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.
2140 experiments · showing 1681–1740
Find community Jev systems through their typed questions, with source snippets and maker posts alongside each listing.
SOURCE SCREENSHOTEXPAND ↗
Speak a voice diary, let Jev sort your thoughts, then review the notes, tasks and reminders before sending them to Notion.
SOURCE SCREENSHOTEXPAND ↗
Search an English–Icelandic dictionary by meaning, with Jev checking whether the shortlist contains a match and ranking the results.
SOURCE SCREENSHOTEXPAND ↗
Bring Jev Search's ranked links into Claude Code, Codex or a terminal through an MCP server, CLI and optional WebSearch hook.
SOURCE SCREENSHOTEXPAND ↗
Play Pong against Jev and watch its movement, shot selection, probabilities, and fallbacks in a live decision monitor.
SOURCE SCREENSHOTEXPAND ↗
Four ways of asking Jev to play 2048, compared with random moves and fixed rules, with replays of the resulting games.
ACTUAL RECORDING12 SEC ↗
OpenRouter's Jev Router picks the model and reasoning effort for each request, and holds a model while its prompt cache stays warm.
SOURCE SCREENSHOTEXPAND ↗
from the source
Jev Router picks the best model and reasoning effort for each request, balancing quality...It runs on Jev, TypeSafe's first System One model, and adapts as your conversation evolv...
OpenCodex asks Jev to choose a model and reasoning level for each Codex request, using the proxy's configured route whenever Jev is unavailable or returns an invalid answer.
SOURCE SCREENSHOTEXPAND ↗
from the source
requestBody = JSON.stringify({ model: JEV_MODEL, state, questions: buildJevRouteQuestion(options.candidates),});
A Phoenix Evals example pits Jev against a small grounded-versus-hallucinated answer benchmark.
SOURCE SCREENSHOTEXPAND ↗
from the source
const jev = typeSafeAi.evaluationModel("jev-latest");const jevEvaluator = createHallucinationEvaluator({ model: jev });const [jevResult, jevMs] = await timed(() => jevEvaluator.evaluate(example));
Docker Agent asks Jev whether a pending tool call is clearly read-only or needs the user's approval, treating uncertain calls cautiously.
SOURCE SCREENSHOTEXPAND ↗
from the source
type: choicechoices: read_only: Clearly read-only, with no credential access or external data transfer. risky: Modifies or deletes data, executes opaque code, accesses credentials, or transf... unknown: Insufficient information to establish the operation's effects.
GPTCache asks Jev whether a cached answer still fits a new request, taking differences in the requests, answer, and dates into account.
SOURCE SCREENSHOTEXPAND ↗
from the source
scores = { key: float(data["answers"][key]["noul"]) for key in self.DIMENSIONS}self.last_scores = scoresreturn min(scores.values())
Agent Native asks Jev which optional tools an agent should see at the start of a request, based on the user's recent context.
SOURCE SCREENSHOTEXPAND ↗
from the source
candidateStateKey: "candidate_tools",answerKey: "best_tool",question:Which tools should be loaded into the agent context first for this task? Pick the most u...limit: prefetchLimit,
Memorax asks Jev whether searching saved coding memories would materially help with the current prompt and, when available, the previous turn.
SOURCE SCREENSHOTEXPAND ↗
from the source
body: JSON.stringify({ model: JEV_MODEL, state, questions: { search_needed: SEARCH_QUEST...const answer = isRecord(body) && isRecord(body.answers) ? body.answers.search_needed : u...const probability = answer.noul;return { ok: true, decision: probability >= 0.5 ? "search" : "skip", probability, model:...
Interactive Shell uses Jev to decide when terminal work needs attention and which pre-authorized action may help.
SOURCE SCREENSHOTEXPAND ↗
from the source
questions.attention = { type: "choice", instructions: `Choose the primary visible state using only explicit current evidence..... criteria: { working: "routine work is active, retrying, progressing, or healthily waiting...",
haiku.rag provides a reusable System One judge for benchmark verdicts such as whether two answers are equivalent.
SOURCE SCREENSHOTEXPAND ↗
from the source
response = await self.client.system_one( self._state(ctx), {name: self._question()}, model=self.model)p = response.nouls[name].noulverdict = EvaluationReason(value=p >= threshold, reason=f"system_one p={p:.3f}")
Kimaki's OpenCode auto-mode asks Jev whether a pending tool action may run without another prompt, using the action, its arguments, and the latest user message.
SOURCE SCREENSHOTEXPAND ↗
from the source
state: serializedState,questions: { allow: { type: 'boolean', instructions: `Should this pending OpenCode tool action run automatically? ${CLASSIF...
ctx sift uses Jev to keep useful or contradictory passages while shrinking an agent's tool output.
SOURCE SCREENSHOTEXPAND ↗
from the source
let judgment = client.select(selection.task(), selection.candidates(text));if let Some(scores) = judgment.scores.as_deref() && let Some(proposal) = selection.propose(scores){ selected_count = proposal.kept_count();
TypeSafe Jev Gate adds a fail-closed review step to Hermes Agent's consequential tool calls.
SOURCE SCREENSHOTEXPAND ↗
from the source
answers = evaluate_tool(tool_name, args, **kwargs)secret_safe = _noul(answers, "no_secret_egress")score = risk(answers)eligibility = answers.get("eligibility", {}).get("choice")
Lintus turns plain-language YAML rules into Jev checks and reports threshold-crossing answers as lint offenses.
SOURCE SCREENSHOTEXPAND ↗
from the source
let mut query = jev::Query::new(format!("File: {path}\n\n{content}"));for rule in &task.rules { query.ask(&rule.id, rule.to_noul())?;}let response = self.client.perform(&query)?;
Search a retained window of Kubernetes or file logs with Jev, then inspect matching, possible, and unrelated evidence.
SOURCE SCREENSHOTEXPAND ↗
from the source
questions.insert(format!("e{i}_relevance"), json!({ "type": "choice", "criteria": criteria, "instructions": "Does events[i] address the log-search query..."}));
mem-jev asks Jev to judge ambiguous agent memories, then folds five semantic signals back into retrieval ranking.
SOURCE SCREENSHOTEXPAND ↗
from the source
evaluation, err := service.evaluator.Evaluate(ctx, request.State)record, err := NewJudgmentRecord(...)return ServiceResult{ Disposition: DispositionCommitted, Features: record.Features,
Mistype a subcommand and jevyoumean asks Jev which documented command you probably intended.
SOURCE SCREENSHOTEXPAND ↗
from the source
"intended": { Type: "choice", Instructions: instructions, Criteria: criteria(commands),}
A JevNPC trial asks whether a threatened Minecraft villager should ignore danger or call the guards.
SOURCE SCREENSHOTEXPAND ↗
from the source
criteria.put("IGNORE", "Start no new action...");criteria.put("CALL_GUARD", "Request help from the listed eligible guards...");String instructions = "Which available response should this NPC take now to stay safe?";
Keel can ask Jev which installed coding provider and model should receive a new task.
SOURCE SCREENSHOTEXPAND ↗
from the source
let input = SelectionInput { state: DecisionState { task, context, state_version }, candidates: prepared_routes,};let outcome = selector.select(input, &DecisionBudget::new(1)).await;
Triage messages, classify CSV rows, interpret voice commands, or let Jev steer Snake from a Java web playground.
SOURCE SCREENSHOTEXPAND ↗
from the source
var state = game.state(semantics);var decision = model.decide(state, Game.QUESTIONS);var choice = Answer.get(answers, Game.MOVE, Answer.Choice.class);var move = Direction.valueOf(choice.choice().toUpperCase(Locale.ROOT));return new Turn(game.step(move, random), ...);
Send text through a named semantic rule or your own question and receive a valid, invalid, or uncertain result.
SOURCE SCREENSHOTEXPAND ↗
from the source
"result": map[string]any{ "type": "choice", "instructions": question, "criteria": map[string]string{"true": "El texto cumple...", "false": "El texto no cump...}
Record live Jev decisions as redacted cassettes, then replay or perturb them without another provider call.
SOURCE SCREENSHOTEXPAND ↗
from the source
JevResponse response = delegate.send(request);repository.save(new Cassette(...));
Watch Jev choose a legal Tetris placement from a heuristic shortlist, then play the selected move.
SOURCE SCREENSHOTEXPAND ↗
from the source
client.choose(QUESTION_ID, JevPromptSupport.state(snapshot), INSTRUCTIONS, criteria);return selected.move();
AutoNomousBot shows Jev a Minecraft world snapshot, lets it choose from available actions, then performs the selection.
SOURCE SCREENSHOTEXPAND ↗
from the source
jevClient.askChoice(state, question, fallbackId)actionExecutor.execute(client, chosen);
Give Jev Pilot a mobile goal and it observes the screen, chooses a bounded UI action, and performs it.
SOURCE SCREENSHOTEXPAND ↗
from the source
let answers = self.judge.evaluate(...)?;let mut decided = catalog.resolve(&answers, &floors);
Add Jev-backed tool approval, chat triage, transcript curation, and model routing to OpenClaw.
SOURCE SCREENSHOTEXPAND ↗
from the source
const scored = await this.client.score({if (scored.level >= this.approvalLevel) {
TweetGuard uses Jev to sort individual X posts into spam categories while leaving uncertain posts visible.
SOURCE SCREENSHOTEXPAND ↗
from the source
const uncertain = answer.confidence < JEV_MIN_CONFIDENCE;const isSpam = answer.choice !== 'normal' && !uncertain;
DLQ Inspector asks Jev whether an ambiguous dead-letter message should be replayed, fixed, avoided, or investigated.
SOURCE SCREENSHOTEXPAND ↗
from the source
resp, err = a.Client.Evaluate(ctx, state, jev.ClassificationQuestions())out.Classification = jevChoiceToClassification(verdict.Classification)
Any Auto asks Jev whether coding-agent tool calls are risky, authorized, and policy-compliant before approving them.
SOURCE SCREENSHOTEXPAND ↗
from the source
json!({"model":self.config.approver.model,"state":input.state,"questions":questions})let mut result = into_review(&response)?;
Supercov asks named quality and security questions about source files, then turns Jev's answers into inspectable findings.
SOURCE SCREENSHOTEXPAND ↗
from the source
if let Some(health) = catalog::health(values) {weighted.push((answer.bytes, health));
Sentio classifies email with Jev and adjusts spam scores from unsolicited-mail and phishing probabilities.
SOURCE SCREENSHOTEXPAND ↗
from the source
let state = self.build_state(message_text, envelope_from, envelope_to);let score_delta = self.score_delta(unsolicited, phishing, confidence);
Run Laya’s typed decisions natively on Apple Silicon, then watch its terminal Snake demo choose moves with safety corrections.
ACTUAL RECORDING12 SEC ↗
Josh Rosen rounds up the first two weeks of JevOps — decision-model calls showing up in telemetry, incident response, SRE agents, CI, and deploys — and argues DevOps is packed with exactly the small semantic judgments Jev is built for.
SOURCE SCREENSHOTEXPAND ↗
Does Jev hold up outside demos? Glean's team benchmarked it on four workloads they run in production and published the wins next to the losses — routing came out ahead of their LLM router, reranking did not beat their search stack.
SOURCE SCREENSHOTEXPAND ↗
The weekly How I AI digest leads with Claire Vo spending a week inside Jev: nine cents to make sense of two years of ChatPRD pull requests, about four dollars for 200,000 classifications, and a working answer to which tasks still deserve a frontier model.
SOURCE SCREENSHOTEXPAND ↗
Claire Vo opens the hood on her week in Jev: nine cents to classify two years of ChatPRD pull requests, her Claude and Codex sessions sorted in minutes, and a real-time voice app where the quote API, not the decision model, was the slowest part.
SOURCE SCREENSHOTEXPAND ↗
Two days after Jev launched, Cribl's AI Research team published a grounded early look at what decision models can and cannot do inside a telemetry pipeline, benchmarking Jev against the purpose-built classifiers and LLM judges they already operate.
SOURCE SCREENSHOTEXPAND ↗
Most log records never need a language model's opinion, so Jev Logs scores each one first: how diagnostically useful it is, how urgent, and whether deeper analysis is warranted. Only confident, low-value records skip the expensive branch, and the archive always gets everything.
SOURCE SCREENSHOTEXPAND ↗
jevbrief fixes the least glamorous part of a Jev integration: what the model actually gets to read. Adapters for CI runs, pull requests, OpenTelemetry exports, and other noisy sources cut state down to what matters, and every dropped line gets a reason.
SOURCE SCREENSHOTEXPAND ↗
jevtraces brings the same annotation idea to spans: an alpha OpenTelemetry processor asks Jev whether an operation looks worth diagnosing, whether it's business-critical, and whether to keep it, then caches the verdicts and stamps them on later matching spans. Every span is retained regardless.
SOURCE SCREENSHOTEXPAND ↗
Datadog's Agent Observability team shows Jev doing double duty in their eval stack: notebooks score production agent spans as they arrive through online evals and run the same focused questions offline, mapping typed answers onto Datadog metrics.
SOURCE SCREENSHOTEXPAND ↗
SREGym gave its incident-solving agent a hesitant reviewer: before any diagnosis gets submitted, Jev has to agree the evidence supports it, and when a test is proposed, Jev ranks which one is worth running first. Pass rate moved from 20 to 24 out of 50.
SOURCE SCREENSHOTEXPAND ↗
jev-oncall routes production alerts with one typed Jev call per alert, then lets boring, auditable code decide what the probabilities are allowed to do. The model judges; a page threshold, a review band, and a fail-open default do the rest.
ACTUAL RECORDING12 SEC ↗
A confidence-gated incident router that consults an ownership registry first and Jev only when routing needs interpretation: choice picks the response team, score places business impact, noul estimates blocked critical work, and low confidence routes to human review instead of action.
SOURCE SCREENSHOTEXPAND ↗
A security-operations experiment set that runs Jev as a first-line decision layer across SOC work: phishing classification, authentication triage, BEC detection, investigation routing, and prompt-injection detection, with a unified PhishGuard pipeline and an eval suite of staged states.
SOURCE SCREENSHOTEXPAND ↗
jev-ci-pathfinder lets Jev decide which optional CI jobs a change actually needs, one allowlisted job at a time. Always-on jobs, path rules, and dependency closure never leave code, and when Jev is unavailable the action runs its configured conservative set rather than guessing.
SOURCE SCREENSHOTEXPAND ↗
This canary-rollout demo keeps the state machine in charge: it decides which transitions exist, Jev weighs evidence only at model-eligible branches, and the captured run holds at 5% traffic before promoting as evidence firms up.
SOURCE SCREENSHOTEXPAND ↗
A demo incident-response agent that puts the loop in Temporal and the judgment in Jev: typed next-action choices, a severity score, and safety probabilities drive durable activities, with retries, worker crashes, and a human-approval signal all surviving restarts.
SOURCE SCREENSHOTEXPAND ↗
A red-team of Jev 1.13 got it to approve harmful tool calls 70.1% of the time under direct misuse and 43.5% under indirect injection. The fix has a nice shape: Jev re-judges its own tool calls before they run, and both attack rates roughly halve.
SOURCE SCREENSHOTEXPAND ↗
Agent payments settle in milliseconds, and x402check spends that instant asking Jev six questions about the payer: known threat, sanctions, laundering, lookalike domain, operation class, trust. Every verdict ships as a signed attestation, so the check is provable, not promised.
SOURCE SCREENSHOTEXPAND ↗
Jevry browses for you one Jev choice at a time: it plans a site's steps first, then walks them with deliberately boring automation.
SOURCE SCREENSHOTEXPAND ↗
An ACP and MCP adapter that hands Jev's three typed question types to any LLM agent, built for Roblox computer use.
SOURCE SCREENSHOTEXPAND ↗
One Minecraft body with a split brain: GPT-6 Astra plans, and Jev answers action, safety, and urgency in a single parallel call.
SOURCE SCREENSHOTEXPAND ↗
CartPole, MountainCar, Acrobot, and FrozenLake trained with reinforcement rewards that come from Jev's calibrated scores instead of a hand-written reward function.
SOURCE SCREENSHOTEXPAND ↗
Freedoom, ViZDoom, and a decision model that has never held a shotgun: the dashboard plays Jev's run back with a dollar figure next to every call.
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.