API

JevAlt speaks TypeSafe's POST /v1/systemone. A request that works with Jev works here without changes. A few optional fields add what Jev does not have.

Request

{
  "state": "Hi, we were billed twice for March. Please refund the duplicate today.",
  "model": "jevalt",
  "questions": {
    "team": {
      "type": "choice",
      "instructions": "Which team should handle this?",
      "criteria": {"billing": "Invoices, payments, refunds", "technical": "Bugs and outages"}
    },
    "urgency": {"type": "score", "instructions": "How urgent is it?", "criteria": ["Can wait", "Today", "Now"]},
    "refund": {"type": "noul", "instructions": "Does the customer ask for money back?"}
  }
}

Field	Type	Notes
state	string, object or array	The evidence. Only put what the questions need in it.
questions	map of id to question	Answers come back under the same ids.
model	string	Accepted and ignored by a local server; the response names the model that answered.

Question types:

Type	criteria	Answer
choice	map of option to description (up to 62 options)	choice, probabilities, confidence
score	ordered list of 2 to 10 level descriptions	score (probability-weighted level), legend, probabilities, confidence
noul	optional {"true": ..., "false": ...}	noul, the probability of yes

confidence is (n * p_max - 1) / (n - 1): 1 when all mass sits on one option, 0 when the distribution is flat. It matches the worked examples in TypeSafe's documentation.

When instructions and criteria disagree

Write the criteria so they agree with the instructions. If a Noul asks "Was it on time?" but its criteria say yes means "arrived after the deadline", the intended answer follows the criteria, and that is where these models still slip. TypeSafe lists the same case as a known weakness of Jev 1.13. JevAlt was trained on such pairs, yet on the held-out inverted-criteria rows the released models answer by the criteria only 43 to 71% of the time, close to the start checkpoint's 40 to 71% (14 to 41 decisions per language, results/ in jevalt-bench).

Extensions

All optional. Leave them out and the response has exactly Jev's shape.

Field	Values	What it adds
reasoning	"off" (default), "on", "auto"	on: the model writes a short reasoning trace, then decides. auto: it decides first and thinks again only for answers below a confidence threshold. Answers that used a trace carry it in reasoning.
abstain	true	Adds an unknown option to every Choice and Score. Its probability comes back as unknown.
coverage	0.8, 0.9, 0.95	Adds set: the smallest group of options that contains the right answer with at least this probability, from conformal thresholds fitted on held-out data. If the requested level has no fitted threshold, the smallest fitted level at or above it is used. The level used is returned as coverage in the answer. If no fitted level is at or above the request, no set is returned and coverage is null.
language	"en", "tr", "de"	Language for the unknown wording and the reasoning trace. Detected from the state when missing.

Response

{
  "model": "Deem-4B",
  "answers": {
    "team": {"type": "choice", "choice": "billing", "probabilities": {"billing": 0.93, "technical": 0.07}, "confidence": 0.86},
    "urgency": {"type": "score", "score": 1.4, "legend": {"0": "Can wait", "1": "Today", "2": "Now"}, "probabilities": {"0": 0.05, "1": 0.5, "2": 0.45}, "confidence": 0.25},
    "refund": {"type": "noul", "noul": 0.97}
  },
  "usage": {"input_tokens": 212, "output_tokens": 3}
}

Errors

Status	Meaning
401	The server was started with JEVALT_API_KEY and the request did not send it as a bearer token.
422	The body is not valid: missing state, empty questions, an unknown type, or too many options. The message names the problem.