Jev Cookbook

Practical, tested recipes for TypeSafe's Jev, the fast decision model on OpenRouter. Fifteen real-world jobs, each with a runnable script, a small labelled dataset and measured results: support triage, database indexing, a file organizer, tagging, category trees, duplicate detection, PII scanning, bank transactions, invoice extraction, search, log triage, moderation, lead scoring, a browser agent and a Gmail labeler that connects to your own inbox.

Created by Jeroen Erne (nexibeo.com · completeaitraining.com), built together with Claude.

What Jev is, in 30 seconds

Jev is not a chat model. You send it some data (the state) and a set of questions with fixed answers, and it returns typed answers with probabilities, usually in under half a second.

• Choice: pick one option from a list you define (up to 255), with a probability for each.
• Noul: the probability that a yes/no statement is true.
• Score: a position on levels you describe, lowest first.

It never writes text, so it can't hallucinate an answer outside your options, and it is cheap: $0.042 per million input tokens, output free. Across these recipes that's 2 to 8 cents per 1,000 items. It can still pick the wrong option, which is why every recipe routes on confidence and measures its errors.

// One call, three questions, typed answers.
const res = await fetch('https://openrouter.ai/api/alpha/decisions', {
  method: 'POST',
  headers: { Authorization: `Bearer ${process.env.OPENROUTER_API_KEY}`, 'Content-Type': 'application/json' },
  body: JSON.stringify({
    model: '~typesafe/jev-latest',
    state: { message: 'I was charged twice for March. Please refund the duplicate.' },
    questions: {
      team: { type: 'choice', instructions: 'Which team should handle `message`?',
              criteria: { billing: 'Payments, invoices, refunds', technical: 'Bugs and outages', sales: 'Pricing and plans' } },
      refund: { type: 'noul', instructions: 'The customer in `message` asks for money back' },
    },
  }),
});
const { answers } = await res.json();
// answers.team   -> { choice: 'billing', probabilities: { billing: 1, ... }, confidence: 1 }
// answers.refund -> { noul: 0.99 }

Two things trip people up. Jev lives at /api/alpha/decisions, not the chat endpoint, which rejects it. And the model id ~typesafe/jev-latest needs the tilde. It always points to the newest Jev.

Quick start

Needs Node 22.9 or newer and an OpenRouter key.

git clone https://github.com/nexibeo/jev-cookbook.git
cd jev-cookbook
cp .env.example .env        # paste your OpenRouter key into .env
npm run quickstart          # one call, all three question types
npm run 01                  # any recipe by number
npm run all                 # every recipe except the browser agent (about 2 cents)
npm run gmail               # label your own Gmail (dry run first; setup in recipes/15-gmail-labeler)

Recipes 01 to 13 and 15 have no dependencies. Recipe 14 (the browser agent) needs npm install and a Chromium: npx playwright install chromium, or set CHROME_PATH. Recipe 15 needs your own free Google OAuth client to reach your Gmail; its sample inbox works without one.

The recipes

Results are from live runs on September 19, 2026 against ~typesafe/jev-latest (answering as typesafe/jev-1.13-20260917). Every run saves its full output in results/.

#	Recipe	The real-world job	Jev questions per item	Result on the sample data
01	Support triage	Route tickets, flag urgency, frustration and refund requests	Choice + 2 Nouls + Score	Right team 34/34; genuinely ambiguous tickets dropped to 0.60–0.67 confidence and went to a person. Urgency 94%, refund requests 100%
02	Database indexing	Turn free-text rows into typed, indexed SQLite columns you can query	2 Choices + 8 Nouls	Category 35/36; 41 tags stored; re-runs only index new rows
03	File organizer	Sort a messy folder into category folders with dated file names	Choice + Noul + date Choice	Category 24/24, document date 23/24
04	Multi-label tagging	Tag articles from a fixed vocabulary	13 Nouls	F1 87% at threshold 0.5; precision 97% at 0.9
05	Category trees	Place products in an 8 × 6 taxonomy, flat or level by level	1 Choice, or 2	36/36 both ways; level by level used 11% fewer tokens
06	Duplicate detection	Decide whether two company records are the same	Score	0 wrong automatic merges; 2 hard pairs sent to review. Without code-checked facts: 1 wrong
07	PII column scanner	Label database columns for a data catalog: type, personal data, sensitivity	Choice + Noul + Score	Type 25/25, sensitivity 25/25, personal data 23/25; all 5 restricted columns caught
08	Bank transactions	Budget categories from cryptic bank descriptors	Choice + Noul	35/36; recurring bills found
09	Invoice extraction	Invoice number, dates and total, in several languages	Noul + 4 Choices over regex candidates	40/40 fields; a quote correctly returned no invoice fields
10	Search re-ranking	Find the help article that answers a question	Choice	Keyword search 2/16 → with Jev 12/16 → Jev over all 30 articles 16/16; unanswerable questions recognised 2/2
11	Log triage	Decide what pages on-call, what becomes a ticket, what's noise	Choice + Score + 5 Nouls	Component 96%. Paging on one severity score caught 4/7 critical events; splitting it into 4 yes/no questions caught 7/7
12	Moderation and guardrails	Screen posts for spam, harassment, prompt injection, doxxing, self-harm	5 Nouls	0 harmful posts published, 0 clean posts blocked; every hazard caught
13	Lead scoring	Rank inbound sales leads, explainably	4 Scores + Noul	The top 5 by score are exactly the 5 priority-A leads
14	Browser agent	Navigate real websites, search and fill forms	Choice + 2 Nouls per step	5/6 live tasks passed, $0.0005–0.003 each; the failure is a documented near-miss
15	Gmail labeler	Connect your Gmail and label every email by type, "needs reply" and "deadline"	Choice + 2 Nouls, plus code-checked header facts	Sample inbox: category 30/30, needs reply and deadline 93%, phishing 3/3 with no false alarms. Only ever adds labels

More every week: three new recipes land every Monday until mid-October 2026. See the roadmap.

Also here: examples/quickstart.mjs, a shared client in lib/jev.mjs, regex candidate finders in lib/candidates.mjs, the full guide, and a case study that files 3,267 real templates.

How every recipe works

The pattern is the same everywhere: your code prepares the data and owns every decision; Jev answers narrow questions.

1. Build the state: the few fields a person would need to judge the item, as JSON. Trim everything else.
2. Ask everything in one call: a Choice for one-answer decisions, one Noul per label for multi-label ones, a Score for anything ordered. Questions run in parallel, so ten more cost almost nothing.
3. Decide in code: thresholds, weights and actions live in your code, so you can tune them without touching the questions.
4. Route on confidence: act when Jev is sure, send the middle band to a person, and log which model version answered.

import { ask, choice, noul, route } from './lib/jev.mjs';

const a = await ask({ message: ticket.text }, {
  team: choice('Which team should handle `message`?', { billing: '...', technical: '...', account: '...' }),
  urgent: noul('The customer in `message` needs a response within hours'),
});
if (route(a.team.confidence) === 'auto') queue(a.team.choice); else askAPerson(ticket);

What we learned

• Describe every option, and say what doesn't count. Jev reads literally. On a real catalog of AI bots, a plain topic question tagged 36 of 100 templates "Generative AI" just for being AI; one sentence saying that doesn't count brought it to 6 (case study).
• Numbers, dates and exact matching are code's job. Let code parse amounts, compare phone numbers or read the sign of a transaction, then hand Jev the result as a fact (06, 08).
• Pick, don't extract. Jev can't copy text out of a document, but it's excellent at choosing among regex-found candidates, and a none option stops it inventing values (03, 09).
• Split "how bad" into specific yes/no questions. A single severity Score bunches in the middle; four concrete Nouls paged every critical event (11). The same idea ranks sales leads explainably (13).
• Confidence tells you where the doubt is. Clear tickets scored 1.00; tickets that fit two teams scored 0.60–0.67 (01). Route the middle band to people.
• Leave a review band around every cut-off. Scores near a threshold can shift a little between runs (a duplicate pair scored 0.49 once and 0.52 the next time).
• Small collections don't need retrieval. Up to 255 items fit in one Choice, and reading all of them beat keyword search plus re-ranking (10).
• Thresholds are a dial. The same answers give precision 80% / recall 94% at 0.3, or precision 97% / recall 64% at 0.9 (04).
• Different mistakes cost different amounts. Moderation blocks only when sure, holds the middle for a moderator, and escalates self-harm early (12).
• A confident answer can still be a near-miss. The browser agent opened "Cold Outreach" when asked for "Cold Email", and both of its checks agreed. Verify outcomes in code (14).

Cost and speed

Price	$0.042 per million input tokens, output free
Typical call	500 to 2,000 input tokens: $0.00002 to $0.00008
Median latency per call	0.34 to 0.45 s in these runs
All of recipes 01–13	425 calls, $0.015
Browser tasks	$0.0005 to $0.003 per task, including the typing model

A note on the numbers

The datasets are small and hand-made (16 to 36 items each), written to look like real data with some deliberately ambiguous cases. The scores show how each technique behaves; they are not benchmarks. Before relying on a recipe, label a few hundred of your own items and measure. Each script prints its errors, so that takes minutes.

Project layout

lib/jev.mjs            the client: ask(), question builders, pickLabels(), route(), mapLimit(), metrics
lib/candidates.mjs     regex finders for dates, amounts and reference numbers
examples/quickstart.mjs
recipes/NN-name/       one folder per recipe: README, script, sample data
results/               the saved output of every run
docs/GUIDE.md          the full guide: connecting, question design, confidence, weak spots, patterns, browser agents
docs/case-study-templatesgrokbot.md

More reading

• docs/GUIDE.md: everything we learned about instructing Jev, in one place
• TypeSafe docs and Jev 1.13 known weak spots
• OpenRouter model page and Decisions API reference
• Browser Use's Jev Ultrafast, the reference browser agent

Credits

Created by Jeroen Erne, of Nexibeo and Complete AI Training, built together with Claude. The case study comes from TemplatesGrokBot. Jev is made by TypeSafe; this project is independent of TypeSafe and OpenRouter.

MIT licensed: see LICENSE. Contributions and new recipes are welcome.