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.Experiments
Every category, from games and tools to repositories and writeups, traced to a primary source.
Real-time research powered by
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.
409 experiments · showing 181–240
Play four-seat Texas Hold'em while comparing Jev's betting judgments with a rules coach and Monte Carlo equity estimate.
SOURCE SCREENSHOTEXPAND ↗
Play terminal Sudoku yourself or watch Jev choose cell-and-digit moves while the benchmark tracks whether it solves the grid.
SOURCE SCREENSHOTEXPAND ↗
Compare Jev with a conventional language model on route planning, Minesweeper, Hanoi, and a simulated drone.
SOURCE SCREENSHOTEXPAND ↗
Jev picks strategic preferences while a traditional engine searches for moves on a chess.com board.
SOURCE SCREENSHOTEXPAND ↗
Compare ways of showing Jev candidate moves in a head-to-head arena for two falling-block games.
SOURCE SCREENSHOTEXPAND ↗
Compare Jev’s Battleship shots with random, hunt-and-target, and placement-density strategies on seeded games.
SOURCE SCREENSHOTEXPAND ↗
Compare 21 ways of describing legal SameGame moves to Jev, with every raw request and response preserved.
SOURCE SCREENSHOTEXPAND ↗
Pit Jev against random and hand-tuned players in Tetris, Snake, and 2048 while logging confidence for every move.
SOURCE SCREENSHOTEXPAND ↗
Triage simulated warehouse-robot incidents with Jev and compare its decisions with a small local classifier.
SOURCE SCREENSHOTEXPAND ↗
Compare Jev with general language-model judges on one multilingual booking-inquiry routing task.
SOURCE SCREENSHOTEXPAND ↗
A Chinese-language experiment suite for Jev's context limits, bilingual judgments, batching and game decisions.
SOURCE SCREENSHOTEXPAND ↗
An original platformer and reinforcement-learning lab testing whether frozen Jev risk features help PPO learn faster.
SOURCE SCREENSHOTEXPAND ↗
from the source
API_URL = "https://api.typesafe.ai/v1/systemone"MODEL = "jev-1.13.0"ACTION_NAMES = tuple(action_names("simple"))ACTION_CRITERIA = { "noop": "Release all buttons; coast or wait, and release jump.",
Political Compass and Pew typology probes for Jev, alongside a deliberately awkward character-by-character text decoder.
SOURCE SCREENSHOTEXPAND ↗
from the source
def choice(instructions: str, criteria: dict[str, str | None]) -> dict: """A Choice question: pick one of up to 255 options.""" return {"type": "choice", "instructions": instructions, "criteria": criteria}
Battleship where Jev returns a probability of a ship for every open cell, creating both the heat map and the shot.
SOURCE SCREENSHOTEXPAND ↗
from the source
def argmax_cell(probabilities: Mapping[str, float], cells: Sequence[Cell]) -> Cell: if not cells: raise ValueError("no cells to choose from") best_cell = cells[0] best_probability = float(probabilities.get(cell_id(best_cell), 0.0))
A roguelike dungeon director where Jev sketches the next room's role, size, danger, enemies, loot and secrets.
SOURCE SCREENSHOTEXPAND ↗
from the source
"room_type": { "type": "choice", "choice": "chamber", "confidence": 0.8, "probabilities": {"chamber": 0.8, "room": 0.2},
Runs Jev and a frontier model down parallel lanes, answering the same timed traffic-hazard choices and comparing the scorecards.
SOURCE SCREENSHOTEXPAND ↗
from the source
export const CALL_CRITERIA = { left: "Swerve left. The hazard is real and the left side of the road is open.", right: "Swerve right. The hazard is real and the right side of the road is open.", hold: "Hold the line. Do not swerve. The markings disagree with the open side, neither...
A controlled city simulation comparing four kinds of institutional memory, with Jev scoring each term beside a deterministic scoreboard.
SOURCE SCREENSHOTEXPAND ↗
from the source
ENDPOINT = "https://api.typesafe.ai/v1/systemone"MODEL = os.getenv("TYPESAFE_MODEL", "jev-latest") "prosperity": { "type": "score", "instructions": "Prosperity over the whole term: jobs relative to population at...
Turns Jev probability distributions into one plain-language sureness verdict, from certain through torn to clueless.
SOURCE SCREENSHOTEXPAND ↗
from the source
QUESTIONS = { "department": {"type": "choice", "instructions": "Which team should handle this mess... "severity": { "type": "score", "instructions": "How severe is the reported issue? If there is no issue, pick th...
Run open-Jev decisions in your browser without an API key or inference server.
SOURCE SCREENSHOTEXPAND ↗
Check generated answers against several rubrics, with Jev judging them in batches.
SOURCE SCREENSHOTEXPAND ↗
Test what an answer means, not just which words it contains, with Jev assertions for Vitest and Jest.
SOURCE SCREENSHOTEXPAND ↗
Compare Jev’s typed classifications with classical ML pipelines across the maker’s eight datasets.
SOURCE SCREENSHOTEXPAND ↗
Replay six Jev chess experiments, from unaided legal moves to tactical filters and Stockfish-assisted variants.
ACTUAL RECORDING12 SEC ↗
Can a cheap classifier label every agent failure in a training run? Applied Compute clusters a sample of traces, then has Jev do the labeling.
SOURCE SCREENSHOTEXPAND ↗
from the source
Each model outputs a floating-point score from 0 to 1 for each of 14 failure modes.A score at or above the decision threshold marks that failure mode as present.
3,608 one-subject-swapped statements put to Jev as forced true/false picks, in three wordings and eleven languages, with every raw answer published.
SOURCE SCREENSHOTEXPAND ↗
from the source
self.url, self.model = JEV_URL, "jev-latest"
Runs the CLASH cross-modal contradiction test against Jev with each image replaced by its COCO caption, then reports accuracy and modality bias.
SOURCE SCREENSHOTEXPAND ↗
from the source
r = self.session.post(f"{BASE_URL}/v1/systemone", json=body, timeout=self.timeout)
Auto-marks 2,054 MR-GSM8K maths scripts against a teacher rubric with one Jev call each, then checks the marks against human annotators.
SOURCE SCREENSHOTEXPAND ↗
from the source
const result = await client.systemOne(
The open data behind an independent benchmark that puts Jev and six LLMs through the same typed questions on PubMedQA, Banking77 and HelpSteer2.
SOURCE SCREENSHOTEXPAND ↗
from the source
run header: {"type":"header",...,"model":"typesafe-ai/jev",...,"probability_source":
The 2026 Korean CSAT Korean-language paper turned into a JSON state and typed choice questions, with Jev's answers, probabilities and error analysis kept alongside.
SOURCE SCREENSHOTEXPAND ↗
from the source
{\"1\": { \"type\": \"choice\", \"instructions\": { \"question\": \
An independent benchmark of Jev 1.13 against a cheap and a frontier LLM on SMS spam and Banking77, measuring accuracy, ECE, Brier, latency and cost.
SOURCE SCREENSHOTEXPAND ↗
from the source
resp = requests.post(\"https://openrouter.ai/api/alpha/decisions\",
A drop-in proxy that answers your app with Jev byte for byte while asking the same question of local doubles, then reports where they would have decided differently.
SOURCE SCREENSHOTEXPAND ↗
from the source
\", \"headers\": { \"authorization\": \"Bearer ${TYPESAFE_API_KEY}\" }, \"model\": \"jev
Measures what the live Jev API actually returns across eight realistic use cases, reporting cost, latency and response shape rather than rate-card arithmetic.
SOURCE SCREENSHOTEXPAND ↗
from the source
ENDPOINT = BASE_URL +
A tree with a place for every closed question a person or program could ask, filled with real questions answered by Jev, plus a 3D explorer where each question is a star.
SOURCE SCREENSHOTEXPAND ↗
from the source
r = await self.http.post(GATEWAY_URL, content=canonical(req.body))
Tests Jev on scientific judgement questions, then follows the consequences: what those choices do to the results that depend on them.
SOURCE SCREENSHOTEXPAND ↗
from the source
ENDPOINTS = {'openrouter': 'https://openrouter.ai/api/alpha/decisions',
Reranks fixed MS MARCO candidate sets by asking Jev one relevance question per query and document pair, then scores the runs with trec_eval.
SOURCE SCREENSHOTEXPAND ↗
from the source
payload = {"model": args.model, "state": {"query": topics[qid]
A first hands-on measurement of Jev: seven small scripts probing its documented jaggedness, latency scaling, calibration, negation consistency and behaviour when every option is wrong.
SOURCE SCREENSHOTEXPAND ↗
from the source
r = client.system_one(A, { (then noul: Noul(instructions=REFUND) and choice: Choice(inst
Asks whether a System One model can do the job of an LLM judge in an agent guardrail, with the hypotheses and decision rules fixed before any run.
SOURCE SCREENSHOTEXPAND ↗
from the source
resp = await self._client.system_one(
Six tracks of experiments against the System One endpoint, from size limits and needle-in-a-haystack retrieval to prompt injection, option bias and a fifteen-job bake-off.
SOURCE SCREENSHOTEXPAND ↗
from the source
resp = client()
A head-to-head benchmark of open Laya against hosted Jev on byte-identical inputs in the same run, with question hashes verified identical before comparison.
SOURCE SCREENSHOTEXPAND ↗
from the source
DIRECT_URL = "https://api.typesafe.ai/v1/systemone"
Puts the Traditional Chinese TMMLU+ exam to Jev, one four-way question per item, and scores it the way the official leaderboard does.
SOURCE SCREENSHOTEXPAND ↗
from the source
ENDPOINT = "/v1/systemone"
Pairs a curated map of the Jev ecosystem with independent, reproducible benchmarks of the three primitives, confidence gating, fan-out latency and agent control.
SOURCE SCREENSHOTEXPAND ↗
Reduces 845 preflop spots to a single fold, call or raise decision, asks Jev five times each, and compares the result with four deterministic reference styles.
SOURCE SCREENSHOTEXPAND ↗
from the source
URL = "https://api.typesafe.ai/v1/systemone"
Turns Jev's probabilities into a routing threshold with a provable bound on how many queries get silently misrouted.
SOURCE SCREENSHOTEXPAND ↗
from the source
DECISIONS_URL = "https://openrouter.ai/api/alpha/decisions"
Benchmarks Jev on Brazil's ENEM 2025 exam against open-LLM baselines, contrasting a raw transcribed state with a structured one and measuring calibration.
SOURCE SCREENSHOTEXPAND ↗
from the source
response = self.client.system_one(state=state, questions={"gabarito": Choice(instruction
A pip-installable benchmark for typed System One models that measures what the probabilities buy: calibration against a noise floor, wording sensitivity, selective prediction and cost.
SOURCE SCREENSHOTEXPAND ↗
from the source
DEFAULT_URL =
Five hospital tasks, a thousand synthetic patients, and Jev answering every one, with Claude writing the scenarios and reviewing the results.
SOURCE SCREENSHOTEXPAND ↗
from the source
response = await client.system_one(state, questions)
Pre-flights a System One question before you put its number behind an if, reporting a decision flip rate rather than a confidence score.
SOURCE SCREENSHOTEXPAND ↗
from the source
resp = await fetch(`${baseUrl}/v1/systemone`, { method: "POST", headers: { Authorization
Tests Jev as a search reranker, and asks whether its confidence can tell you which queries are worth spending more compute on.
SOURCE SCREENSHOTEXPAND ↗
from the source
r = await self.http.post( URL, json=payload, headers={"Authorization": f
Tests Jev pairwise rankings of FOMC statements against rate decisions, FedLock scores, and a Haiku comparison.
SOURCE SCREENSHOTEXPAND ↗
Measures how Jev accuracy and calibration change when the same tasks are presented in English and Spanish.
SOURCE SCREENSHOTEXPAND ↗
Reports a preregistered adversarial evaluation spanning calibration, batching, answerability, option counts, and out-of-domain logic.
SOURCE SCREENSHOTEXPAND ↗
Run Jev-compatible typed judgments locally in a browser or Node server using small natural-language inference models.
SOURCE SCREENSHOTEXPAND ↗
Compare Jev with VIGILIA’s own labels on a sealed set of risky and harmless agent actions.
SOURCE SCREENSHOTEXPAND ↗
Petru placed a local classifier behind a Claude pre-tool-use hook as a Jev-free comparison.
SOURCE SCREENSHOTEXPAND ↗
Vincent put Jev inside a browser agent and reported faster steps and fewer model calls, with lower accuracy.
SOURCE SCREENSHOTEXPAND ↗
An early optimizer compares Jev, classic algorithms and an LLM agent on data-center load balancing and power management.
ACTUAL RECORDING12 SEC ↗
Nick Khami’s experimental endpoint applies inference engineering to make an open model behave like Jev.
ACTUAL RECORDING12 SEC ↗
A fraud pipeline lets Jev classify emails quickly, then sends uncertain cases to Kimi K3.
ACTUAL RECORDING12 SEC ↗
Florian Darroman compares Jev and Fable 5.1 on a post-scheduler build task and reports Jev finishing work faster.
ACTUAL RECORDING12 SEC ↗
WOLF backtests a Jev trading setup and reports better results after removing time data and most technical inputs.
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 experimentSign in first so we can reply to you. Submissions are reviewed, not published automatically.