A playful collection of decision demos spanning dispatch, moderation, search, desktop control and a 32-agent arena.
ACTUAL RECORDING12 SEC ↗
The directory02 / 02 · Experiments
Find something worth exploring.
Every category, from games and tools to repositories and writeups, traced to a primary source.
Same ideas.
Different paths.
This is the full registry, not just the featured picks. Figures like speed, cost, and stars are maker-reported or captured snapshots, not JevMade measurements.
1509 experiments · showing 1–60
Browser Use's agent asks Jev to choose the next browser operation and target element from a fresh table on every step.
SOURCE SCREENSHOTEXPAND ↗
from the source
git clone https://github.com/browser-use/jev-ultrafast.gitcd jev-ultrafastuv synccp .env.example .env# Add TYPESAFE_API_KEY and TEXT_MODEL_API_KEY.
Speak an English or Ukrainian drawing command while pointing, and Jev translates it into an action on a tldraw canvas.
ACTUAL RECORDING12 SEC ↗
Jev plays headless Doom from object bearings, distances, health, and ammo rather than game frames.
ACTUAL RECORDING12 SEC ↗
Jev plays NES Super Mario Bros. by choosing controls from emulator state rather than pixels.
SOURCE SCREENSHOTEXPAND ↗
from the source
NES emulator -> telemetry/RAM parser -> structured JSON -> Jev Choice -> controller inpu...
A simulated Panda robot asks Jev how to move a block over a barrier.
ACTUAL RECORDING12 SEC ↗
Jev picks a flight path through each pipe, then Flappy Bird turns that route into precisely timed flaps.
ACTUAL RECORDING12 SEC ↗
Jev steers a Pac-Man recreation using structured game state.
ACTUAL RECORDING12 SEC ↗
Jev places Tetris pieces by considering holes, stack height and the current board.
ACTUAL RECORDING12 SEC ↗
A Composio provider that asks Jev which tool to use and which arguments to supply.
SOURCE SCREENSHOTEXPAND ↗
from the source
decision = provider.decide(tool_set, REQUEST) if decision["kind"] == "abstain": print(f"Jev abstained: {decision['reason']}") if decision["risk"] != "read_only": print(f"Not executing {decision['tool']}: its risk class is {decision['risk']}")
Demonstrates a bounded Jev action chooser inside Cua Driver loops, with mock and credentialed live paths.
SOURCE SCREENSHOTEXPAND ↗
Run Convai Innovations' open-weight typed-decision model locally with a Jev-shaped request format; Laya is an independent model, not TypeSafe's Jev.
SOURCE SCREENSHOTEXPAND ↗
Train and self-host Qwen-based decision models that answer Jev-style typed questions in one pass; kev is an independent implementation, not Jev.
SOURCE SCREENSHOTEXPAND ↗
Reads supported Android chat screens, judges conversational intent, and offers replies that users may fill before sending manually.
SOURCE SCREENSHOTEXPAND ↗
Vercel's open framework for building agents.
SOURCE SCREENSHOTEXPAND ↗
An agentgateway example that screens LLM requests and responses with Jev guardrails.
SOURCE SCREENSHOTEXPAND ↗
A Claude Code plugin that uses Jev to decide which context survives compaction.
SOURCE SCREENSHOTEXPAND ↗
from the source
claude plugin marketplace add tamaratran/fast-jev-compactionclaude plugin install fast-jev-compaction@fast-jev-compaction
NanoJev trains a 0.6B parallel decision model and compares it with Jev across maze, Snake, and ViZDoom tasks.
SOURCE SCREENSHOTEXPAND ↗
An AI stock-research team that can use Jev to rate an investment and flag risks before writing its report.
SOURCE SCREENSHOTEXPAND ↗
from the source
"rating": { "type": "choice", "instructions": "综合团队讨论结论，该股票的最终投资评级是哪一档？", "criteria": dict(RATING_MAP), },
SemIf runs open models locally as probabilistic semantic conditionals on an RTX 3090.
SOURCE SCREENSHOTEXPAND ↗
Play Snake or request typed decisions from Core ML ports of Laya on Apple hardware; the models are open-weight alternatives, not Jev.
SOURCE SCREENSHOTEXPAND ↗
Drive a Mac through its accessibility tree, with Jev choosing an operation and a live element on each turn.
ACTUAL RECORDING12 SEC ↗
from the source
const req = buildRequest("save the file", { app: "TextEdit", window: "Untitled" }, offer...assert.deepEqual(Object.keys(req.questions), ["target", "command", "present", "destructi...assert.equal(Object.keys(req.questions.target.criteria).at(-1), "none");assert.equal(req.questions.present.type, "noul");assert.equal(req.questions.target.criteria[`@${S}:e3`].what, 'button "Save"');
Symfony's first-party bridge for asking Jev several choice, score and yes-or-no questions about one shared state.
SOURCE SCREENSHOTEXPAND ↗
Jevlike trains a compact model to choose among a list of text options that can change from request to request.
SOURCE SCREENSHOTEXPAND ↗
from the source
uv venvsource .venv/bin/activateuv pip install -e '.[dev]'jevlike-data synthetic --output data/syntheticjevlike-train data/synthetic/train.jsonl \
A trading bot asks Jev what to do on every Monad block.
SOURCE SCREENSHOTEXPAND ↗
A Vercel Labs command-line tool for generating content from the terminal.
SOURCE SCREENSHOTEXPAND ↗
choice
Which team should handle this request?
Run a local, non-autoregressive System One model behind Python and JavaScript clients compatible with Jev’s request shape.
SOURCE SCREENSHOTEXPAND ↗
A macOS menu-bar companion that turns a typed click task into a short sequence over locally detected controls.
SOURCE SCREENSHOTEXPAND ↗
from the source
"absent": .noul( instructions: "Is the control needed to make the next bit of progress on... ), "kind": .choice( instructions: "How should that element be acted on?",
Astra plans a Minecraft route while JEV chooses bounded actions for a recorded, source-verifiable playthrough.
SOURCE SCREENSHOTEXPAND ↗
Uses local OCR and Jev to judge Windows chat messages, rank three drafted replies, and fill one without sending it.
SOURCE SCREENSHOTEXPAND ↗
Simple Jev exposes open Hugging Face models as structured classifiers and includes playgrounds, evaluations, and an autonomous-driving demo.
SOURCE SCREENSHOTEXPAND ↗
A cautious Codex computer-use skill that chooses UI targets from accessibility text while local policy guards consequential actions.
SOURCE SCREENSHOTEXPAND ↗
Adds typed probability readouts and calibration heads to ordinary language models, with Jev-style evaluations and saved results.
SOURCE SCREENSHOTEXPAND ↗
Indexes Jev SDKs, integrations, applications, evaluations, and learning material with source-backed annotations.
SOURCE SCREENSHOTEXPAND ↗
Collects Jev projects, reusable agent skills, and editable decision recipes with an optional validated Python client.
SOURCE SCREENSHOTEXPAND ↗
Rizzo Flow serves local Spark models through a Jev-compatible API and visual playground without generating answer text.
SOURCE SCREENSHOTEXPAND ↗
See where a Claude Code session made progress and where it went in circles, with Jev reviewing a redacted trace.
SOURCE SCREENSHOTEXPAND ↗
from the source
key: "progressingEfficiently", label: "Progress", question: "Did the agent make steady, meaningful progress toward the user's task... higher_probability_is_better: true, answer: MetricAnswer::Noul,
Shows intent, risk, and ranked reply suggestions beside supported macOS chats while keeping message sending under user control.
SOURCE SCREENSHOTEXPAND ↗
Browse a bilingual, source-linked index of Jev projects grouped by decision pattern, with dated link and call-site checks.
SOURCE SCREENSHOTEXPAND ↗
A Hermes toolkit for model routing, retrieval triage, memory, skill selection, context trimming and interface decisions.
SOURCE SCREENSHOTEXPAND ↗
Identify IRS forms and schedules from their text with Jev.
SOURCE SCREENSHOTEXPAND ↗
from the source
const q1: Record<string, ChoiceQuestion> = { kind: { type: 'choice', instructions: 'What kind of page is this?', criteria: KIND_C... form: { type: 'choice', instructions: FORM_INSTRUCTIONS, criteria: firstList }, } const r1 = await opts.backend.ask(state, q1)
A Jev-compatible server reads answer-label logits from Qwen3.6-35B-A3B while SGLang reuses shared prompt prefixes.
SOURCE SCREENSHOTEXPAND ↗
An MCP server that lets a coding agent ask Jev whether a claim is supported, a page is worth reading, or its own patch is safe.
SOURCE SCREENSHOTEXPAND ↗
choice
How does the evidence relate to claim `claim0` (`<claim text>`)? · supports / contradicts / says_nothing
A software-factory foreman built around TypeSafe's Jev model.
SOURCE SCREENSHOTEXPAND ↗
from the source
TYPESAFE_API_KEY=your-key-here
Tunes Jev decision functions from labelled examples by surfacing uncertain cases and proposing revised definitions through GEPA.
SOURCE SCREENSHOTEXPAND ↗
A 151M non-autoregressive decision model pairs typed outputs with benchmark receipts, calibration analysis, and browser-oriented artifacts.
SOURCE SCREENSHOTEXPAND ↗
A Mac menu-bar assistant that uses Jev to choose what to click.
SOURCE SCREENSHOTEXPAND ↗
from the source
nonisolated static let maxChoices = 255 nonisolated static let doneThreshold = 0.70 nonisolated static let absentThreshold = 0.50 nonisolated private static let noneKey = "__none__" endpoint: URL = URL(string: "https://api.typesafe.ai/v1/systemone")!) {
A staged Jev code-review workflow with a local dashboard.
SOURCE SCREENSHOTEXPAND ↗
from the source
Noul risk matrix -> Choice + Score file profiles -> Choice evidence selection -> Choice mechanism classification -> Score severity
TypeSafe agent skills
officialOfficial agent skills for writing software that uses TypeSafe's System One API.
SOURCE SCREENSHOTEXPAND ↗
from the source
claude plugin marketplace add typesafe-ai/skillsclaude plugin install typesafe@typesafe-ai
Point DocJev at a PDF and a list of categories you wrote in plain English, and Jev names the category or marks where each new document starts.
ACTUAL RECORDING12 SEC ↗
noul
Does page number 5 start a new source document, rather than continue page number 4?
skillbox is a self-hosted, versioned Agent Skills library with MCP access, scoped clients, and optional Jev recommendations.
SOURCE SCREENSHOTEXPAND ↗
AbdelStark curates official references and community work around TypeSafe, Jev, and System One models.
SOURCE SCREENSHOTEXPAND ↗
A macOS automation loop that reads the screen with OCR, asks Jev what to do next, and clicks the chosen target.
SOURCE SCREENSHOTEXPAND ↗
from the source
git clone https://github.com/awlevin/typesafe-computer-usecd typesafe-computer-useuv synccp .env.example .env # fill in the keys
Audits each LLM code-review finding with one credibility question before deterministic CI rules decide whether to fail.
SOURCE SCREENSHOTEXPAND ↗
from the source
"questions": { QUESTION_ID: { "type": "noul", "question": VERIFY_QUESTION, "instructions": VERIFY_INSTRUCTIONS,
Suggests a safe moment to compact an agent session by combining two Jev judgments with context usage.
SOURCE SCREENSHOTEXPAND ↗
Port browser-use's Jev-driven browser agent to local Laya on Apple Silicon while retaining an optional hosted Jev mode.
SOURCE SCREENSHOTEXPAND ↗
A browser video editor whose Jev-powered Director turns plain-language requests into structured timeline operations.
SOURCE SCREENSHOTEXPAND ↗
from the source
// Shared shapes for the Director's timeline operations and vault-media placement.// Filled by Jev on the server (POST /director/route), executed in Home.tsx.export type TimelineOperation = | 'delete' // remove clip(s) | 'split' // cut clip(s) at a time
Notra turns material from your work into publishable content.
SOURCE SCREENSHOTEXPAND ↗
Adds typed tool, skill, owner, and quality decisions to agent harnesses without taking over their planning or execution.
SOURCE SCREENSHOTEXPAND ↗
A PostgreSQL extension for asking questions about tables in plain language with Jev.
SOURCE SCREENSHOTEXPAND ↗
Jev, briefly
A decision model, not a chatbot.
A chat model generates text token by token and hopes you parse it. Jev never generates a word. Your code sends the state of the world plus typed questions; Jev returns every answer in one parallel pass — typed values with calibrated probabilities and a confidence score your code can trust. About 70–500 ms end to end, $0.042 per million input tokens, output free. Trained with what TypeSafe calls Reinforcement Learning for Calibrated Decisions.
choice Pick one of these.
You define the options. Jev returns the chosen option, a probability for every option, and a confidence. The workhorse of routing, agents, and games — the answer space is always legal, so there is nothing to parse and nothing to hallucinate.
score Rate this on a rubric.
You describe ordered levels — say trivial / normal / critical. Jev returns the level, the probability of each, and a confidence. Scores turn fuzzy judgment (“how severe is this log line?”) into a number-free decision your code can branch on.
noul Is this statement true?
You assert a statement; Jev returns the probability that it is true — a noul. Moderation, verification, guardrails, “does this diff actually fix the bug?”: one question, one calibrated probability.
Official material lives at typesafe.ai and docs.typesafe.ai. JevMade is an independent community registry — not affiliated with or endorsed by TypeSafe AI.
Recurring lessons
Patterns that keep showing up.
One decision per tick
Games, drones, trading bots, and browser agents all converge on the same loop: serialize the state, ask one decisive question, act, repeat. Jev’s latency makes the loop feel instant — the model lives inside the control loop, not outside it.
Confidence as a gate
The answer says what; the confidence says whether to act. The most reliable entries threshold on confidence to route edge cases to a slower model or a human — automation with an honest escape hatch.
Shrink the choice space in code
Makers rarely ask Jev to pick from everything. Local tactics prune 225 gomoku moves to ~40; DOM filters turn a page into an element table; code narrows options, Jev judges within them.
Jev judges, LLMs talk
A recurring split: Jev makes every decision cheaply and instantly, and a small LLM is only invoked when a human-facing string must actually be written. Decision and generation are separate budgets.
Made something with Jev?
Send the link — repo, live demo, post, or video — plus a line on what it does and which primitives it uses. Every entry is verified against its primary source before it ships.
Submit your experiment → hello@JevMade.com