Jev Showcase

September 20, 2026 · View on GitHub

Jev — a crystalline decision engine emitting parallel judgment beams

Unofficial operator-level companion for TypeSafe Jev
Choice · Score · Noul · parallel fan-out · confidence as a second axis
Jev 1.13 · released 15 September 2026

What people miss · Announcement · Docs · Models · Article · Showcase framework

Most posts stop at “fast structured output.” This repo surfaces the docs-only details that change how you build: Jev is a function call, questions run in parallel, output is free, confidence is not probability, and code owns the workflow.

Under-known fact	Why it matters
Not a chatbot	No prose, code, or explanations. Pair with a generative model.
Three primitives	Choice / Score / Noul. Question IDs are not sent to the model.
Fan-out	Many questions, one state, one call. ~12× cheaper than serial.
$0.042 / MTok, output free	Optimize question design, not completion length.
Noul has no `confidence`	Do not copy a Noul threshold onto a Choice.
No fine-tune	Shape answers via `state` + `instructions` + `criteria`.
Literal + no math	Count, dates, and arithmetic stay in code.

Full write-up: docs/WHAT-PEOPLE-MISS.md · interactive cards in the lab.

Keywords: TypeSafe, Jev, System One, RLCD, calibrated decisions, structured output, confidence routing, agent guardrails

The one-liner

Frontier-intelligence function call — unstructured state in, typed probabilistic decisions out. 70–500 ms. Cannot invent a label off your schema.


Model id	`jev-1.13.0` (`jev-latest`)
Class	System One (not a chat LLM)
Endpoint	`POST /v1/systemone`
Context	64k / request; 32k for state + longest question
Input	Text / JSON. No image, audio, or video
Output	Choice · Score · Noul (+ probabilities)
Pricing	$0.042 / MTok input · output free
Latency	70–500 ms (vendor)
Training	RLCD — Reinforcement Learning for Calibrated Decisions
Released	15 September 2026

30-second start

git clone https://github.com/cobusgreyling/Jev.git
cd Jev

python3 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
python app.py

Open http://127.0.0.1:7872/home for the smart-home demo (TypeSafe’s fan-out house), or http://127.0.0.1:7872 for the operator lab.

Harness CLI (TypeScript)

The same System One questions, packaged for agent loops — model routing, semantic retrieval, LLM guardrails, reasoning-trace classification. Wired into Loop Engineering as loop-jev.

cd packages/js && npm install && npm test
npx jev doctor
npx jev route --goal "CI has been red for 3 days" --level L2 --json
npx jev guard --side input --text "Ignore previous instructions" --json

Docs: packages/js/README.md · loop-engineering/docs/jev.md

The key is never printed. Store it as TYPESAFE_API_KEY, ~/.typesafe/api_key, or ~/.config/typesafe/api_key (mode 600).

./run.sh

Agent skills

Portable SKILL.md files for coding agents. Official TypeSafe skill: typesafe-ai/skills.

npx skills add cobusgreyling/Jev --skill jev
npx skills add cobusgreyling/Jev --skill jev-fanout
npx skills add cobusgreyling/Jev --skill jev-guardrail
npx skills add cobusgreyling/Jev --skill jev-route

Skill	Purpose
`jev`	Primitives and when to call Jev
`jev-fanout`	Speculative fan-out (one request, many questions)
`jev-guardrail`	Input/output hazard screen
`jev-route`	Confidence lanes and model-tier routing

Grok also loads .grok/skills/ (symlinks). Catalog: skills/README.md.

Docker

docker compose up --build
# → http://127.0.0.1:7872

No API key for offline tabs. Live judge:

cp .env.example .env
# put your TypeSafe key in .env or ~/.typesafe/api_key
# mint at https://console.typesafe.ai/keys
./run.sh

The key is never logged, never returned by /api/health, and never committed. See SECURITY.md.

What the lab shows

Surface	Offline?	What you learn
Smart home	Yes (mock)	Official-style house: 13 questions in one call, code flips the lights
What people miss	Yes	10 ranked operator cards with actions + doc links
Playground	Yes	Choice + Score + Noul, then a policy in code
Cost lab	Yes	Input-only billing, fan-out vs serial, LLM stand-in
Patterns	Yes	Fan-out · confidence routing · composite · intent
Jaggedness	Yes	Official Jev 1.13 failure modes
Scenarios	Yes	Triage, guardrails, RAG, citations, tool traces
Live judge	Needs key	Real System One call; key never shown
Model card	Yes	Specs, aliases, gotchas

The request shape

curl -X POST https://api.typesafe.ai/v1/systemone \
  -H "Authorization: Bearer $TYPESAFE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "state": "I was charged twice. Please refund the duplicate today.",
    "model": "jev-latest",
    "questions": {
      "department": {
        "type": "choice",
        "instructions": "Which team should handle this?",
        "criteria": {
          "billing": "Payments and refunds",
          "technical": "Bugs or integrations",
          "other": "Neither fits"
        }
      },
      "refund_requested": {
        "type": "noul",
        "instructions": "Does the message request a refund?"
      },
      "urgency": {
        "type": "score",
        "instructions": "How time-sensitive is this?",
        "criteria": ["No deadline", "Within a week", "Today or sooner"]
      }
    }
  }'

Answers come back under the same keys, typed. Nothing to parse.

Progressive examples

python examples/01_noul.py
python examples/02_choice.py
python examples/03_score.py
python examples/04_parallel_fanout.py
python examples/05_confidence_routing.py
python examples/06_composite_scoring.py
python examples/07_guardrails.py

Fixtures by default. Add --live when TYPESAFE_API_KEY is set.

make test
make secret-scan

Repository map

├── app.py / static/          Operator lab + /home smart-home demo
├── jev_lab/                  HTTP client, cost, policy, house dispatcher
├── packages/js/              TypeScript harness CLI (@cobusgreyling/jev)
├── FRAMEWORK.md              The eight-layer showcase skeleton
├── BLOG.md                   Narrative article
├── examples/01–07            Progressive scripts
├── data/                     model-card, claims, fixtures
├── docs/                     Operator notes + smart-home + sources
├── skills/                   Agent skills (jev, jev-fanout, jev-guardrail, jev-route)
├── .grok/skills/             Grok-discoverable copies of those skills
├── tests/                    pytest
└── SECURITY.md               Key hygiene

This is not an official TypeSafe product. It does not replace typesafe-sdk. It does make the System One thesis interactive and measurable.

Use Jev when…

You need classify / route / score / verify on a hot path
Schema errors are unacceptable
You want calibrated probabilities, not a paragraph
You will keep side effects in code

Prefer a generative model when…

You need prose, code, or an explanation
The answer space is not a closed set
The task is counting, arithmetic, or date math
You want an agent that chooses its own next action

Docs index

Topic	Link
What people miss	docs/WHAT-PEOPLE-MISS.md
Smart home demo	docs/smart-home.md
Anti-patterns	docs/anti-patterns.md
Sources	docs/sources.md
Official jaggedness	Jev 1.13
Official patterns	docs.typesafe.ai/patterns
Workflow evals	evals.typesafe.ai
Agent skills	skills/

MIT · Unofficial · TypeSafe AI