Jev 1.13

Jev 1.13 (typesafe/jev-1.13) is the first TypeSafe System One decision model. It is not an LLM: instead of generating text, it answers typed questions about the state you send — with probabilities your code branches on directly. Context: 32,000 tokens.

What it is not

• Not chat: there is no POST /v1/chat/completions for it — calls go to POST /v1/decisions.
• It does not generate text or explain its decisions — probabilities only. Need a rationale? Ask a chat model afterwards.
• Not a tool for “write me an email” — it is for decision points in code: routing, classification, action gating.

Three primitives

Primitive	Question	Answer
noul	Does this condition hold?	Probability of yes, 0…1
choice	Which one of these options?	The pick + full distribution + confidence
score	Where on the scale?	Weighted position + per-level probabilities

Any number of questions of any type per request — one answer each.

When to pick Jev over an LLM

“Asking an LLM a narrow question and parsing a label out of its answer” is an anti-pattern: slower, costlier, unstable. Jev does the same faster and predictably:

• Routing — which team/queue owns an item.
• Agent gating — is a tool call safe: reversible? within scope? Run / refuse / ask a human by a noul threshold.
• Classification and tagging — one category plus any number of binary tags in a single request.
• Cascades — draft with a cheap model → verify with Jev → escalate to an expensive one only on failed checks.

Quickstart

curl

curl https://api.mixen.ai/v1/decisions \
  -H "Authorization: Bearer $MIXEN_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "typesafe/jev-1.13",
    "state": "My checkout page shows a blank screen after I click Pay. Tried two browsers.",
    "questions": {
      "is_bug": {
        "type": "noul",
        "instructions": "Is the customer reporting a software defect?",
        "criteria": {
          "true": "Describes broken or unexpected product behavior.",
          "false": "Asks a question or requests a feature."
        }
      },
      "urgency": {
        "type": "score",
        "instructions": "How urgent is this ticket?",
        "criteria": ["Can wait", "This week", "Blocking revenue"]
      }
    }
  }'

Python

import requests

r = requests.post(
    "https://api.mixen.ai/v1/decisions",
    headers={"Authorization": f"Bearer {KEY}"},
    json={
        "model": "typesafe/jev-1.13",
        "state": "Task: clean up accounts. Tool call: delete_rows(...)",
        "questions": {
            "safe_to_run": {
                "type": "noul",
                "instructions": "Is this action safe without human approval?",
                "criteria": {
                    "true": "Reversible, low-impact, within the task.",
                    "false": "Destructive, irreversible, or broader than the task."
                },
            }
        },
    },
)
noul = r.json()["answers"]["safe_to_run"]["noul"]
if noul < 0.8:
    escalate_to_human()

One answer per question:

{
  "answers": {
    "is_bug":  { "type": "noul", "noul": 0.96 },
    "urgency": { "type": "score", "score": 2, "probabilities": {"0": 0, "1": 0.01, "2": 0.99} }
  },
  "usage": { "cost": 0.00002, "input_tokens": 476 }
}

Writing good questions

• criteria define the decision boundary, not option descriptions: a true/false pair for noul, one sentence per option for choice, ordered levels for score. The more concrete, the stabler the distribution.
• state is context, not a prompt: ticket text, a state object, an array. instructions may reference nested fields.
• Pick thresholds from probabilities/confidence, not just the top answer: low confidence is the signal to hand off to a human or re-ask. Run a dozen labeled examples and tune the threshold to your cost of error.

Pricing

Only input tokens are billed (state plus questions); output tokens are free. A typical request is a few hundred tokens — a fraction of a cent. Charging follows usage.cost from the response.

Limitations

• Context: 32K tokens.
• The model does not modify state and does not call tools — it only answers your questions.
• Full request/response schema — in the Decisions guide.