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 61–120
Maps 1,000 recent AI papers into 24 topics with Jev, then compares a seeded sample with an LLM judge.
SOURCE SCREENSHOTEXPAND ↗
jevfire batches many Jev-style decisions over one shared context on CUDA language models.
SOURCE SCREENSHOTEXPAND ↗
from the source
git clone https://github.com/kikoncuo/jevfire.gitcd jevfirepython -m venv .venvsource .venv/bin/activatepython -m pip install -e '.[dev]'
LegalForecastBench provides an alpha benchmark and evaluation workflow for LegalForecast-MTD.
SOURCE SCREENSHOTEXPAND ↗
A word-level language model that drafts with n-grams and asks Jev to check chunks of the result.
SOURCE SCREENSHOTEXPAND ↗
from the source
jev-lm probe "the little dog was" --top 5jev-lm verify "the capital of france is" "paris" "london"jev-lm gen "the little dog" --words 18 --tracejev-lm gen "she went to the door" --draft --tracejev-lm eval --file data/heldout_en.txt --positions 60
typesafe-local asks a local MLX model typed questions and returns calibrated probabilities without generating or parsing prose.
SOURCE SCREENSHOTEXPAND ↗
from the source
{ "model": "mlx-community/Qwen3-1.7B-bf16", "answers": { "refund_requested": {"type": "noul", "noul": 1.0, "raw_mass": 1.0}, "department": {"type": "choice", "choice": "billing",
Verdict is a 151M-parameter ModernBERT decision engine with calibrated uncertainty and no autoregressive generation.
SOURCE SCREENSHOTEXPAND ↗
Organizes independent Jev studies by calibration, consistency, prompt injection, abstention, and other failure modes.
SOURCE SCREENSHOTEXPAND ↗
Reimplements the Jev request contract over existing SGLang or vLLM deployments by scoring finite candidate tokens.
SOURCE SCREENSHOTEXPAND ↗
Tests whether filtering coding-agent tool results through Jev improves task success, context size, or cost against an unfiltered arm.
SOURCE SCREENSHOTEXPAND ↗
Split local Laya batches across Apple Neural Engine and GPU with a Jev-shaped API; it accelerates an independent model, not Jev.
SOURCE SCREENSHOTEXPAND ↗
Rev trains three Qwen decision models and compares their accuracy, latency, cost, scaling, and Snake play with hosted Jev.
SOURCE SCREENSHOTEXPAND ↗
Follow Eikos and Jev as they paper-trade the same Hyperliquid markets from identical snapshots and precommitted rules.
SOURCE SCREENSHOTEXPAND ↗
How well does Jev understand Korean, including medical text? This benchmark publishes questions, results, runtime and cost.
SOURCE SCREENSHOTEXPAND ↗
from the source
.venv/Scripts/python.exe -X utf8 -m jevbench.overview_figure
system-one-open recreates Jev-style calibrated decisions with small open Gemma models.
SOURCE SCREENSHOTEXPAND ↗
from the source
modal run build_data.py # ~5 min, CPU onlymodal run typesafe_eval.py # secondsS1_GPU=H100 modal run train.py --name lite-270m --model unsloth/gemma-3-270m-it --tier l...S1_GPU=H100 modal run evaluate.py --run lite-270mS1_GPU=H100 modal run train.py --name e2b-full --model google/gemma-4-E2B-it --tier full...
LitJev wraps ordinary language models in a Jev-like layer for typed choices, scores, and binary judgments.
SOURCE SCREENSHOTEXPAND ↗
choice
What does the customer want?
jevcal derives confidence thresholds, checks calibration, and watches typed decision models for drift against an LLM teacher.
SOURCE SCREENSHOTEXPAND ↗
from the source
pip install "git+https://github.com/abhixhek/jevcal" # corepip install "jevcal[anthropic] @ git+https://github.com/abhixhek/jevcal" # + Claude a...
Surveys evidence on Jev and compatible decision models, separating calibration claims, selective control, and open implementations.
SOURCE SCREENSHOTEXPAND ↗
Test whether Jev can choose compiler optimisation passes from program features, with its decisions compared against measured outcomes.
SOURCE SCREENSHOTEXPAND ↗
Build and test Jev decision workflows in Python, then use them from the command line or an MCP client.
SOURCE SCREENSHOTEXPAND ↗
choice
What is the tone? · calm / frustrated
A closer look at Jev 1.13.0, with controlled prompts and the raw answers available to inspect.
SOURCE SCREENSHOTEXPAND ↗
jev-chat constructs chatbot replies by repeatedly turning Jev probabilities into hierarchical speculative decoding choices.
SOURCE SCREENSHOTEXPAND ↗
from the source
git clone https://github.com/adhyaay-karnwal/jev-chatcd jev-chatuv sync --extra devcp .env.example .env # set TYPESAFE_API_KEY
Can Jev find better agent skills than embedding search? This evaluation tests both on Chinese and English queries.
SOURCE SCREENSHOTEXPAND ↗
from the source
uv venv --python 3.12 .venv && uv pip install --python .venv/bin/python -e ".[dev]".venv/bin/pytest -q # 28 tests.venv/bin/python -m jse robustness # re-scores cached runs, no API key...export OPENROUTER_API_KEY=... # env only; never written to disk.venv/bin/python -m jse pool # ~30 min first time (bge-m3 over 3...
jev_stock experiments with short-term market-direction forecasts derived from structured financial data.
SOURCE SCREENSHOTEXPAND ↗
This open-weights Jev alternative returns typed, calibrated choices in one Hugging Face or vLLM forward pass.
SOURCE SCREENSHOTEXPAND ↗
from the source
pip install open-alternative-jev # Hugging Face backendpip install "open-alternative-jev[vllm]" # + vLLM backendpip install "open-alternative-jev[quant]" # + bitsandbytes 8-bit / 4-bit loading
Sort through training data with a Rust tool that asks Jev to score rows and filter out unwanted examples.
SOURCE SCREENSHOTEXPAND ↗
Split source code mechanically, ask Jev what each fragment is, and render the guesses as syntax highlighting.
SOURCE SCREENSHOTEXPAND ↗
Compare Jev with another judge in a blind arena of bounded answer choices, then reveal the models after voting.
SOURCE SCREENSHOTEXPAND ↗
Tests Jev and the open-weight Laya counterpart as typed judgment layers in a look-ahead-free quantitative research stack.
SOURCE SCREENSHOTEXPAND ↗
Ask typed questions directly about images through a shared Qwen3-VL encoding instead of first producing captions.
SOURCE SCREENSHOTEXPAND ↗
Calibre measures which model suits each part of a dataset, then routes requests using those results.
SOURCE SCREENSHOTEXPAND ↗
Test Jev on chess puzzles and on figuring out which game character you're talking to.
ACTUAL RECORDING12 SEC ↗
from the source
jevcommon/client.py rate-limited async TypeSafe client; persists every raw request/...jevchess/ chess benchmark state.py board → state at three levels (fen / ascii / rich); move descri... experiments.py A–G: Choice, Score fan-out, hierarchical, perception, mate-in-1... engine.py positions.py Stockfish ground truth; position generation
Eight small examples apply Jev to mechanical and electrical engineering decisions.
SOURCE SCREENSHOTEXPAND ↗
jev-gate experiments with assigning Claude Code tasks to different models according to Jev's judgment.
SOURCE SCREENSHOTEXPAND ↗
jev-harness adds confidence gates, shadow runs, reusable recipes, and evaluations around Jev decisions.
SOURCE SCREENSHOTEXPAND ↗
jevgpt coaxes a non-generative decision model into chatting by choosing the response one step at a time.
SOURCE SCREENSHOTEXPAND ↗
trade-jev backtests Jev as a buy, sell, or hold trader on Nasdaq futures order-book data.
SOURCE SCREENSHOTEXPAND ↗
from the source
uv run python -m trade_jev.run --days 2026-06-23 # one...uv run python -m trade_jev.run --days 2026-06-08,2026-06-09,2026-06-10 # sev...uv run python -m trade_jev.run --all # all...uv run python -m trade_jev.run --days 2026-06-23 --policies hold,random,imbalance # bas...
jevify helps agents spot suitable Jev tasks, frame typed questions, and draw on recent community experiments.
SOURCE SCREENSHOTEXPAND ↗
noul
Does `passages.a` help answer `query` with a step, prerequisite, or constraint?
Play chess against Jev or Stockfish, or let the two models play each other.
SOURCE SCREENSHOTEXPAND ↗
Turn NYC 311 complaint text into explorable maps of reporting volume and Jev-estimated impact.
SOURCE SCREENSHOTEXPAND ↗
Before building an RPG, this project tests whether Jev's characters can react to context and keep a secret.
SOURCE SCREENSHOTEXPAND ↗
from the source
// Retries are handled here, not in the SDK, so a recorded latency is one attempt.const client = new TypeSafeClient({ defaultModel: MODEL, retry: { maxRetries: 0 } });
Metask Jev-Lab trains open typed-decision models and publishes calibrated option probabilities alongside reproducible JevBench evaluations.
SOURCE SCREENSHOTEXPAND ↗
DecisionBridge gives existing language models an interface for explicit choices, scores, calibration, and human-review thresholds.
SOURCE SCREENSHOTEXPAND ↗
from the source
import osfrom decisionbridge import Choicefrom decisionbridge.backends.api import APIDecisionModelmodel = APIDecisionModel("openai", os.environ["OPENAI_MODEL"], mode="json")choice = Choice("What is the main intent?", {
This benchmark compares Jev with a strong language model at attributing agent failures in the text subset of Who&When Pro.
SOURCE SCREENSHOTEXPAND ↗
This research notebook pairs runnable Jev demos with a close reading of the model's public claims.
SOURCE SCREENSHOTEXPAND ↗
Build a chat response one character at a time, with every character chosen by Jev.
SOURCE SCREENSHOTEXPAND ↗
from the source
git clone https://github.com/kesku/jev-freeform.gitcd jev-freeformnpm installexport TYPESAFE_API_KEY="your-key"npm start
Little Airways puts Jev's decision-making into a small browser-based flying demo.
ACTUAL RECORDING12 SEC ↗
from the source
demo/ the runnable demo (three.js WebGPU via CDN, zero build step)research/ the Jev capability dossier (live-captured API behavior, pricing, patterns)resources/ dream-loop artifacts (target, prompts, rounds) and screenshotsoriginals/ the Astra single-file build (fully offline) and the GLM round-3 build
This benchmark compares Jev with Cohere, ZeroEntropy, and a chat model on 14 reranking datasets.
SOURCE SCREENSHOTEXPAND ↗
from the source
uv synccp .env.example .env # JEV_API_KEY, OPENROUTER_API_KEY, ZEROENTROPY_API_KEYuv run candidates/build.py # downloads the datasets, builds BM25 top-30 and the "ab...uv run candidates/build_nevir.pyuv run run.py --model jev-score-batch --dataset all --workers 16
This harness evaluates Jev Ultrafast on curated research-browser cases and generates a field report from each run.
SOURCE SCREENSHOTEXPAND ↗
from the source
export JEV_ULTRAFAST_ROOT=/path/to/jev-ultrafastpython scripts/run_suite.py \ --jev-root "$JEV_ULTRAFAST_ROOT" \ --cases cases/research_browser_v1.yaml \ --out results/latest
jev-sec-bench tests Jev blindly on prompt-injection and vulnerable-code detection tasks.
SOURCE SCREENSHOTEXPAND ↗
This demo compares Jev's abstract-screening decisions with gold labels from the ASReview SYNERGY dataset.
SOURCE SCREENSHOTEXPAND ↗
This static dashboard compares Jev and two Luna settings on Japan's 2026 Common Test.
SOURCE SCREENSHOTEXPAND ↗
Explore a Monte Carlo tree search that uses Gemini to propose paths and Jev to judge them.
SOURCE SCREENSHOTEXPAND ↗
from the source
TYPESAFE_API_KEY=your_typesafe_api_key_here# Optional model overrides:AGY_MODEL=gemini-3.8-flash-medium
PadFlow contributes anonymized land-development decisions for testing confidence-aware models such as Jev.
SOURCE SCREENSHOTEXPAND ↗
qwen-rlcd trains Qwen3.5-0.8B to make Jev-style choices, scores, and binary judgments with calibrated confidence.
SOURCE SCREENSHOTEXPAND ↗
choice
Which team should handle this?
RISC-jeV asks how far state-in, decision-out inference can go by wiring Jev up as a RISC-V CPU.
SOURCE SCREENSHOTEXPAND ↗
These OpenJev experiments classify sentence pairs, including whether one statement follows from or contradicts another.
SOURCE SCREENSHOTEXPAND ↗
from the source
from classify import ClassifierFactory, OpenJevSettings, TextPairclassifier = ClassifierFactory.create(OpenJevSettings())predictions = classifier.classify([ TextPair("A man is playing a guitar.", "Someone is making music."), TextPair("A man is playing a guitar.", "A man is sleeping."),
A practical guide and prototype for putting Jev triage in front of Herdr-managed coding agents.
SOURCE SCREENSHOTEXPAND ↗
Experiment with a Jev-powered Codenames player and a separate CLI for checking determinism, sensitivity and calibration.
SOURCE SCREENSHOTEXPAND ↗
from the source
questions[word] = style === "compact" ? { type: "noul", instructions } : { type: "noul",
An experimental pharmacy checker that asks Jev the same question five ways and looks for agreement.
SOURCE SCREENSHOTEXPAND ↗
from the source
//! - **Not deterministic.** Identical requests returned 0.03, 0.03, 0.03,//! 0.04, 0.04. This is the entire reason the protocol needs a stability//! gate; the client must not paper over it with caching.//! - **Billing is `input_tokens` only**, with a ~281-token fixed overhead per//! call and state billed once regardless of question count. Batching many
A record of five attempts to find a worthwhile job for Jev—and the reasons each stayed out of production.
SOURCE SCREENSHOTEXPAND ↗
from the source
const r = await client.systemOne({ state: { command }, questions: { harm: noul(question) }, ...(model ? { model } : {}), });
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