Read & learn
Written guides.
Understand Jev, one idea at a time. Walkthroughs, recipes, and writeups — our notes first, the original next.
Explore the Jev ecosystem.
Learn, build, and play with Jev — from deep technical guides to creative experiments.
Watch & learn
Video guides.
See an idea take shape. Tutorials, demos, and deep dives, organized by topic and credited to their creators.
Make & explore
Experiments.
See what builders made with Jev. Games, tools, repositories, articles, and more — traced to their sources.
The directory03 / 03 · 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.
268 experiments · showing 1–60
World Monitor benchmarks Jev's headline severity and topic labels against cached labels and a saved judged set.
SOURCE SCREENSHOTEXPAND ↗
from the source
async function callJev(titles) { const t0 = performance.now(); for (let attempt = 0; ; attempt++) { const r = await fetch(JEV_ENDPOINT, { method: 'POST',
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 ↗
NanoJev trains a 0.6B parallel decision model and compares it with Jev across maze, Snake, and ViZDoom tasks.
SOURCE SCREENSHOTEXPAND ↗
Nimble compares Jev's typed probabilities with locally trained schema adapters on curated decision datasets.
SOURCE SCREENSHOTEXPAND ↗
from the source
payload = request_payload(row, model)response = request_with_retry(transport, payload, attempts=attempts, sleep=sleep)validate_teacher(row, response, model)return jev_row(row, response, time.perf_counter() - started)
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 ↗
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 \
Run a local, non-autoregressive System One model behind Python and JavaScript clients compatible with Jev’s request shape.
SOURCE SCREENSHOTEXPAND ↗
Adds typed probability readouts and calibration heads to ordinary language models, with Jev-style evaluations and saved results.
SOURCE SCREENSHOTEXPAND ↗
Mini Taiwan Pulse tests whether Jev can pick the map layers relevant to a visitor's question.
SOURCE SCREENSHOTEXPAND ↗
from the source
const probabilities = parseNoulAnswers(body, batch.questionIds);for (const candidate of batch.candidates) decisions.set(candidate.key, knownDecision(probabilities.get(questionIdForLayer(candid...
Compare Jev's market probabilities and trading choices in a Polymarket benchmark that never places real-money orders.
SOURCE SCREENSHOTEXPAND ↗
from the source
probability = noul_probability(answers, "probability")action = choice_value(answers, "action")confidence = choice_value(answers, "confidence")
A 151M non-autoregressive decision model pairs typed outputs with benchmark receipts, calibration analysis, and browser-oriented artifacts.
SOURCE SCREENSHOTEXPAND ↗
Vector Graph RAG can rerank graph relations with Jev and compare the resulting retrieval recall.
SOURCE SCREENSHOTEXPAND ↗
from the source
"reranker_provider": self.rag.settings.reranker_provider,"jev_model": self.rag.settings.jev_model if self.rag.settings.reranker_provider == "jev"..."jev_threshold": self.rag.settings.jev_threshold if self.rag.settings.reranker_provider...
See Jev's published chess row alongside language models in a benchmark that separates wins, losses and illegal moves across repeated games.
SOURCE SCREENSHOTEXPAND ↗
An Apple Silicon experiment scores visual candidates directly against shared context using a local model.
SOURCE SCREENSHOTEXPAND ↗
from the source
jev-visual examples/photo-request.json --model-path .models/Qwen3.5-0.8B-4bit
JevK5 trains and serves an open-weight decision model that returns closed-choice probabilities in one forward pass; it is not TypeSafe's Jev.
SOURCE SCREENSHOTEXPAND ↗
A benchmark of how Jev and a language model choose among 100 tools for a personal assistant.
SOURCE SCREENSHOTEXPAND ↗
from the source
AGENT_MODEL=mock npm run eval # llm-directAGENT_MODE=jev-classifier AGENT_MODEL=mock JEV_STUB=1 npm run eval # jev-classifier
Reimplements Jev-style bounded decisions locally, exposing a compatible contract over prompt-only and open-weight backends.
SOURCE SCREENSHOTEXPAND ↗
Reflex is an open Qwen3.5-based model that maps state and typed questions to calibrated probabilities.
SOURCE SCREENSHOTEXPAND ↗
choice
Which team should handle this?
Replay Jev and chat-model judgments side by side on comment datasets imported from CSV or Excel.
SOURCE SCREENSHOTEXPAND ↗
Run an open Thai-and-English decision model that answers Jev-compatible typed questions in one forward pass.
SOURCE SCREENSHOTEXPAND ↗
A Python library for checking an agent's work with several Jev questions in one call.
SOURCE SCREENSHOTEXPAND ↗
from the source
async def evaluate(self, state: Any, questions: dict[str, Question]) -> BackendRespo... body: dict[str, Any] = {"state": state, "questions": questions_to_typesafe(quest... if self.model: body["model"] = self.model headers = {"Authorization": f"Bearer {self.api_key}", "Content-Type": "applicati...
Goodwatch's search arena measures Jev's latency, token use and request size on movie-search questions.
SOURCE SCREENSHOTEXPAND ↗
from the source
const [attribute, fingerprint] = await Promise.all([ call(attributeRequest(q)), call(fingerprintRequest(q)),]);
Run Nano, Small or Large open-weight models behind Jev-style Choice, Score and Noul requests; these are independent models, not TypeSafe's Jev.
SOURCE SCREENSHOTEXPAND ↗
A gateway that makes a Cerebras-backed model answer in Jev's format, for comparison with the real thing.
ACTUAL RECORDING12 SEC ↗
from the source
packages/demos/ Side-by-side benchmark UI, workloads and native Jev adapterpackages/api/ Cerebras structured-output adapter, validation and supporting APIscripts/ Offline benchmark summarizers and media toolingdocs/benchmarks/ Raw exports, source hashes, metrics and quality comparisons
OpenJev is a Jev-compatible decision server built on DiffusionGemma.
SOURCE SCREENSHOTEXPAND ↗
noul
Does the customer need a reply within the hour?
Compare Jev and two language models as they control the same simulated xArm7 apple-placement task.
ACTUAL RECORDING12 SEC ↗
An open reimplementation trains a lightweight decision head for Choice, Score and Noul questions.
SOURCE SCREENSHOTEXPAND ↗
score
The capital of France is
Checks Jev questions against labelled data and reports whether each can gate decisions, rank examples, or carries no useful signal.
SOURCE SCREENSHOTEXPAND ↗
Provides a fast random baseline that speaks the Jev protocol without reading the prompt.
SOURCE SCREENSHOTEXPAND ↗
Qwen Choice reads option logits from a local vision-language model to classify one image without generating an answer.
SOURCE SCREENSHOTEXPAND ↗
Maps where Jev succeeds or fails using reproducible API receipts, external studies, and bilingual explanations rather than a leaderboard.
SOURCE SCREENSHOTEXPAND ↗
Decider fine-tunes Qwen3.5-2B to make typed choices with calibrated probabilities in one pass.
ACTUAL RECORDING12 SEC ↗
JevMLX adds parallel constrained decisions and schema-valid JSON to MLX models on Apple Silicon.
SOURCE SCREENSHOTEXPAND ↗
from the source
pip install git+https://github.com/bnsd55/jevmlx # libraryuv tool install git+https://github.com/bnsd55/jevmlx # CLI onlygit clone https://github.com/bnsd55/jevmlx && cd jevmlx && ./setup.sh # dev
A compact PyTorch teaching implementation explores a Jev-inspired encoder, shared state cache, and typed readout heads.
SOURCE SCREENSHOTEXPAND ↗
Mini-Jev reads option-letter logits from a frozen Qwen3-4B instead of asking the model to generate JSON.
SOURCE SCREENSHOTEXPAND ↗
from the source
git clone https://github.com/r-ms/mini-jev.git && cd mini-jevuv sync # torch 2.5.1, transformers 4.57.6, xgrammar 0...MINIJEV_DEVICE=mps uv run python demo/server.py # or MINIJEV_DEVICE=cuda# first start downloads Qwen/Qwen3-4B-Instruct-2507 (revision pinned) and warms up, then...open http://127.0.0.1:8765/
Compare Rust plus Jev with Python plus DeepSeek as both sides flag simulated Pix scams and trace the same criminal network.
SOURCE SCREENSHOTEXPAND ↗
Explores an agent runtime where Jev selects from bounded, pageable actions while local code manages context and execution.
SOURCE SCREENSHOTEXPAND ↗
Run Laya's typed-decision model on Apple Silicon with selectable memory modes and a Pong latency demo; this is not TypeSafe's Jev.
SOURCE SCREENSHOTEXPAND ↗
Evaluates whether classifier confidence is calibrated and derives human-review thresholds from the cost of mistakes.
SOURCE SCREENSHOTEXPAND ↗
This study benchmarks parallel typed decisions on unmodified 1.5B–8B models running on Apple Silicon.
SOURCE SCREENSHOTEXPAND ↗
from the source
git clone https://github.com/rorshopping/jev-on-a-laptopcd jev-on-a-laptop./setup.sh # creates .venv, installs mlx-lm, clones the upstream engine./run_benchmark.sh # default: Qwen2.5-1.5B-Instruct-4bit
An experiment in getting structured JSON answers from DiffusionGemma, with benchmarks against Every and Jev.
SOURCE SCREENSHOTEXPAND ↗
Studies calibration-aware reinforcement learning for adaptive decision systems through training and benchmark harnesses.
SOURCE SCREENSHOTEXPAND ↗
Train a Qwen-based closed-choice model by applying Brier loss directly to candidate-token probabilities; it explores Jev's decision-first idea but is not Jev.
SOURCE SCREENSHOTEXPAND ↗
A benchmark labels 1,000 app reviews with Jev and Gemini 3.8 Flash side by side.
SOURCE SCREENSHOTEXPAND ↗
from the source
git clone https://github.com/goodrahstar/jev-column-race.gitcd jev-column-racecp .env.example .env# Add TYPESAFE_API_KEY, and LLM_API_KEY for the Gemini lane.node server.mjs
An independent text-scoring model that tries to improve on the jevlike starter project.
SOURCE SCREENSHOTEXPAND ↗
from the source
$ jevbetter-benchmark --reference /path/to/jevlike --epochs 8
Compare decision models at Texas Hold'em through cash and sit-and-go leaderboards, hand replays and tables for custom agents.
SOURCE SCREENSHOTEXPAND ↗
Self-host a Jev-compatible decision engine over Ollama or OpenAI-compatible models, with typed outputs and calibrated probabilities; it does not run TypeSafe's Jev.
SOURCE SCREENSHOTEXPAND ↗
Serve Laya typed decisions through a Rust Candle runtime for local CPU or GPU inference; this is a Jev-compatible alternative, not Jev.
SOURCE SCREENSHOTEXPAND ↗
Add natural-language constraints to JEPA planning by asking Jev whether probe-described imagined states violate each rule, then folding probabilities into planning cost.
SOURCE SCREENSHOTEXPAND ↗
TypeAR studies type-safe constrained decoding for autoregressive language models.
SOURCE SCREENSHOTEXPAND ↗
A reproducible evaluation suite measures calibration, selective risk and latency in probabilistic decision models.
SOURCE SCREENSHOTEXPAND ↗
Evaluates synthetic refund-support traces with four typed criteria and records auditable results as an Opik experiment.
SOURCE SCREENSHOTEXPAND ↗
OpenVons answers finite-choice questions with probabilities across text, images and Japanese voice commands.
SOURCE SCREENSHOTEXPAND ↗
An independent bilingual decision model inspired by Jev, designed to run locally.
SOURCE SCREENSHOTEXPAND ↗
system-one performs batched, single-token choice inference with open language models through a TypeSafe-compatible interface.
ACTUAL RECORDING12 SEC ↗
from the source
from system_one import SystemOnefrom typesafe_sdk import Choiceengine = SystemOne.from_pretrained("Qwen/Qwen2.5-0.5B-Instruct")result = engine.system_one( state="My shoes arrived in the wrong size. Can I exchange them?",
Run the open-weight Laya decision model locally on Apple Silicon through an Apple-focused runtime; it is Jev-compatible but does not use Jev.
SOURCE SCREENSHOTEXPAND ↗
A bilingual English-Chinese guide explains notable Jev applications, how they work and their trade-offs.
SOURCE SCREENSHOTEXPAND ↗
Pit local Laya against Jev in matched snake races and fighting-game rounds, or assign either model to both sides.
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