry, how urgent. Each answer carries its probability, so a router can send p(cancellation) > 0.3 to a human and the rest to a bot.
Where this earns its keep: routing before you spend a frontier chat call, moderation gates with a tunable threshold, lead and deal scoring against a rubric an operator can read, validating another model's output.
A species with modality "decision" must produce calibrated probabilities. Naming a chat model here is a 404, not a fallback.
You can, and for free-text output you should use /v1/chat. But a chat model's "probability" is a plausible number, not a
calibrated one, and its label choice drifts with phrasing. A decision model is built to return a distribution you can threshold.
Jev by TypeSafe (jev) is the default and currently the only onboarded decision model, served through Eden AI. Fire's request
shape is its own contract — if the upstream endpoint changes, the adapter absorbs it and /v1/decide doesn't move.
Up to 32 questions per call; names 1–64 characters of letters, digits, _ . -; choice criteria 1–255 labels; score criteria 2–10 levels;
~64K tokens of context (32K for state plus the longest question). Text and JSON only.
Yes — POST /v1/calls/{call_id}/report accepts wrong_answer, miscalibrated,
ignored_criteria, missing_answer or other for decision calls.
Free to sign up — you only pay when you top up. No card required to get your first key.