Jev Noul: the probability that a statement is true

Independent and unofficial. Jev is a model by TypeSafe AI; Jagent is not affiliated with or endorsed by them. The field-by-field reference is TypeSafe's Noul documentation.

What 14 Noul calls showed

All calls went to TypeSafe's API on 2026-11-12, on one theme (a SaaS team reading cancellation risk in support messages). Every one was answered by jev-1.14.0. Cost is input tokens at $0.042 per million; the response also reports output tokens, which are not billed. Milliseconds are the full round trip from our machine through a proxy, so they run slower than TypeSafe's 70–500ms end-to-end. Raw requests and responses are in data/measured-noul-2026-11-12.json.

There is no confidence field

Each answer came back as { "type": "noul", "noul": 0.87 } and nothing else, on the alias and on the pinned version. Code written for a Choice that reads answer.confidence gets undefined here, and a check like confidence >= 0.8 then silently fails every time. The same request sent twice also moved one answer from 0.85 to 0.87, so do not test for exact values.

Note what a short message costs. A 12-word thank-you note came to 346 input tokens, because the question text and request framing are counted too. Budget per question, not per word of input.

A statement and its negation do not sum to 1

TypeSafe lists this in its jaggedness notes, and it held on 1.14. Clear messages sum to about 1. The outage complaint, which never mentions cancelling, gets a firm no to the positive phrasing and a shrug, 0.58, to the negated one. The negation adds a hop of reasoning, and the hop costs certainty. Ask the positive version and compute 1 - p yourself if you need the other side.

A Noul and a yes/no Choice agree on direction, not on size

The last column asks the same thing as a Choice with two options, yes and no. Across five messages (the four above plus a borderline refund question, 0.22 as a Noul and 0.06 as a Choice) the two never pointed opposite ways. But the Choice pushed every answer toward the ends: the outage complaint is 0.18 as a Noul and 0.01 as a Choice, with a Choice confidence of 0.98. TypeSafe documents cases where the two disagree outright. Either way, a threshold tuned on one question type means nothing on the other, and a yes/no Choice hides exactly the middle cases you would want a person to see.

Criteria wording moves the number more than the message does

One instruction, “Is this customer at risk of leaving?”, asked three ways about two messages: the cheaper-plan question and the outage complaint.

The two messages score almost the same under each wording, while the wording alone swings each by 0.5 to 0.6. The criteria are your definition of “yes”, and Jev applies it literally. Without them, it uses a reading of “at risk” you never chose. Write criteria that match the action your code will take: the strict version suits an alert to an account manager, the broad one suits a weekly report.

Screen first, then spend

One Noul on the bare message (“considering cancelling or moving to a cheaper plan”) sorted four messages cleanly: 0.98 and 0.97 for the two churn messages, 0.03 for a how-to question and 0.02 for a thank-you. Only the clear yes went on to the full request, which fetched the account record and thread and asked five questions. It flagged that only 9 of 25 paid seats were in use (0.95) and picked right-sizing at 0.80.

The saving from the gate is smaller than it looks when the next step is also Jev. At about 350 tokens per gate and 876 per full request, gating four messages and running two full requests costs 3,187 tokens. If each full request were the size of ours, running it on all four would cost about 3,504, roughly 10% more. The gate pays off when the step behind it is expensive in some other way: a database lookup, an LLM drafting the reply, or a person's time.

Thresholds: 0.7 / 0.3 or 0.6 / 0.4?

TypeSafe's docs and agent skill suggest yes above 0.7, no below 0.3 and a review band in between. Our Jev MCP server uses a narrower band, 0.6 and 0.4, because an agent asking a quick yes/no question would rather get an answer than “too close to call”. Neither is a rule. The strict-criteria answer for the outage message, 0.36, goes to review under the first pair and counts as a no under the second. Which is right depends on what a missed churn costs compared with a wasted call from an account manager. Label a few hundred of your own messages, then move each cut-off on its own.

How to call a Noul

With a key from this site, on our endpoint (same body; model is optional and defaults to the current Jev):

curl -X POST https://jev-agent.com/api/v1/systemone \
  -H "Authorization: Bearer $JEV_AGENT_KEY" \
  -H "content-type: application/json" \
  -d '{
    "state": "Is there a cheaper plan? Half the team has left.",
    "questions": {
      "churn_risk": {
        "type": "noul",
        "instructions": "The customer is considering cancelling or moving to a cheaper plan."
      }
    }
  }'

With a TypeSafe early-access key, on their host, naming the model:

curl -X POST https://api.typesafe.ai/v1/systemone \
  -H "Authorization: Bearer $TYPESAFE_API_KEY" \
  -H "content-type: application/json" \
  -d '{"model": "jev-1.14.0", "state": "...", "questions": {"churn_risk": {"type": "noul",
       "instructions": "The customer is considering cancelling or moving to a cheaper plan."}}}'

Read answers.churn_risk.noul. The full request and response shapes are in the API reference, and the playground lets you edit the question and rerun it.

When a Noul is the wrong question

• The answer is one of several. Five yes/no questions for five categories can all come back high. Use a Choice, which makes the options compete.
• You want a degree. 0.6 does not mean “somewhat at risk”. It means 60% likely to be at risk. For levels, use a Score.
• Two conditions in one sentence. “Angry and asking for a refund” is two Nouls, combined in code.
• Counting, dates and arithmetic. These are documented weak spots; see Jev limitations.
• Fixed labels at huge volume. A trained classifier can be cheaper per call; the trade-off is on Jev vs a classifier.

Nouls used in practice on this site: five signals per email in phishing detection, checks on an agent's claims in output guardrails, and a diff review and permission gate in Claude Code.

Jev Noul questions

Sources

• TypeSafe AI: Noul
• TypeSafe AI: Confidence
• TypeSafe AI: Screen with a Noul first
• TypeSafe AI: Jev 1.13 jaggedness
• Our 14 calls, 2026-11-12: data/measured-noul-2026-11-12.json