Lesson 03 · Jev 101

Writing questions: Choice, Score, Noul

The 101 skill is writing questions your code can act on. This page restates TypeSafe’s primitives, State, and how-to-build docs. It is not a substitute for the official pages.

Official sources:Primitives · Choice · Score · Noul · State · How to build.

Ask for one snap judgment

Official primitives page: System One is built for a judgment a knowledgeable person makes in a second, given the right context. “Does this message convey urgency?” is a good question. “Analyze this message and determine the best course of action” is not — that needs slow reasoning, so break it up and compose in code.

Official how-to-build: “This is probably the most important concept in this guide.” Broad questions hide several judgments. Atomic questions expose them so code can inspect, tune, and combine.

Every question has four parts

Pick the type your code can act on

State first: string, object, or array

Official State page: state is the content you evaluate — a support message, a passage, or application state. All questions in a request see the same state and run independently. Text only; no image, audio, or video yet. English is primary.

Official State page: one object can hold a conversation, an order, and a policy. Point questions at fields with a backticked dot-and-index path.

• Does `ticket.messages[0].text` request a refund?
• Does `refund_policy` support the refund requested in `ticket.messages[0].text`, given `order.charges`?

Keep content in state and judgments in questions. Filter first — official name: context rot. Do not rely on model weights when your knowledge base has the current fact.

Official Python questions.md: state cannot be None, but values inside an object may be. You can mix question objects (Noul / Choice / Score) and dictionaries that carry a type key in the same request.

Writing a Choice

Official choice.md: the model never sees the question id. Names and descriptions both go in. Official JS choice() requires instructions and criteria; the ChoiceQuestiontype still allows instructions to be omitted ornull.

• The selected `choice` is the option with the highest probability. Official: `probabilities` cover every option and sum to 1. `confidence` is computed from how peaked that spread is — a 0.61 / 0.35 split is not a 1.0 answer.
• Option names and descriptions both go to the model. Write descriptions that separate options. When the names are already clear, official examples use `null` descriptions (`calm` / `frustrated` / `angry`).
• Ask every Choice your code might need in one call, including speculative ones. Code ignores the unused answers. Official: adding options costs a few tokens each; give the full list (up to 255) plus `other` when the set might miss an input.
• When two options keep getting confused, describe each with an object: what it covers, what belongs next door, a few examples. Official: field names such as `what` / `not_for` / `examples` are yours — none are reserved.

Official triage example: five Choices in one call. Code readsdepartment, ignores shipping_issue unless that team won, copies a second team when its probability is over 0.25, and asks the customer when resolution confidence is under 0.5. Those cutoffs are the docs’ illustration, not a TypeSafe SLA.

Community cookbook (chr-kelly/jev-cookbook): Choice criteria is a map of option → description (up to 255). The TypeSafe API returns 422 on a malformed Score, but can silently accept {"options":[…]} as a Choice withone option named options — and answer every message with it at confidence 1.0. Author: we shipped that bug;python eval/lint.py catches it. Their waitlist note is stale vs official X Sep 20. Need more than 255 options?jev-tree walks a taxonomy (default 32/call) and returnsunavailable instead of inventing a leaf. Under the cap,jev-palette sends all 77 commands as one Choice and treats the returned distribution as the ranking. Need meaning checks after a Zod shape?zod-jev semantic() is Noul-only — Choice and Score still need the raw SDK. Use the primitive that fits the question:jev-drone saw a muddy Choice (p=0.24) and a clean Noul (target_truly_lost=0.12) on the same scene — search is gated on the Noul. Need seconds from a transcript?Sponsor Skip renders L042| lines and lets Jev pick IDs — code maps them back to time (official semantic-find pattern). Jev does not write timestamps. Need a quote, not a paraphrase?jev-reviewer lets Jev pick A001-style line ids; code copies the span verbatim. Not found is an answer. Need more than 255 tools?jev-eval-agent sends 100 mocked tools +respond_to_user as one Choice — look-alike distractors count as wrong tools. Need a draft scorecard, not a rewrite?vibecheck asks every rubric in one request; code writes the verdict. Jev never sees pixels. Need Korean vs English?jev-korean-benchmark says translating the instructions buys nothing — fine-grained meaning is shaky in both languages. Need a none-of-these on a local clone?openvons always includes it;zhihz/openjev does not unless you add the option. Need coordinates, not pixels?jev-behavior-study Snake and city labs send board or scene state — Jev is not vision. Prose framing reached 14/16 foods; the original prompt reached 1/16. Need a line quote without generation?commit-sentry and send-guard let code number the spans; Jev only picks an id or answers yes/no. Need fifty scores at once?turbo-rerank sends one Score per BM25 hit plus a Noul “does any answer exist?” Need seven labels per email without writing a reply?inbox-blitz computes priority in code; category confidence below 0.7 goes to a human — and its UI name Sift is not bohutang/sift. Need a squiggle that dies when the line changes?jev-lint discards stale in-flight answers; Noul ≥0.60 raises a marker. Distinct from perch. Need a canned reply without generation?agent-assist lets Jev pick a macro id; the template stays hand-written. Need same-incident grouping?log-sentinel does that in code — Jev only scores the event. Distinct from jevlogs. Need hold-before-publish that never bans self-harm?ModStream sends care, never a ban, at self_harm ≥0.5. Need a voice endpoint without sending audio?jev-voice-turn fires at turn_complete ≥0.85 plus 250 ms pause. Distinct from jev-voice-browser. Need to drive a Mac without screenshots?jev-ax-pilot walks the Accessibility tree and lets Jev pick an element id — notJevPilot (tables, not pixels). Need seven judgments on a 911 transcript?jev-dispatch merges when the duplicate Noul is ≥0.6. Need every chat line judged, then re-filtered locally?jev-firehose stores the probabilities; sliders do not re-query. Distinct fromModStream. Need thirty scores on a keystroke?jev-instant-search sends one request of 30 Scores plus five query questions — notturbo-rerank. Need “the last PDF I opened”?jev-launcher parses the time window in code; Jev only picks. Enter stays explicit. Need 32 independent policies?jev-swarm batches up to 8 agents per request; physics stays in code. Need a radar call that is not geometry?jev-tower lets code predict the conflict; Jev picks one instruction. Distinct fromjev-drone. Need a whole column from a header?judge-sheets maps JUDGE/PICK/RATE onto Noul/Choice/Score and batches by source text. Need notes while the sentence is still in the air?Fern (folder live-minutes) resolves dates and owners in code. Distinct fromjev-voice-turn. Need “make this louder” without the command name?nl-palette asks 66 commands plus four argument slots in one request. Distinct fromjev-launcher. Need a spoken Mac action?say lets Jev pick the action; OpenAI transcribes the audio and writes the reply. Need a paste that never invents?exact-paste copies a clipboard slice; fail → the whole chunk. Distinct fromsend-guard. Need a paper FX decision that is not a win-rate?JevPip asks two independent Choices; code owns quantity and TTL. Need one animal instinct without leaking stats?jev-ecosystem-sim sends coarse text bands — notjev-swarm. Need a reply made of dictionary Choices?JevChat does that as an experiment — official Jev is not trained to generate text. Need a ticket department plus urgency in one call?jev-demo-router auto-dispatches at ≥0.75; 100/100 is 100 generated tickets. Need jailbreak and verdict as separate questions?jev-demo-guardrails lets a 95% jailbreak with 37% action confidence go to review. Need a prompt checklist without generation?pi-prompt-enhancer asks 15 questions; the hint copy is a static table. Need a resume scored against one JD?jev-demo-screener asks eight questions; swap the JD and the verdict flips. Need comment spam and tone as separate axes?jev-demo-moderator does one call per line — notModStream. Need retrieved chunks gated before an LLM?jev-demo-rag keeps a chunk only at relevance ≥ 2 and injection < 0.5. Need many conclusions from the same premises?jev-demo-logictest asks up to 12 Nouls; the 70/30 colors are UI, not the model. Need every sentence labelled?semantic-microscope sends one request per sentence — not one document dump. Need a Pi gate that never treats silence as yes?pi-jev-permit asks one Noul and blocks the rest. Need an editorial pass that does not count?jev-editor-skill keeps counts in lint.py. Need a Tetris piece placed without look-ahead?typesafe-jev-dojo lets code shortlist ~10 legal placements, then one Choice — notjev-tetris. Need a statute shortlist reranked, not a legal answer written?NRS Navigator asks three questions on up to 30 candidates; Jev cannot recover a keyword omitted. Need six ticket judgments in one call?hj01857655/jev-router asks team, priority, urgent, auto-reply, mood, and complexity together — and its README $0.0578/M is not official $0.042/M. Need a receipt classified without generating a category label?jev-receipt-classifier asks category, necessity, and fixed_cost in one call. Need a citation checked against its own passage?TheWayWithin/jev-bench asks one support Choice — 85.7% is 36/42, author numbers. Need a Choice that can refuse every listed option?jev-first-look saw an all-wrong menu still pick shipping untilnone_of_these existed. Need seven AML flags in one call?the UiPath demo asks five Nouls, one Score, and one Choice — the LLM still does the arithmetic. Need last-mile exceptions without letting Jev own the verdict?LogiPulse asks four questions; code applies B › D › C › A. Need 12 shortlisted moves, not a written plan?jev_practice asks one Choice — notjev-tetris. Need 43 independent questions about one trip?nak1b/jev-experiments asks them in one reading — cost follows the question count. Need a queue of independent yes/no items, not one person waiting?jev-ultralightspeed packs them; packing a single waiting item is slower. Need inbox priority and category without a chat model writing labels?bo7/jev_test asks two Choices — noteddydong/JevTest. Need Fact / Definition / Reason for RAG routing?QA-Classifier-Jev is one Choice; --dry-runprints the payload. Writing a Choice whose menu might miss a real case?wellposed flags a missing other before you send it — a closed menu can answer at 1.00 and still be wrong. Need one sentence about whether a name still matches the body?mizchi/jev-lint asks that Noul — the finding is a candidate, not a verdict. Need a local loop of expected answers you wrote yourself?jevals keeps those labels in your workspace — they are not Jev’s. Packing a whole eval library into one request?openlayer-ai/jevals does that — TypeSafe backend mock-tested, not live. Need a yes/no that will not snap to 0 and 1?should-ai-kill-us-all asks the Noul under the Choice for that reason.Auto-Guard sends five questions in one flight; JevShield sends Choice + Noul + Score. Need a menu larger than one Choice can hold?tax-doc-classifier keeps the first list under 255 and adds a second Choice only for five corporate forms.Simple Jev caps Choice at 50;openjev-sglang rejects 65 answers;jev-visual caps a question at 26 options.decider accepts Choice 2–255 and isolated Score levels — packed order can change up to 12% of answers.Nimble caps an enum at 26 and rejects prompts over 2,048 tokens.JoshuaSP/open-jev lets fields share attention — that is not official independence.wondertwins found one Choice over legal chess moves beat Score-per-move, and packed every NPC Noul into the same call.jev-secret-detection is one Noul per snippet — burst 100 triggered 529.quackd only offers Jev verbs whose parameters are a closed set.jev-libero stacks three Choices (intent → family → one of 27 inputs).jev-robot-control uses two sequential Choices — not one packed request.evaluate rejects Choice outside 1–255 and Score outside 1–10 before the API.qwen-rlcd has no Choice/Score/Noul heads yet — only a prefix-fork engine.padflow-jev-evals publishes three JSON schemas, not live Jev questions.pngwn’s scorer trained with an option cap of 16 and then evaluated banking77 at 77 — tickets_queue drops to 0.234.system-one-mini only answers five frozen heads.eve-rlcd caps Choice at 26 letters and writes Score as a 0–1 mean index — not official Score.Gemma E2B adds an Independent multi-label type that TypeSafe does not ship.decido names the accessors binary / choice / ordinal.jod keeps the Noul threshold on the schema — it is never sent to Jev.super-jev packs independent questions in one request; a tool result needs the next iteration.typesafe-ai-rs accepts a one-level Score the way Python does — official JS still wants two.typesafe-client types the key so a Choice answer will not compile as a Noul.codeitlikemiley/typesafe-sdk-rust has no noulsdict — callnoul("name").Jev4Mellea maps Noul to pass / fail / uncertain — the band is never a silent pass.typesafe-assist fans out intent / target / device, then maps onto HA intents — free-text still needs a fallback LLM.judging-with-typesafe saw a one-question rubric fail (sciency nonsense 0.74–0.82 vs textbook 0.48) — their working set is four questions.a0-typesafe-ai packs independent questions in onetypesafe_query.nibir1 measured a one-level Score live — official prose still wants two.jevmetrics asks four questions and filters on only the keep Noul.qualm names Noul isand requires anunsure branch.jev4k checks Choice 1–255 and Score 2–10 locally.unimtx accepts a one-option Choice locally.typesafe_api attaches a derived confidence to Noul — official Noul has none.typesafe.zig makes a one-level Score a compile error.rashedInt32/jev-mcp andjev-reach split Choice lists at the official 255 cap.jev-sift caps a call at 1–8 questions.typesafe_api_sdk README prints noul asfalse — official Noul is a probability.y0usaf/typesafe-cli rejects a one-option Choice and a one-level Score locally.jtsang4/jev-cli names Noulboolean and calls P(true) uncalibrated — official still claims calibration.hfiguera accepts a one-option Choice locally.TypeSafeAI.Sdk gives Noul no confidence field.valksor Go SystemOneAsmissing answers become zero.neo4jev asks Choice plus Noul in the same hop.askjev treats Noul 0.5 as a coin-flip, not medium.rhighs decide wants 2–100 Choice labels locally — official cap is 255.jev-recall asks one Noul per memory.BorisLeMeec find asks one Noul per file, then a Choice on chunks.jev-dsh-decision caps a call at 64 questions.typesafe-skill-router chunks at 240 — official Choice cap is 255.hermes-jev-plugin maps Scorelevels to officialcriteria.actiongate-jev asks six Nouls in one request — never one vague “is this safe?”.patdown asks one Noul per file × rule: “does this file violate …?”git-jev-stage asks one Choice per hunk withinclude /exclude /mixed.riff asks one Noul per semantic rule, plus one document-type question.rajdhakad router does not pick Jev’s top tier automatically — expected loss does.construct-auto-classifier asks one allow/deny Choice plus nine risk Nouls.jev-skill-gate asks one Noul per skill in one call.jevprune asks one Noul per remaining line.jevex lets Jev pick the action; the writer never emits a tool-call object.typesafe-migration-guard asks one SAFE/DANGER Choice.JevPromptCoach asks nine Nouls in one request.jev-workbench gives Noul probability only — no confidence field.jev-pref conditions are Noul; choices map labels to outcomes.agent-chaperone asks a small battery of typed screens per call.jev-pr-labeler asks typed label Choices; size is conceptual.jev-skillful asks one Choice over a shortlist plusnone, then Nouls for runner-ups.last-exit asks two Jev calls per turn — observations, then one available action.openpoke-meets-jev asks four email Nouls in one call; Noul has no confidence field.switchboard asks three risk Nouls plus one small/large Choice.typesafe-local reads one logit position; read raw_massbefore confidence.jevegis builds one Jev request from a CHECKS array.research_desk fans out seven questions in one exposure call; composition is code.magic-jev asks one Choice over four verdicts; code picks the phrase.jev-chat samples next units from Choice; Jev never emits the string.jev-phishing-bench asks nine questions in one call — a verdict, a mirror Noul, five signals.jev-browse asks operation and speculative targets in one request.jev-autopilot asks a Choice per stick plus two Booleans.lucianfialho/jev-model-router asks five request questions and never sends the catalog.Kamil’s compaction does not send tool-result bodies to Jev.lomeshdutta/skill-router keeps every question and threshold inquestions.py.typesafe-mod rides every dialog question on one request.TypeSafe Code Guard uses Choice + Noul, then a keyword scan for the “exact line.”omni-/ask-jev batches 13 Nouls plus one outcome Choice.dgui-hypermem logs analyze / rerank / supersede as instruction rows.jev-engineering asks two questions after hard rules and an allowlist.az9713 asks one Choice over four tiers.Mandrilsquad asks eleven questions in one call; code ranks.jev-organize picks dates and counterparties from candidates code already found.jev-predict-skill asks a breakdown Noul then a Choice over the target’s own labels.omp-typesafe ends every battery with a severity Score.stepwarden asks five questions before the tool runs.agent-handoff-gate uses Choice on claims; examples are fixtures.check-risk scores a policy index; Noul has no confidence.Frost asks Jev to assess the task, then Go picks a profile.hermes-jev-router exposesjev_route_planonly.jev-builder-loop you write the questions and priors; Jev returns probabilities.jev-firewall asks one Choice over allow / ask / block after rules.jev-judge takes 1–16 named questions and prints answers, never the state.hermes-jev admits a turn as OFF / WATCH / ON, then batches later questions.thiago-ss/jev-review asks finite typed questions, then code applies gates.omp-jevens-classifier asks one Choice, nine hazard Nouls, and a blast Score.pi-jev-code ranks search, gates edits, and reviews the task-baseline diff.twilwa/pi-typesafe batches four hazard Choices, then four Score critics.taste-lint sends rule context through Vercel AI Gateway for meaning judgments.typesafe-demo-mcp exposestypesafe_askover one shared state.typesafeai-review asks the catalog once per hunk, then code composes.zcode-jev batches one Noul per acceptance criterion indone-gate.hermes-jev-skills routes, filters memory, and sorts mail with typed questions.jot asks Jev which tool and arguments to run next.x-scanner asks six questions per post in one call.live-jev asks lane, speed, hazard, and yield in one call. Notokinaaudio/live-jev.jev-ego asks operation plus speculative targets.xtags asks one intent Choice and three Nouls.AskJev asks next action, target, irreversible, and goal-done.aside-jev asks Jev to pick a candidate id.vinilana/jev-browser asks Jev for the next action and field, then a separate eval call.computer-use-jev asks action, target, goal-done, and text-needed.kyrylo jev-browse asks goal, blocked, and next action.jev-mail asks category, priority, and five Nouls — spam is composed in code.jev-windows-voice asks nine-to-eleven questions per transcript update.YT Time Saver asks a usefulness Score and three Nouls.tontoko/jev-browser asks Jev to pick supplied candidates, then asserts in code.jev-builder lets you add as many Noul / Score / Choice questions as you like.jev-frontend-qa asks operation plus target in one request.vmendes jev-shield batches Nouls on feed cards.JevArena asks which legal direction button to click.JevOnly asks done, off-path, and next action.MCode Jev pack asks operation plus target, then a local model types.Otto asks Jev to pick a concrete native action; form fill asks none.jev-turbo asks which candidate and whether the goal is met.ThinkyMiner Winnow asks up to 11 questions on a page, or 7 per feed item.jev-ra asks operation, target, value, prev_ok, and goal_achieved.Barrunto asks nine Nouls on a post or comment.Cerebellum-2B is a local pointer net — it is not a TypeSafe question set.jevcumber asks action, control, literal, and submit.terryds/jevplayground lets you add boolean / choice / score on the Gateway.psearch asks per-page and per-link scores in one batch.KesavanKing/jev-browser asks operation plus target.Mahmoud/jev-browser asks action, element, prev_ok, and goal_achieved.Page Verdict asks nine questions in one request.CorieW/JevTest asks Jev to pick an action, then asserts in code.tylergibbs1/sift asks four Nouls and one depth Score per result.Sloppy Jev’s asks two Nouls: AI prose and ads.jeff speaks Choice / Score / Noul on a local/v1/systemone. NotAlurith/jeff.kevinbadi/jev-voice asks about 15 questions per utterance.Jev Trades asks four Choices, two Nouls, and five Scores.invalidate asks six Nouls, then code rules.jev-mail-classifier asks one Noul per category.ilyamk jevMail asks one Choice over up to 12 labels.tumf/jev-cli asks Choice, Score, or Noul from the PATHjev command.hev/reranker asks up to 30 Nouls in one request.AI Elo Ranker asks two Choices: winner A/B, then a closed reason set.jev_antispam_bot asks seventeen spam questions plus one linkage.jev-grug asks Choice next word, then Nouls for new vocabulary.AI Provider for Jev exposesask_noul /ask_choice /ask_score /evaluate.Mintzs/jevify asks local choice / noul / score — not TypeSafe.SmartSwitch asks one Choice over 10 apps.BTC–Jev Signal asks higher-or-lower questions.Jev-AV asks Choice, Score, and Noul on extracted features.jev-cvss asks one Choice per metric, plus v4 safety Nouls.JEV Document Classification asks category, confidentiality, injection, and subject.jevibe-check asks one question with 2–12 labels.commentcop asks accuracy and usefulness.jev-audio-beeper asks a Boolean and a four-level Score per word.JevTicketRouter asks two Choices, one Score, and two Nouls.DraftPulse asks category plus hook, specificity, reply, share, dwell, and slop.Emoji Jev asks eight questions in one pause.jev-A-share-trader asks Choice per dimension plus a volatility Noul.jev-resume-analyzer asks one batched rubric over the whole CV.Mimicry asks four Scores and two Nouls.pkg-gate judges install-script intent, then code routes allow / warn / block.s1s asks typed judgments over file evidence; answers are paths, not generated explanations.Scam Shield asks eight Nouls, one Score, and one Choice.Transcript Scorecard scores each revealed turn against a weighted card.typesafe-comment scores five comment heuristics.triage-guard batches ticket, alert, and deploy-gate questions in one call each.EmbodiedJev asks staged skill or XYZ-step Choices; default is the rule baseline.mario-jev asks one movement Choice and three jump Nouls.Askable arm picks one primitive from a closed menu.jev-tetris (trungdq88) asks placement, strategy, board health, and next-piece fit.OneVOneJev fans out move, aim, ADS, fire, and jump.jev-doom-agent asks one Choice macro; code owns the controller.JevBall asks one Choice per player from at most 14 local options.gg-friggin-ez screens profanity, toxicity, and severity in one call.Jev Lab (jammaru) asks the next NPC action or a legal shogi reply.jev2048 (erhanmeydan) asks one Choice over four directions.弈瞬 asks one Choice from a shortlist, not every empty cell.jev-play-ping-pong asks serve plus two return Choices.jev2048 (KyleKreuter) asks one Choice over valid moves only.tsai-civ2 asks named Choice vectors; displayed values are action probabilities, not P(victory).TypeSafe Minecraft demo asks one Choice over direct 250 ms primitives.Agent-JEV-Tetris asks one Choice over seven precomputed legal moves.BeatJev asks one Noul: is this spam, a scam, or deceptive marketing.casse-brique-typesafe asks one Choice over 16 paddle zones.Cyber-Breach asks maneuvers, threat Scores, and Nouls.jev-games asks per-title questions; Scripted asks none.jev-gomoku (mizchi) asks one Choice from nearby cells.Jev Snake asks one direction per tick.jev-t-rex-runner asks jump, duck, or keep_running, plus short or full jump.jev-tetris (ML-Nerd) asks one Choice over enumerated placement IDs.P(DOOM) PROTOCOL asks Jev on the p2 slot; Mock is not Jev.pong-jev asks up, down, or stay.ps2-ai-agent asks one Choice over five controller labels.River Oaks asks batched reaction Choices.River Run asks lane, fire, and speed in one request.RoverLab asks one Choice over bounded rover actions.Shady Town asks one structured judgment per round.SIEGE asks allow, block, or escalate.snake-jev asks nine assessments; there is no move Choice.Terrarium asks one Choice over eleven controls plus per-object Scores and Nouls.typesafe-3d-chess is AI vs AI by default.typesafe-chess asks Choice + Score + Noul in one request; it is notmalDuffin.jevcache recalls a cached decision; local/mock is not Jev.jev-me grills a plan as a design-tree interview.Instinct asks hierarchical 2–3 calls over 48 components.TypeSafe AI Playground runs one request over selected questions; Enter without a key browses only.IntelliPrompter asks one Score per open talking point; 1.5 is the author’s.jev-notion asks typed judgments per Notion row; ≥60% is the author’s.Jev System One asks mode, depth, and review Nouls; OpenAI only writes.jev-got asks five questions after Luna writes; inFiction < 0.5 is the author’s.GPT vs JEV asks three support Nouls; presets need no API.guard-jev asks 7 Noul + 1 Score; mock is not Jev.harden-jev-decides asks one Choice over eleven ideas plus per-idea Scores.jev_projects asks one typed decision per pasted text.jev-cloud-quiz asks which cloud a feature belongs to.JEV Language asks bounded semantic judgments; it is not a finished language.jev-pii-checker asks 13 Nouls plus a sensitivity Score.jev-playground (Little-Planet-Labs) mixes Noul, Choice, and Score in one request.ToolGate asks allow, ask_human, or deny.typesafe-arena does not ask Jev; it searches a local docs mirror.CartShield asks a checkout-fraud disposition.The Council asks 36 independent questions in one request.HarnessJudge asks ok, retry, escalate, or stop.HireSignal asks a resume first-pass score.human-compiler asks about thirty Nouls, three Choices, and four Scores; thresholds are opinions.Jev Data Analysis asks typed questions about a CSV; visiting never starts a paid run.jev-ad-preflight does not ask Jev yet — README-only.Jev CSV Workbench asks up to 20 questions per selected row — the author’s.jev-bun1 asks the official billing sample, then extra Nouls.jev-demos asks the next N maze moves as N questions.jev-dev asks eight Score axes on one user line; it is notjev-dev-kit.Jev Paper Judge asks fixed rubric questions the client never sees.jev-should-i-apply andjev-user-jury do not ask Jev yet — README-only.JevPlay lets you invent Choice, Score, and Noul questions.Job Risk Analyzer asks one Score, one Choice, and one Noul.LaneBreak asks ticket priority and routing.McpMatch asks which MCP capability matches a goal.Probably asks buy or wait; it does not predict returns.PulseLane asks a clinic triage.Search-Function-Test is a blog-search experiment; no root README.SpendBrake asks continue, downgrade_model, or stop.SwarmRouter asks which agent should take the task.TrustGate is an indie media T&S gate — Choice / Score / Noul.typesafe-image-diffusion asks 256 parallel Choice questions, one per pixel.razorback16/openjev speaks the same wire types; it is not TypeSafe weights.reflex does the same on Qwen3.5-4B.Jev Capability Atlas maps when those typed questions hold up — not a leaderboard.jev-as-a-judge asks quality and does_pass on frozen agent runs.open-alternative-jev asks Choice letters in-process — not the official SDK.sgoedecke/system-one batches single-token Choice.jev-column-race asks 80 typed questions per request.system-one-open speaks Choice / Score / Noul on Gemma.Jev_apps catalogs those question shapes — it does not call Jev.TypeLLM asks JSON Schema fields on SGLang, not the TypeSafe API.utk2103/jev-studio wraps verify / screen / classify and friends — leftover of the PyPI kit.jevfire scores finite fields on CUDA/vLLM — not TypeSafe weights.jev-rag-benchmark asks batch Noul over hybrid top-20 — not a first-stage retriever.jev.nu builds noul / choice / score in Nushell.jev-lm asks a 229-option Choice as an output layer.claude-jev asks real / reachable / already_handled on review findings.Little Airways asks route / broadcast / response / clearance per plane.cobusgreyling/Jev replays official fan-out questions in a lab.realZachi/jevtest asks one Noul per expectation.mcts-agent asks batched Noul prune, Choice priors, and Score value.jev-codes asks pack questions per hunk.jev-exploration recomputes other people’s questions — it is a ledger, not a TypeSafe card.jev-for-engineers asks FEM / DFM / BOM questions in one call.MongLong0214/jev-gate asks whether to decompose and which tier to run.jev-mode asks one item per batch call, or many questions on one item.jev-oas-sentinel asks bounded questions about changed OpenAPI prose.larguesa/jev-search asks a Noul per line.jev-starter wraps official questions in a policy contract.JevGPT asks “Next word?” over a 20,000-word dictionary (Choice 255 official).ORIGIN-CIVILIZATION asks one Choice per voluntary action — LLM only proposes.siftr asks search / read / pick / filter questions.tenbin asks lint / evaluate / rank / taxonomy at design time; production uses the official SDK.trade-jev asks BUY / SELL / HOLD on NQ L10.ZeroSweep asks inbox-triage questions on procedural emails.DecisionBridge asks one question per existing-LLM call — not a Jev reproduction.jev-agent-failure-benchmark asks three Choice questions per failed trace (who / when / what).carryforward asks one yes/no per ledger entry on recall.hemanth/jev-chess asks intent, sharpness, theme, and king risk.jev-freeform asks the next character (98 options) then complete / incomplete / invalid.jev-labs asks five paraphrased pharmacy questions and gates the margin.jev-reward-model-evaluation asks pairwise and process-verify questions across eight tracks.jev-routing-experiment asks difficulty or per-model quality for a frozen pool.jev-synergy-screening asks include / exclude plus eligibility Nouls.jevchess asks one Choice over legal moves and one Score evaluation.kyotsu-ai-bench asks one Choice per Common Test blank.openjev-experiments asks NLI labels on a local checkpoint — not TypeSafe Jev.RISC-jeV asks 22 Boolean gate combinations.typesafe-vs-deepseek asks invoice / email / rerank questions beside DeepSeek.jev-alpha-bench asks bullish/bearish plus already-priced.jev-anotacao-sentencas asks 12 typed variables per TJSP sentence.jev-deferred-crispification does not call Jev — it audits hop-ECE vs TCE / AMS.jev-headline-bench asks which of two headlines won the A/B test.jev-jp-address asks Choice over leftover Ken_ALL candidates.jev-pick-and-place-study asks one bounded Fetch action every 0.2 s.jev-report asks 50 Chinese Choice / Score / Noul items.jev-shadcn-lint-eval asks whether a lint finding is right and how helpful the message is.jev-trace-classifier asks one Noul: was this wiki page written by an agent?PocketJev asks local Qwen3-VL option logits — not TypeSafe Jev.shade-arena-jev-monitor asks whether a tool call is sabotage.system-one-adapter-rust asks an LLM the same Choice / Score / Noul shapes.thaiexam-jev-charts asks one Choice per ThaiExam item.Typesafe_chess_eval asks win/draw/lose plus one UCI move.typesafe-oracles asks whether a commit message matches its diff.jev-skill asks Choice / Score / Noul from the installed scenario; Mode B is not a Jev call.fable-jev asks a shortlist Choice then a fitness Noul before the harness commits.LiteLLM JEV classifier asks one Auto Router Choice over authored tiers; the 95.00% match rate is LiteLLM’s on those labels, not general accuracy.jev-robotics-demo asks Jev which small move, grasp, and whether the stack is done. NotRoboJEV.jev-physical-ai asks escalate Noul, team Choice, and urgency Score in one call.Jev-sample compares one 4-option Choice to four precondition Nouls.nassim-arifette/jevgrep scores eligible fragments for a behaviour question.savka777/jev-use asks one Choice over numbered Accessibility targets.jev-gateway asks which tool fits, then fail-opens if Jev is unsure.SystemOneHarness compiles finite actions into typed questions each step.pZacca/askjev infers noul / score / choice from free text — you cannot force the type.philippdubach/pi-jev-router asks five classification questions, then picks the Pareto knee.davila7/jev-explained walks Noul / Choice / Score on four live examples.jcr asks Jev which catalog child to open at each node.jevvy asks four shipped Nouls before auto-approving a shell.oxlint-plugin-jev asks one Noul per match: yes means report.tiab-review-plugin asks one include Noul at default threshold 0.3.Pi Jev Guard asks four action Nouls in one request.jev-ts-repl lets you shape Noul / Choice / Score before you write code.w3cj/jev-chat asks one Choice over tools plus Nouls for missing details.tripwire asks seven checks in one call.triagedy asks five typed questions per alert.jev-rules asks one Noul per rule.compact-adviser asks two Nouls, then code composes a /compact score.jev-pong asks one Choice overup /down /stay.lurk asks Jev to judge titles, posts, and comments; the 0–100 feed score is product-side.RoboJEV asks intent, then XYZ + gripper.doc-router asks whether a page needs OCR.live-jev (Ableton) asks Jev to place a sentence on an allow-list of Live commands.call-coach-ai asks next-action Choice + buying-stage Score + 7 Nouls in one request.abide asks one question per rule on the diff.macbrow asks one Choice over tools plus speculative argument Choices.jevscan-evm asks 2 general + 14 category + 343 detector questions.crush-monitor asks emotion and intent Choices per line; the affection total is product-side.Alurith/jeff asks four read-only semantic rules.ST-jeved lets you write Score / Choice / Noul sensors.pi-jev-router (sugarforever) asks route / on_topic / needs_history / one_off before the prompt lands.ai-backends exposes/api/evaluatefor Choice / Score / Noul.vexjoy-agent /dasks Jev to classify then gate the restated intent.sharp asks a classifier whether a post should stay.orchestkit asks one category Choice, or nineteen route classes plus four side judgments.pi-fabric asks four Nouls in one auto-approval request.agent-dispatcher asks which of 27 roles, then skills and tools, only when a scope is enabled.jev-social asks the next socai CLI operation from a changing allow-list.hermes-web-search-plus asks Jev to confirm anewssearch type.JevSeek / JMP asks Jev which action to run next.WrongStack asks Choice / Score / Noul on brain, memory, topic, and lint seams.jevkeep asks which excerpts still matter before compaction.open-jev-approvals asks risk, authorization, hazards, and scope on each tool call.forma-system1-experiment asks about 250 layout / block / theme questions in one request.jev-tool-router asks one Choice over tool names plusnone_of_the_above.jev-stage asks one Choice over 14 performance classes.jev-browser-pilot asks one Choice over ids your code built.newsscore asks Score + relevance Noul + event Choice + novelty Noul per article.jevlens asks political / spam / adult-promo labels on a post.beat-jev asks a typed shot or save zone.filedcom/playjev asks check / choose / rate / act over numbered nodes.decide-mcp asks one Choice over 2–64 ids.JEVRACE asks three Choice questions per driver call.tenet asks one probability per ≤254-line window.trading-bot-jev asks asset / sentiment / materiality / regulatory risk — not a price.github-star-organizer-jev asks one Choice over existing Lists plusNo matching category.jev-flash-router asks noul / choice / score viaevaluate_decision.picaye/jev-compaction asks two Nouls per tool call.typesafe-jev-examples asks seven questions on one ticket, or one Noul per article.kiarina/labs asks department Choice + urgent Noul + frustration Score, or one Noul per moderation category.jev-agent-router picks one agent or abstains.pi-fast-jev-compaction (QuentinDanblon) asks two Nouls per old tool call.jev-review-action asks one Noul per source candidate, then checks plus one Choice.jevframe packs every question for one row on the same request.jevcache (kushals256) asks same-intent, not cosine similarity.jevcheck replays fixture Choice / Noul / Score.swift-typesafe uses@QuestionSet— pass the type, do not construct an instance.cursor-clijev-compaction asks two Nouls per old tool call.dsh-jev-tools exposesjev_askandjev_gate.jev-guard (Clemens) asks workspace Noul + destructive Score 0–3 + violation Choice.how-sure-is-jev reads a finished probability vector; it does not ask Jev.ProgressGate asks six atomic signals; the policy picks HALT.jev-issue-radar asks four Choice questions per pair.leanest asks whether a change could affect one test.jevtriage asks one Choice: ready / needs_review / risky.cua uses Jev only for stuck detection.agentic-harness-cli maps boolean → noul, choice → choice, ordinal → score.jevassert requiresunknownon every closed set.jev-packs ships nine SPEC v0 packs.jev-mcp-spring exposes classify / score / check / gate.moongate asks one Choice per rule.req_llm names noul:boolean.git-judge-jev asks a fixed hunk pack; gates never see the PR description.typesafe-screening-mcp asks match Noul plus relevance Score 0–2.jev_vampire_survivors asks character, stage, level-up, and walk direction.dspy-jev mapsbool → noul andLiteral → choice.slop-filter asks about 18 narrow Noul/Score tells.diff-risk-sentinel asks four semantic Nouls per touched function.jev-browser-qa asks expect Noul plus click/goal Choice.jev-outlook-organizer asks three Choice plus two Noul.survey-qc-with-jev asks open-end Choice, coherence Noul, and attentiveness Score.dsh-jev-guard asks one irreversible-delete Noul.work-with-jev asks related / action / urgent / worth-reading.rocky asks thirteen router questions in one request.jevseo asks about 25 questions, then a leave/improve/rewrite Choice.lgtm asks whether a test actually tests the implementation.jev-adrank asks tap and buy Nouls for every eligible ad in one call.jev-subagent-router asks one Choice over model+effort combos.jev-board-game asks a Choice over the other players.semantic-live-caption asks key Noul plus emotion/intent Choice.approval-judge-bridge asks one approve/deny/escalate Choice.whoa asks five digest questions only after a counter fires.jev-debtgate asks a fixed debt-kind set over a local ast-grep fact pack.rspamd-jev asks a ham / spam / phishing Choice that never changes the score.jevdevice asks one action Choice over the live tree, then a safety Noul.whatitdid asks a parallel label set and never shows the agent’s words to that pass.JEV-TheFishGame asks action + target Choice and a panic Noul per fish.jev-android asks several questions in one Jev request over the current controls.jev_playground asks the same pair with and without measured state to show framing flip.jev-typesafe-real-financial-use-cases asks the official support-ticket trio plus trade Choices over candles.rag-jev asks a rerank Choice and a support Noul.dynamic-context-engine asks one Noul per paragraph.jev-crawlers asks verdict + risk Score 0–3 per node.jeffrey asks tool Choice, progress Score, risk Score, and stuck Noul each step.jevdash asks a macro Choice, a jump Noul, and a danger Score 1–10.InfamousCube/JevPilot asks one action Choice over the live list.public-browser asks a ref Choice plus an already-done Noul.clash-jev asks strategy, card, and square as three Choices.jev-voice-control asks one command Choice over the config menu.jevbro asks one action Choice over the current controls.jev-leftpad asks one Choice space_0–space_10.actionreflex asks many policies in one request.codex-cua-jev asks an AX-tree element Choice plus a risk Score.jev-ads-analysis asks a shadow action Choice after the rule layer.jev-realtime-sdk asks a verb Choice on each decision tick.kaustav1996/reflex asks gate, monitor, and router questions in parallel.tegersdorfer-collab/jevkit fans many questions into one request, then bands.Jev-Checkpoint asks one route Choice.discoprint asks theme, mood, complexity, explicit, and first-person per song.JevRAG asks a usefulness Score, an action Choice, and a condition Noul.prasanthj/duckdb-jev asks Noul, Choice, and Score from SQL.jevbus asks one question per subscription.loom asks archetype, theme, block, and social-proof Choices.comment-opportunity-radar asks intent Choice, purchase Score 0–4, and urgency Noul.jev-multimodal asks Noul and Choice over compact evidence — Score is not implemented.jev-extract asks sequential token start then end Choice.anuran-de/tripwire judges a rolling stream window.jev-factorio-agent asks goal, action, and stuck questions.is-jrecord asks one Noul.jev-ai-skill asks keep/drop per tool call.jev-table asks every pack question in one call per unique row.nf-jev exposes noul / choice / score / jev as Nextflow functions.jevi forwards the question JSON as written, including key order.jev-in-mcp asks Jev to pick the next tool.ririumu/jev-demo shows four recipes as JSON.spring-ai-typesafe packs Noul / Choice / Score in one Java call.n8n-nodes-jev-classification exposes Classify / Score / Check / Ask Questions.jev-wide splits a Choice when candidates exceed 255.coffee-under-fire asks one Choice per live NPC action.sudoku-jev asks Jev among legal placements.model-switch asks reuse vs refresh.jevfeed asks one ranking Choice per batch.jev-newbie walks ask / pick / rate.jevtown asks a crowd of personas.jev-downloads-sorter asks one folder Choice per file.neurolink exposes boolean / choice / score asdecide().home-assistant-typesafe fans out Choice / Noul / Score over devices.x-reply-filter asks four Nouls per leftover reply.tidy judges each channel.DataJev asks CONTINUE / SWITCH / VERIFY / STOP.pi-Jev-browser asks one Choice over page actions.jevmory grades durability and support.jev-browser-wingman asks one step on the open page.svitaTLCO/jev-skill gates a local SLM swarm.jev-divination-lab normalizes two fictional methods.jev-complete.nvim ranks buffer words.hearth-jev-rental-search picks browser steps on rental pages.cartpole-jev asks the same five questions every frame.sf-early-learning-mcp scores licensing and budget.jev-wrapped asks four questions per Telegram post.jev-skill-router names at most one skill.slither-jev picks a tactical route.Jev-LLM-Playground asks department, urgent, and frustration.jfind asks one Noul per file.jev-grand-prix asks line, pedal, and trouble.jev-2048 asks one Choice over four directions.pi-jev-effort scores prompt hardness.feed-rubric asks one Noul per category.JohnsonRan/pi-jev asks four safety Nouls.playwithjev asks one Choice over legal moves.Jev-MongoDB asks topic, sentiment, urgency, and safety.jev-torneo-animales asks which animal wins.jev-hierarchy asks one Choice per tree node.hfjev adapts Choice, Score, and Noul.jev-hanko asks 41 clause types at once.jev-agent-hooks asks skill fit then model tier.zm2231/skill-router scores then reranks skills.clay-jev-people-ranker asks founder class and Noul.jev-linter-action asks yes/no per file set.ask-jev-series asks one Noul per title.opencode-jev-compaction asks keep-call and keep-result.typesafe-curl shows raw Noul, Choice, and Score curls.jury.nvim asks which fix to show.hermes-slash-router asks which slash command.the-claw weighs prizes in a pit.yuru-poll asks one Choice over poll options.poker-theory-knowledge-base asks five poker Choice / Score / Noul scenarios.jev-craft asks after reflex, before a slow LLM.typesafe-decision-skill does not run System One.jev-wechat-live asks reply class on 11 messages.jev-study mocks a DecisionClient.pi-ext-model-router asks which model for 120 prompts.xinwang-nwpu/jev-mobile asks operation and target in one call.dsh-jev-verify asks Choice / Score / Noul then verifies live.ego-jev asks which DOM step.fewhnhouse/jev-review-action asks Noul then impact 0–3.jevmoji asks a 0–3 Score per emoji group.jev-mailroom asks 11 questions per email.okooo5km/jev asks yes / pick / score from the shell.jev-toolkit asks viatypesafe_ask.Jev-Examiner asks atomic violation dimensions.yibie/pi-jev-browser asks one Choice per browser step.scam-checker-with-jev asks 22 questions per SMS.mcp_typesafe asks 12 MCP tools.jev-fetch asks keep or skip per section.jevis asks the next Flutter action.jev-folio-recursive-classifier asks a Choice at each FOLIO level.typesafe-scheduler-diagnostics asks why a Pod is Pending.Tech-Byte-Frontier/jevgate asks three maintainability verdicts.ibmi-jev asks Noul / Choice / Score in SQL.ActionJev asks screen / locate / judge questions.typesafe-ai-jev-example asks six demo patterns.jev-town asks one judgment per resident.auto-jev-codex-for-paseo asks the task category.cloudforge-jev asks three Nouls plus depth.sava-software/typesafe-client asks typed Java questions.traderai asks thesis / pulse / mover questions.imsukhe/jev asks keep or classify effort.obsidian-auto-tagger-jev asks one tag Choice plus None.gut asks likely / classify / rate.totally-tim/jev-gate asks review-rule observations.sahajamit/jev-lens asks four questions per post.jev-evaluation asks nine pre-registered experiments.ajayk/jev-go-sdk asks typed Go questions.jevrouter-prompt-tier-extension asks one capability-tier Choice.k4its1t/jevlens asks YAML questions per record.jev-agent-eval-example asks grounding / tool-use / usefulness.inbox_classification asks one inbox tag Choice.jev-acento asks Spanish XNLI / Noul pairs.hermes-and-jev-play-minecraft asks one Mineflayer action.jev-harness-router asks tools + skill in one call.x-bookmark-sorter asks up to 20 folder Choices.jev_dart asks Choice / Score / Noul from Dart.dsh-jev-decide asks jev_decide (noul / choice / score).gualican/jev-model-router asks tier + stakes.jev-algorithms asks pairwise comparisons.typesafe_agent_gates asks four Noul shell gates.themsquared/jev-benchmark asks one tool-call risk Choice.jev-state asks the next state plus stay.doceval asks ten Score dimensions.pilot-typesafeai-jev asks intent + urgency + injection.system1-system2 asks one issue label Choice.jev-crypto-scout asks news sentiment / catalyst / confirmed.jevmod asks one Noul per category.jev4j asks typed Java handles.jevlang asksx ~ question.jev-policy-engine asks YAML Noul / Choice / Score.jev-net asks typed .NET handles.typesafe-sdk-ruby (afurm) asks Ruby Choice / Score / Noul.Mrribvar/hermes-jev-plugin asks onejev_evaluate.dsh-plugin-jev asksjev_decide.cursor-jev asks one hook Noul.jev-table-tennis asks paddle centre y.jev-measured asks eight fixture packs.jev-injection-bench asks one injection Noul.profanity-checker asks one or two Noul.zsh-jev-suggest asks next_command + intent_fits.jeveryword asks word-id Choice.dsh-jev-hook asks route / prune / retry.jev-verifier-skill asks one advisory Score.system-one-playground asks SOS judgments.demo-symfony-typesafe asks intent + urgency + bug.jev-the-spire asks one legal action.jev-flappy-bird asks FLAP or WAIT.openrouter-jev-mcp asks jev_check / jev_classify / jev_score.claude-code-jev-compaction asks still needed.jevstudio asks per funnel node.zurfyx/jev-browser-skill asks operation + targets.playwright-jev asks one actionId.dsh-memory-jev asks worth_keeping + supersedes.jevSweeper asks one Noul per cell.jevqa asks the next action + oracles.jevlergy asks 29 allergen stages.explore-typesafe-ai asks NEWS2 / meds / inbox / notes.jev-fastloop mixes noul / choice / score.jev-feels asks feels? / decide / score.prompt-demux asks same_task + tier.grok-jev-router asks lane / risk / reversible.jev-judge-mcp asks judge.jev-rust-review asks triage + verify.jev-eyes asks decorative / kind / needs_vision.jev-form-filler asks which profile field.jev-source-evaluator asks publisher + same-meaning.Jev-Mem asks write + retrieve.jevfanity asks one Noul per category.tempo-jev-demo asks the view plan.typesafe-pi asks typesafe_judge.typesafe-jev-gate asks allow / deny / approval.quick-decision-assistant asks Choice / Noul / Score.jev-audit asks where Jev fits.mcp-server-jev asks jev_evaluate.typesafe-jev-mcp asks evaluate.Jev-chooses-a-LLM asks COST / BALANCED / INTELLIGENCE.obsidian-jev-archiver asks which folder.Jevsus asks true / false.priorbench/jev asks 21 experiments.jev-vs-llm asks 77 intents.lkclean asks bait / promo / interest.jev-writer asks a writing rubric.fotocopiatrice asks same result.jev-reality-check asks count / date / arithmetic.Running-Dolphins/jev-bench asks twelve public tasks.jev-rag-benchmark asks candidate relevance.graphql-classifier asks PII / auth-gap.jev-lateral-thinking-quiz asks yes / no / irrelevant.jev-cite-check asks 지지 / 모순 / 무관.PageLens asks Score 0–3.adaptmypage asks six intent questions.jev-market-reflex asks BUY / SELL / HOLD.HA-SystemOne asks jev.noul / choice / score / ask.realbogart/jev asks Applicative triage.Postmark asks tone / virality / cringe.jev-kit asks PR triage.stil-lint asks Norwegian style.typesafe-jev-ruby asks Choice / Score / Noul.virolea/jev asks ask / choose / score.fizzbuzz-jev asks Fizz / Buzz / FizzBuzz.jev-categorization-test asks 15 categories.boldbug1/jev-triage asks category / urgency / frustrated.tidy asks which folder.zammad-jev-dispatcher asks which Zammad group.jev-calibration-audit asks bilingual probes.Jev-is-odd asks odd or even.jev-access-day asks triage / checkout / dunning / PR review.jevscope asks named project questions.longjev asks over kept chunks.jev-systemone asks Choice / Score / Noul.jevskill asks which log lines matter.jev-spam-lab asks legitimate vs phishing.Jev-Rug-Checker asks narrative / risk / exit.jev-voice asks six voice axes.jev-atlas maps Choice / Score / Noul fit.jev-classification-guide asks company / title.everything-about-jev asks ticket / claim / route.jev-technical-term asks jargon vs everyday.judgement asks Choice / Score / Noul.jevymarket asks resolves_yes / answerable / clarity.pi-jev-model-router asks task_kind / complexity / capability / needs_deep_reasoning.jevsome-projects does not ask Jev — it catalogs who does.jev-usecases sketches vendor vs reported questions.typesafe-experiments asks through Claude Code / Agy.LocalSafeKidsEngine-Open asks only when the local SLM is below 0.88.openclaw-plugin-typesafe-ai asks intervene / risk / prune.openclaw-typesafe-ai asks only what the agent supplies.brain-speeddial-eeg-jev asks whether the action makes sense now.jev-decision-gateway asks tool / continue / verify / tier.jev-workbench (Chunky83) asks over supplied evidence.jev-vphone-cli asks next tap and whether the goal is met.omni-scraper-mcp asks over normalized documents.openclaw-jev-compaction asks keep-call / keep-result.tab-bouncer asks useful + kind per tab.jev-linkedin asks fit / outcome / gap / worth applying.jev-cyrillic-audit asks one Choice per paired item.AnchorLint asks promise vs destination.jev-graph-walk asks one Score per edge.jev-agent-toolkit asks Choice / Score / Noul through a skill.jevoracle asks whatever you paste.jev-preflop-poker asks fold / call / raise.claude-code-jev-guardrails asks 5–8 Nouls per hook.jev-field-tests asks twelve scenario packs.Linkedin-Slop-Detector asks is_slop + boilerplate.mailverdict asks phishing / credentials / attack type.opencode-intent-gate asks four Nouls before tools.jev-codex-bridge asks model + effort.omp-jev-extensions asks accepted / rejected.RCIL asks how to frame the next question.opencode-context-pruner asks keep-call / keep-result.jev-reranking asks one Noul per passage.omp-jev-web asks operation + target.umby-jev-stack asks six independent Nouls.jev-enterprise-decision-fabric asks a versioned pack.cartas-ciberseguridad-jev asks NICE / ECSF / SFIA.typesafe-computer-use-windows asks operation + target.pi-jev-skill-picker asks one Score per skill.LiveSignal asks 26 judgments per comment.jev-lab (Pasblinn) asks model + health.jev-study asks Chinese wording + game loops.jev-1.13-mini-benchmark asks noul / choice / score on 34 cases.pi-jev-session-router asks model + thinking.jev-tab-grouper asks one Choice per tab.jev-domain-auctions asks brand + commercial + cleanliness.maestro-jev asks goal + next tap.jev-master asks queue / urgency / frustration.jev-playground-openrouter asks any preset.pi-jev-skill-bench asks which skill.claude-jev-warden asks exquisite + verdict.jev-test (clduab11) asks search / keep / support.battleship-vs-jev asks the next cell.skillpick asks gate + shortlist.minesweeper-jev asks the next guess.omp-greenlight asks human / severity / scope.commitjev asks eight Nouls + a headline.pi-jev-approver asks risk 0–2.omp-jev-watchdog asks contradiction.pi-jev-skill-suggestion asks gate + skill.Metis asks category + missing details.jev-feed-filter asks filter probability.jev-router-playground asks one Noul per candidate.jev-x-blocker asks five narrow Nouls.FirstScreen asks six first- screen questions.jev-tic-tac-toe asks the best cell.jobbyjev asks fit + would_interview.cc-mod-jev asks keep-call + keep-result.jev-pacman asks one legal direction.astra-jev-new-vegas asks a typed action.jev-corrective-rag asks triage + grade + verify (stubbed).jev-plays-tetris asks the best placement.jev-fanout asks six probes plus fillers.jev-lint (vayun) asks contradiction + stale.HighwayEnv-But-Jev-is-Driving asks manoeuvre + pedals.rag-jev (Emre) asks one Noul per passage.typesafe-tetris asks one of twelve landings.jev-hands asks which element.FocusTube asks learning usefulness.jev-labeler-action asks one Noul per label.jev-prompt-optimization asks one Choice.prompt2jev asks the converted set.jev-vs-luna-phishing asks one phishing Noul.PlotVeil asks one spoiler Noul.awesome-jev (AgentBuff) lists catalogs, not questions.jev-screen-mcp asks spam / toxic / category.jev-review-mcp asks a patch verdict.jev-skill (abhishek) rewrites a question.mcp-server-jev (codeChap) exposes noul / choice / score.jev-decision-skill sends a payload.jev-arcade asks one legal Choice.JevSharp asks Choice / Score / Noul.jev-dev-harness asks file relevance.vibecheck (psyb0t) has no live questions yet.agentic-rl teaches when not to train.Jevons asks skill relevance.hippo-memory asks a rerank.jev-clinic-triage-eval asks four clinic questions.jevgo (devbackend) asks Choice / Score / Noul.askgrep asks one score per chunk.jev-symfony-bundle asks a validator.astrbot typesafe autoreply asks whether to reply.evoke asks which reflex.awesome-jev-apps lists questions other apps ask.local-jev (amithgc) asks Choice / Score / Noul locally.AnyDecisionModel asks Choice / Score / Noul.decision-backend-http asks a protocol question.jev-arm-lab asks intent / grasp / progress.jev-zork asks one valid command.jev-doombot asks action / danger / shoot.jev-telegram-channel-router asks a filter.jev-table-import-mapper asks one Noul per pair.jev-is-you asks one verified action.Jev-Flywheel asks sentiment and topic.qqbot (Mist-wu) asks whether to reply.commentlab asks an attitude.jev-as-a-judge (sjyangkevin) asks Noul / Score.jev-intent-review asks whether intent is met.lus222-jev-challenge asks a mission question.jev-hub (antTing) forwards Choice / Score / Noul.jevnav asks which element.node-red-contrib-jev-decisions asks a pack question.jev-perfume-advisor asks one Noul per scent.awesome-jev (BeatAPI) lists questions other apps ask.jev-mcp (Freepik) asks Choice / Score / Noul.jevernetes asks whether a log matters.Jev.UI asks a workbench question.jev-calibrate asks the labelled question.jev-skill-selection asks keep or drop.jev-decision-lab asks a Studio pattern.jev-doom (olivier-motium) asks which motor.ormus-jev-cua-linux asks which legal action.jevable-code drafts a contract.japanese-jev-lint asks typo / twist / repeat.herdr-jev-router asks which harness.jev-router-dashboard shows a past Choice.jev-chat-JARVIS asks intent and danger.jev-benchmark (arifulislamat) asks four ticket questions.jev-spire-brain asks seven Spire decisions.jev-lamplight-house asks mood / welcome / cat / nudge.dsh-jev-adapter asks Choice / Score / boolean.jev-bench (cocodedk) asks a classifier label.jev-esp32s3-gateway asks which bounded command.jevagent (Lilithys) asks which skill.Jev_model asks how hard this turn is.jevish asks a label.jev-agent-control asks Choice / Noul / Score.playground-Jev asks six presets.awesome-jev (jtnkminimal) lists questions other apps ask.awesome-jev (robokrunch) lists measured questions.Awesome-jev-use (APA) lists persona questions.decio asks Choice / Noul / Score.clue asks a relevance Score.Maybe asks yes/no or a match.jev-doc-classification asks one Noul per type.worth-replying asks five tweet questions.jev-team-calendar asks whether a resume shows a skill.jev-cli (gmaxxxie) asks Noul / Choice / Score.naiou asks one Noul.jev-news-cold-start asks appeal / broad / clickbait / emotional / useful.jev-support-pulse asks topic and widespread.jev-code-review-benchmark asks four rules.chakravyuha-jev asks one Choice per cell.jev-issue-pulse asks kind / frustration / blocking.jev-eval asks a public-set label.jev-gamebenchmark asks a Score per drop.jmarket asks Noul or Choice.your-signal asks relevance scores.jrx asks risk dimensions.jev-approval asks allow or deny.questions asks Zod or native batches.sawyer-plugins asks project / machine / model.prism asks many ids in one jev() call.argus asks a template or five yes/no claims.system-one-security-triage asks one five-way Choice.hermes-jev-north-star asks met / progress / weakest.slopcheck asks 15 nouls.JevBird asks a path plus danger.jev-tweet-radar asks worth engaging plus tags.pr-sieve asks .jev.yml rules.JevFind asks path then window scores.spend-gate asks fit and value.jgrep asks 16 nouls per batch.jevfill asks which note matches a field.paper-radar-jev asks P(True) relevance.ego-jev asks an opaque operation.jev-context asks Yes or No per passage.foreman-jev asks completion plus the acceptance command.azdaja asks judge_many / jev batch.delegador asks dimension / complexity / volume.jev-security-scan asks suspicious behavior on redacted snippets.jev-auto-approve asks ready / tests / human review.mk-jev-fly-brain asks one of seven moves.jevscan asks ape / watch / avoid plus four scores.tokengate asks five 0–1 scores on the buffer.typesafe-go asks Noul / Choice / Score handles.typesafe-ai-ruby asks department / frustration / urgency.jev-classifier asks the next tool plus done.jev-research asks route / risk / readiness.typesafe-sdk-swift asks category / urgent / priority.codex-jev-compaction asks whether a tool pair stays.pi-jev-compaction asks keep on each pair.fast-jev-compaction-pi asks keepCall / keepResult.elons-job asks suggestive replies.bluenoise asks unmatched replies.jev-bridge asks, then caches.jev-snake asks move / strategy / danger / escape / food.TrashCompact asks excerpt scores.ruleworld asks 24 Nouls for physics.jevkit-js never asks Jev.awesome-jev lists projects.jev-shortlist reuses cached rankings.duckdb-jev asks per row.typesafe-go asks category / urgent / frustration.jev-pilot asks next_action.jev-harness specs seams, not a runtime.polar_llama asks one row, many questions.jev-mcp asks evaluate.jevkit asks activity / useful / severity.jev-sentinel asks intent and risk.awesome-jev-use-cases lists demos.Jevest asks surface questions, not defects.opencode-langsearch asks whether a hit stays.jev-xiangqi scores each legal move.jev-browser-use asks action and done.goodall asks department and urgency.jev-creature-forge asks family and trait.SpecPi asks the advisor layer.oh-my-pi names typesafe as a judge provider.jev-screening-benchmark asks include / exclude / uncertain.jev-oasis asks lurk vs generate.jev-minesweeper asks one Noul per hidden cell.jev-webui asks a batch of questions.jev-radar asks intent / emotion / risk.awesome-jev lists usage.jevgrep asks one Noul per line.clash-jev asks strategy, card, and square.jevriel asks a Flight Test.Jev-CustomerService-Demo asks queue, urgency, mood, and escalation.jev-gate asks five PR questions.Footwork asks operation and six yes/no checks.pytest-jev asks holds and lacks.jev-foundation-models maps @Generable fields.jev-pre-compaction asks should_reduce and retain.arbitype asks evaluate tools.tabjev asks one Choice per tab.jev-skill-scout asks rank then verify.ghost-route asks which route.lightjev asks hit, score, and next link.Jev-Cognigy-QA-Suite asks a rubric library.dsh-fact-check asks three Nouls.jev-computer asks four NAND cases.jevkit asks check, choose, and guard.awesome-jev indexes catalogs.zjev asks Choice, Score, and Noul.typesafe-laravel asks intent, score, and noul.jev-starter asks three typed questions.jev-docs does not ask Jev.SystemOneDotNet asks Choice, Score, and Noul.dsh-jev asks evaluate, skill, and tool.guideme asks noul, choose, and score.jevselector asks one Noul per skill.TypeSafe.Jev asks enum Choice, Noul, and Score.jev-go asks department, frustration, and urgent.typesafe-php asks billing, team, and urgency.jev asks Noul, Choice, and Score.jev4jobhunter asks worth opening and apply.jev-connect-four asks column and intent.typesafe-laravel asks refund, team, and urgency.TypeSafe-as-a-Judge asks route, rank, and verify.jev-slot-machine asks bet, pace, and action.JAST asks candidate and support.tinystruct-typesafe asks action and arguments.web_attack_detection_jev asks attack, block, and review.pi-jev-multi-provider asks compact and gate.jev asks Choice, Score, and Noul.jev-triage asks flake and owner.jev-architect asks whether a step needs Jev.jev-usecases asks department, urgency, and contain.gosys1 asks billing, tone, and urgency.jev-cli asks noul, choice, and score.Slop Mop asks tells and usefulness.Jev Voice Browser asks intent, target, and complete.Jev.Sdk asks urgent, department, and frustration.jev_blog asks triage and skill.jev-mcp asks Choice, Score, and Noul.n8n-nodes-jev asks department and urgent.let-jev-speak asks the next word.DroidJev asks which element.Jev demos asks match and movie.jev-agent-browser asks operation and complete.jev-mail asks type and needs action.pi-jev-router asks tier and consequence.Sysone asks check, classify, and score.Jeff asks safety and match.Jev vs Laya asks the same eight tasks.jev-awesome-skills asks pick, determined, and harm.overseer-judge asks session, task, and review.typesafe-mcp asks check, classify, and score.Laya vs Jev asks steer and sprint.typesafe-rust-sdk asks urgent, department, and frustration.mac-voice asks action and app.Jeview proxies whatever you send.ai-type-safe-platform asks an Evaluation.jev-mail asks action, important, and bucket.jev-mcp asks evaluate.İlan Eleme asks 22 questions in one call.awesome-jev-agent-skills asks review and evidence.jev-skill asks intent, rank, and safety.jev-harness asks a policy pack.jev-axi asks choice, noul, and score.usher-point asks target and skills.jev-patent-disclosure asks completeness and type.jev-figure-router asks a visual branch.dual-process-ai asks route and risk.awesome-jev asks department, frustration, and urgency.jevrb asks noul, choice, and score.docfriction asks friction type, severity, and prerequisites.openclaw-jev-router asks a tool family.no-hallucination asks support and citation.opnroute-jev asks noul, choice, and score.jev-trader asks BUY, WAIT, or SELL.jev-lab asks sentiment.Seek asks kind, time, and fit.jev-tradingview-signal asks direction and strength.r2r-jev asks noul hazards.Elmridge asks action and content.doom-war asks a macro.remember-me asks hydrate and admit.judgevet asks noul, choice, and score.tink-route asks specialist then skill.jev-router (hyspacex) asks a task assessment.Model Picker asks a model/effort pair.jev-verdict asks a gate judgment.Readwise classifier asks 23 tags then a verdict.JEV Smart Router asks a named option.node-red-contrib-jev asks noul, choice, and score.Jev Stories asks winner and filters.Scribe Verify asks support.hunch asks operation and element.web-search asks source and recency.blockly-jev asks noul, choice, and score.jev-crash-course asks the three primitives.jev-trade-cc asks entry and exit.Jev Voice asks an action.omp-classifier asks noul, choice, and score.jev-shadow asks the same question you already answered.Feedlens asks one label.Jev IDS asks is_attack and category.jev-sdk-java asks noul, choice, and score.moviejev asks fit, reason, and exclusion.jev-mcp-server asks verify, classify, screen, decide, and extract.horse-racing-ai-jev asks which horse wins.Jev the Janitor asks bucket, persist, and secret.jev-certify asks intent and scope.Jevflix asks constraints and one film.Smart Paste asks a field and a resume gate.ApplyGuard asks five safety checks.spec-ptc-jev asks whether to start early.jev-judge asks noul, choice, and score.jev-cli asks the same three.jev-classification-prompting asks with rewritten criteria.Microsoft.Extensions.AI.Jev asks from a schema.jev-safety-benchmark asks typed safety questions.jev-codex-router asks model and effort.RFxCheck asks seven procurement checks.jev_codex asks a work category.Yapp asks an action.typesafe-jev-opencode asks noul, choice, and score.DocJev asks a category and a boundary.Codex Sift asks hardness and stakes.jev-triage-cn asks a Chinese label.jevals-data asks the same typed board.dsh-typesafe-ask asks choice, score, and noul.dsh-auto-review-jev asks several risk noul.Commander Oracle asks four card questions.Book of Answers asks a choice.typesafe-jev-tools asks a band.jfrader/jev-mcp asks four MCP tools.openjev-mcp asks through OpenJEV.Ask JEV asks a plain-English question.jevyoumean asks a subcommand choice.jev-oncall asks four triage questions.use-jev asks noul, choice, and score.cpa-plugin-jev asks through CPA.jev-signals asks seven market questions.jev-go asks from Go.jev-mcp-server asks choice, score, noul, and classify.jev-gtm asks 50 prospecting questions.Jev DecisionOps asks through a gateway.jev-cli asks noul with 0.5.maibot-jev-timing-gate asks whether to speak.Jev MAGI asks three independent Nouls.Jev-Integrate asks whether to call a writer.jev-skill-router asks which skills fit.jevbrief asks one question on a brief.jev-judge-mcp asks ten judgment tools.jevskillz asks several yes/no phrasings.jev-claw asks task type, complexity, and risk.Jev Inbox asks critical, urgency, and labels.Bagel Review Board asks eight policy questions.AI Decision Engine asks one choice and six nouls.S1Rank asks one noul per document.Barq asks which click.hermes-jev-tool-gate asks authorized and risk.jev-scheduling-evaluator asks three scores and two nouls.jev-playground asks what the box is.claude-herdr-jev-compaction asks done and waiting.jevmem asks whether to save a line.Cube Rule Oracle asks where the starch sits.NaturalCodz asks is, pick, and rate.SwiftJev asks typed Swift questions.tone-your-mind asks how close a rewrite is.mind-your-tone asks how close a phrase is.typesafeai-cli asks whether to exit.Relay asks whether to act.jevkit asks typed Rust questions.Decision System Forge asks legal actions.ghtriage asks type, area, and severity.jev-hearth asks stop or add a word.OpenClaw catalog router asks which tool.WestAI OpenClaw Jev asks yes, pick, and rate.Jev Yarn asks which line wins.Customs Authority asks whether it is sure.agent-fastpath asks ship, risk, and claim.Auto Mode for Paseo asks which Codex lane.jev-site-auditor asks brand and quote risk.jev_project_context asks load or skip.kb-jev asks the subject.swarm asks vetoes and tractability.Agent to Trust asks success, partial, or failure.jev-browser-skill asks the next click.jev-mcp asks which element.JevMatch asks must-haves and skills.browser-jev asks which control.Jev Autopilot asks on or off.Jev Exam asks each rubric point.Rules Are Lava asks if the answer obeys.jevelry asks team and urgency.JevApi asks typed questions.Longform Production Method asks each gate.jev-scout asks where Jev fits.SiteClarity asks if a section answers.Reflex Control asks accept or escalate.jev-shadow asks five hazards.Jev Flight Agent asks the next action.claude-router asks which tier.Bouncer asks seven hazards.jev-dom asks which control.Vampire Survivors asks which upgrade.MedHallu asks if the answer misrepresents.AII codequality asks fault statements.Jev Replay asks the same item twice.jev-fanout-bench asks one-call vs N-call.reddit-jev-tool asks if you can answer.dsh-jev-interceptor asks if the tool is safe.Tavli asks which point.BookmarX asks which folder.jev-accessibility-tree asks which node.Home Jarvis asks which card.jev-pilot asks which plan.omarchy-beam asks which action.claude-jev-routing asks which tier.jev-model-router asks which model.Jev City asks what to do next.jev-research-pipeline asks if the source keeps.jev-skill asks which tier.jev-verify asks if the identity holds.VotingAid asks how well it matches.jev-coherence-bench asks if the claim is supported.Jev Gatehouse asks if the call is destructive.opencode-jev-router asks how much effort.Bullshit Detector asks how much bullshit.JEV-Arcade asks which link.jev-backend-qa asks the real risk.Jev Flow asks which branch.ATHENA asks whether to go.firefox-jev-mcp asks which element.jevr asks how relevant.jev-opus asks how much effort.Doorman asks if it is real.jev-claude-statusline asks how far.jev-dingtalk asks who waits.Jev for Gmail asks how urgent.jev-claim-processor asks approve or review.rolldown-triager asks the priority.astrbot Jev gate asks whether to speak.Sieve asks if it is worth reading.jevlint asks whether the query is written badly.TypeSafe C++ SDK asks a typed question.jev-precision-router asks which model and effort.langchain-skill-router asks which skills.Kkdev92.Jev asks against a plan.can-jev-bayes asks which arm.pi-jev-router asks how much thinking.Slop Alarm asks which habit.jev-browser-use asks which click.Codex Task Router asks which worker.Mina’s Terrarium asks what she feels.Spark Jev Email Triage asks the category.Codex Subagent Router asks which profile.Paper Radar asks which paper.Zajszűrő asks how clickbaity.jev-voice-agent asks which tool.jev-why asks which span.pi-jev-core asks the question.jev-phone asks which tap.Jevcast asks which action.jev-cli asks the label.jev-x-filter asks which junk.JEV CI Pathfinder asks which job.Flick asks which control.jev-mcp asks the batch.Tweet Radar asks which post.jev-browse asks which element.Jev as LLM asks which word.jev-reflex asks which excerpt.JEV Security Sentinel asks PASS or BLOCK.Jev Draft Advisor asks which Codex model.jev-hermes asks approve or block.zot-jev asks if a tool is destructive.Snap Judge asks the same typed questions side by side.Jev Challenge Lab asks a drafted challenge.Jev Agentic Workflows asks which candidate.Jev X Sentiment Analysis asks buy or hold.jev-judge-experiment asks which track.awesome-jev-robustness indexes a robustness test.jev-in-the-wild indexes a wild build.awesome-typesafe-jev indexes a field guide.awesome-jev-essentials indexes a shortlist.typesafe-sdk-go asks which team.omp-jev-decision asks which queue.JEV Switch asks who should take the turn.jev-model-router-for-claude asks which tier.Jev Game Theory Arena asks which legal move.JEV Model Navigator asks which model.jev-cli (plurp911) asks noul, choice, or score.typesafe-playground asks action, review, or other.jev-harness (cyanheads) asks typed questions over rows.jevmcp asks whether code matches a spec.cog-typesafe asks a bound System One turn.jev-pi-model-router asks how hard the turn is.dsh-agent-kit asks whether a draft is wrong.jev-figma-design-plugin asks what to draw.jev-paste asks which field this clip is.jev-security-prioritization asks how urgent a finding is.Should I Jev? asks nineteen fit questions.jev-decision-mcp asks one batched decision.Jev Computer Use asks which app you meant.Let Jev Decide asks yes or no.jev-skills asks a typed noul.gmail-classifier asks which label.Jev Arena asks which legal move.dsh-skill-suggest asks whether a skill is needed.a2h-onus asks whether evidence holds.crush-monitor-with-jev asks how they feel.Jev from Java asks which desk.jevex asks what a row means.jev-testing asks how interested they are.copilot-jev asks which worker is next.jevkit asks which desk should act.跃然 StylePop asks which style matches.Reply Guy Repellent asks how generic a reply is.Dertek asks which route to take.jevcompat asks whether a server matches.jev-songwriter asks which note comes next.Jev The Spire 2 asks which move is safe.jev-vs-llms asks whether a prior holds.jev-internal-links asks which page to link.stock-lite asks which of four buckets.jev-poc asks seven shapes.Shapeshift asks which card you mean.ReflexGate asks seven dimensions at once.jev-cc asks which tier and whether to hold.claude-jev-funnel asks a batch in parallel.Jev vs Laya Tetris asks where the piece goes.go-jev asks noul, choice, or score from stdin.JevPDF asks whether this line answers.Jev steer or queue asks steer, queue, or interrupt.jevsearch asks which page answers.jev-pi-orchestrator asks whether a slice is done.jev-skills (eran-broder) asks whether this skill is needed now.JEV Inbox Router asks which folder a note belongs in.Sensibility asks whether a change is in scope.jev-enforce asks whether a rule was broken.jev-claude-code asks which model and effort.Jev Intent Triage asks department, priority, and PII.SwiftJev asks noul, choice, or score in Swift.TypeSafe.Jev .NET asks the same three primitives in C#.jev-skillbench asks whether a skill is malicious.Jeverative UI asks which component fills a cell.jcol asks each codebook column.macos-computer-use-kit asks whether the target is still right.好感度分析 asks emotion, intent, and warmth.Jev-cu asks which AX element to click.TypeSafe AI Playground asks the catalog questions.awesome-jev (JohnDotOwl) is a catalog, not a question.awesome-jev (MumuTW) is the same kind of catalog.TypeSafe AI Implementation Skill asks where a seam belongs.agent-workflow-typesafe-ai asks task class, interaction, and risk.kojev asks Choice, Score, and Noul as enums.pydecide asks the same three primitives over a chain.Nightwatch asks retry now, later, dead-letter, or a person.jev-swap asks the same decision the LLM was making.taurus-jev-sdk-go asks Noul, Choice, and Score in Go.jev-kmp asks the same three from Kotlin.vigia asks offensive and category.jev-sap-commerce asks abuse, spam, personal data, and on-product.jev-triage (CMaintz) asks type, priority, and security.typesafe-ai-php-client asks Noul, Choice, and Score in PHP.pi-jev-todo-audit asks whether the board still matches.jev-brotato asks which legal action to take.jev-reviewer-navigator asks which allowlisted reviewer fits.adk-go-typesafe asks the same three primitives from Go.Reclaim asks what kind of directory this is and whether it can come back.jev-flaky-detective asks regression, flaky, environment, or unknown.guideme-python asks Noul, Choice, and Score as Python values.jev-ql asks CHOICE, SCORE, and NOUL in SQL.ask-jev (Aether) asks one or more typed questions over MCP.Slop Fold asks whether a post was AI-written and which tell.jev-test-intelligence asks which allowlisted groups to run.jev-resource-rightsizer asks scale-down, keep, scale-up, or review.guideme-typescript asks Noul, Choice, and Score as TypeScript values.Jong asks which zone and which aim.pi-ask-jeff asks yes/no, a label, or a level.jev-loop (King4s) asks route, done, and recovery.typesafe-jev-dotnet-sdk (mehdihadeli) asks Noul, Choice, and Score from C#.JevNet asks the same three as JevNet questions.adk-go-jev-demo asks intent, subsystem, severity, and outage.jev-cu-jp asks which labeled control is next.jev-graphify asks a name, a fit, and a subsystem.polyjev asks whether a market is underpriced.yuru-come asks kind and priority.jev-plays (mansicer) asks which macro option.AD Sentinel asks fraud, fake warning, and hype.Tabwerk asks which tab and how important.jev-datafusion asks noul, choice, and score in SQL.jev-triage-benchmark (SENZO) asks urgency as situations, not degrees.Invisible CAPTCHA asks person or bot.jev-crush asks emotion, intent, and affection.jev-edge asks what the request is trying to do.xjevboost asks which rows and columns matter.typesafe-mcp (nirvana124) asks Noul, Choice, and Score over MCP.minesweeper-thinking-lab asks which cell to open.bloxmith-jev asks whether the message requests an action.jev-native-mcp asks which public item to read next.dsh-system1 asks Noul, Choice, and Score from Cordis.jevqa (sairohith) asks the next browser step.jev-issue-router (buckmoon) asks which model and effort fit the issue.jev-mcp-middleware asks six risk signals.jev-lab (tanayvasishtha) asks whether Jev is calibrated.claude-code-jev-smart-router asks which Claude tier the request needs.jev-skill-selection (redreamality) asks which skills to keep.jev-webmcp-extension asks which WebMCP tool matches the typed line.jev-cli (shaharia) asks a noul from a shell.jev-code-context-router asks which symbols to inject.Codex Jev Router asks Luna or Sol for the subagent.jev-browser-skill (hqman) asks which in-page action to take.typesafeai-sdk-rust-community asks Noul, Choice, and Score from Rust.typesafe-sdk-elixir asks the same primitives from Elixir.pi-jev-router (cskwork) asks which Pi model and effort fit.jev-model-router (lastlad) asks what the turn needs.jev-model-router (satviksinha) asks which Claude tier and effort.jev-router (reallygood83) asks home, implement, research, or write.pi-jev-router (toorop) asks route, clarity, complexity, and category.TypeSafe.AI.Sdk asks Noul, Choice, and Score from .NET.jev-lab (BrendanH18) asks your own questions in a workbench.jev-llm-router asks which Codex pair to use.typesafe-sdk-go (blazskufca) asks Noul, Choice, and Score from Go.typesafe-sdk-go (dwisiswant0) asks the same three kinds.typesafe-sdk-go (guchengod) asks from a PATH CLI.jev-mail-router asks intent and next action.jev-ai-sdk-form-router asks which team should receive a form.pi-jev-task-router asks task kind and complexity.claude-jev asks whether an edit breaks a rule.jev-recipes asks a catalog question.jev_model_routers asks which model to use.clouter asks which visual model fits.typesafe-sdk-rust (volker48) asks Noul, Choice, and Score from Rust.jev-apply asks which saved fact belongs in a field.typesafe-ai-sdk (aoprisan) asks Noul, Choice, and Score from Rust.typesafe-sdk-scala asks the same three kinds from Scala.hermes-switchyard asks which skill to load.jev-chatgpt-visibility asks how ChatGPT treated a brand.factory-dispatch asks which team should assess first.jev-webmcp asks which WebMCP tool to run.jev-cannibalization asks which page should own a search.astra-jev-harness asks which files the coding model needs.pi-model-router asks category and demand.jev-delegate asks which hit ranks first.dsh-model-router asks which answer model to pin.skill-picker asks which skill fits.JevRev asks which path is cheaper.routr asks whether the brief is standalone.DrissionPage-Jev-Browser-Harness-Skill asks which control to click.dsh-jev-prune asks whether a result is still needed.dagos asks whether a node is active.context-router asks whether a file belongs.partway asks whether to open an app.gatekeeper asks which handler owns the request.stingray asks whether the turn stopped half-done.sizeup asks which size can do the job.similarity-ts-jev asks how strongly to merge a pair.tdd-gate asks whether the test or the code is wrong.jev-playgroud asks a noul, choice, or score.jev-ai-project asks how urgent an alert is.typesafe-tutorial asks the official pattern questions.jev-inbox-queue asks whether you still need to act.vellum asks which block is the title.toolgate asks whether a call exfiltrates.agent-seek asks whether a snippet is enough.jev-toto asks whether a number looks hot.jev-demo asks which ref to click.dsh-context-curator asks whether a call is still needed.convoai-jev-vad asks whether the speaker is done.jev-skill asks which skill fits.jev-gtm-cookbook asks whether the new role buys.ego-jev asks which control to press.jev-rubiks asks whether to coach.dsh-completion-supervisor asks whether the claim is true.dsh-jev-plugin asks which item failed.inbasket-triage asks who owns the message.jev-showcase asks which team owns the ticket.tb-spam-guard asks whether the mail is phishing.play-with-jev asks which legal move to play.jev-paper-screener asks whether the paper matches.pi-jev-route asks whether to steer now.flick asks which control to tap.jev-cua asks whether the utterance is a command.jev-approach-control asks which aircraft to turn.jev-web-agent asks which element to click.langchain-jev-tutorial asks whether the ticket is urgent.esketcher asks which material to paint.jev-guard asks whether the call is destructive.typesafe-rust asks a typed question from Rust.jev-explained asks which label fits.jev-hermes asks whether the command is risky.typesafe-jev-benchmarking asks action, review, or other.luna-occulta-poker asks which action to take.trading_bot asks whether the setup is a trap.Prompt-Router asks which model to keep.jev-router asks which skill fits.ai-dj asks which move to play.jev-calib asks the same question again.jevkit asks whether the lead is worth outreach.jev-ai-use-cases asks which team owns the ticket.jev-mcp asks a typed question from an agent.jev-mcp asks the same System One body.llm-typesafe asks whether the message wants a refund.JevRPS asks rock, paper, or scissors.rust-bevy-jev-ants asks which intent to take.jev-relevance asks whether the chunk answers.typesafe-ai-go asks a Noul from Go.obsidian-smart-tagger asks which tags fit.tally-android asks which emotion fits.bili-hardcore-jev asks which option is right.jev-calculator asks which digit is next.jevai asks a typed question from Rust.chat2jev asks which department owns the ticket.Jev-Anti-Netabare asks whether the post spoils.dsh-jev asks which mood the turn has.jev-tmmluplus-eval asks which letter is right.jev-search-cli asks which result is on topic.opencode-jev-guard asks whether the command is ordinary.comms-check asks whether the line is AI filler.jev-playground asks the typed question you typed.veil asks whether the reply matches.file-guide asks which folder fits.learn-jev-end-to-end asks which team owns the ticket.jev-wingman asks whether a reply is due.jev-skills asks which route fits.jev-harness asks whether the patch is in scope.jev-lab asks whether the turn is in policy.Rime-Capsule-AI asks which word fits.dsh-teacher-consult asks whether a teacher is worth it.pairsort asks which of two is better.jev-rerank-bench asks whether a passage is relevant.typesafe-sdk-rust asks billing, tone, and urgency.pi-jev-router asks which agent and model.ego-jev asks which operation and target.pvz-jev asks which legal move.building-with-typesafe-jev asks which primitive fits.paperless-jev asks which correspondent and tags.jev-quant asks whether to trade.jev-codex-plugin asks whether the claim holds.jevzf asks whether a line matches.jev-ios-bridge asks whether the screen matches.jev-ticket-triage asks which queue and priority.n8n-nodes-jev-ai asks which route fits.jev-gmail-filter asks whether a topic matches.jev-desktop asks which operation and target.opencode-auto-jev asks which model to use.jev asks one noul, choice, or score.jevfilter asks which topic matches.jev-java asks noul, choice, and score.ralph-jev asks whether the objective is met.mastra-inbound-router asks which category and how urgent.lisa asks whether the diff adds a secret.GeoDeploy asks which region wins.jev-ragcheck asks whether a sentence is supported.jev-driver asks typed enums.unwedge asks whether the agent is stalled.jev-test-impact asks how likely a test is affected.jev-req-gate asks whether a requirement is feasible.jev-effort asks how much effort this turn needs.jev-expert asks whether a question is stable.jev-pi-token-reduction asks how much of a chunk to show.pi-typesafe-approve asks whether a command is routine.jev-frontier-bench asks the same 200 questions as the LLMs.pi-shift-router asks which tier this turn needs.jev-search-intent-classifier asks which intent and product.semcull asks what the output reports.baton asks whether this is a checkpoint.honeytongue asks how persuasive the line is.jev-game-engine asks which goal to run.pi-jev asks whether a block still matters.scrutus asks how accurate and useful a comment is.ha-jev-conversation asks which intent and device.jermes asks which skill fits.better-cheaper-llm asks whether a claim is grounded.jevort asks which effort tier.jev-workflow asks whether a contract still holds.computah asks which control to hit.Jev-Frame asks whether a candidate is supported.jev-fast-judgements asks whether an item is relevant.jev-ecc asks whether a tool may run.jev-riffs asks how significant a motif is.jevcut asks whether a clip stands alone.jev-tool-lint-eval asks whether a sentence is supported.jev-lifecycle asks whether a shadow path agrees.jev-harness asks whether a command may run.JEV_ADK asks whether a command is destructive.jev-review asks how severe a finding is.jev-watch asks whether the worker is looping.jev-2048-experiment asks which legal move.opencode-jev-plugin asks how much budget a turn needs.jev-rust asks a typed question.jev-controller asks which step comes next.jev-harness asks how hard a task is.voicevox-jev-proxy asks which reading is right.jevextract asks which class a span is.quicksilver asks whether an item matters.jev-hub lists a sourced page.jev-router asks which tier a turn needs.jev-browse asks which click is next.system-one-connector asks whether a ticket is urgent.jev-client asks which category a ticket is.jev-fabric asks whether a process is healthy.x-jev-gate asks whether a draft may go out.jeverifier asks whether a function breaks a rule.jev-worker asks a typed question.jev-model-routing asks which tier a prompt needs.dsh-jev-router asks how hard a turn should think.sooth asks whether a claim is supported.PRESS

Writing Score levels

Official Score page, beginner wording. A level’s number is its index incriteria, starting at 0. The model gets the descriptions and nothing else. The score is a position on your levels and can sit between two of them. Different distributions can produce the same score — read probabilities andconfidence with it. Confidence 1.0 means all probability sat on one level; it is not a guarantee the answer is correct. Use the score to rank or threshold — not to reconstruct “40% of the way from frustrated to angry.” Official: it does not measure a fraction of customers. Independent eval (jev-orderby-bench): ORDER BY a two-decimal Jev probability is not always a total order — the author saw 53 of 360 rows tied at 0.99.PRESSA score of 1.0 can mean all probability sat on level 1, or half on 0 and half on 2 — read the distribution. Official JSON keysprobabilities and legend by string level numbers; the Python SDK keys them by integers.

• Describe situations, not degrees. “Broken, but a workaround exists” beats “moderately severe.”
• Every level is evaluated separately. The model does not see a level’s number or its neighbours — “worse than the previous level” means nothing.
• Numbers-only levels fail. Official demo: “Rate 0 to 2” with criteria ["0","1","2"] split probability; the same report with descriptive levels scored 0.0 at confidence 1.0.
• One dimension per Score. “Punctual and smart and experienced” is three questions hiding in one.
• Use as many levels as you can describe distinctly, up to 10. Three is fine. Don’t invent levels you can’t describe.
• Give rare extremes their own level if you must act differently (e.g. “abusive or threatening” vs “very angry”).
• If there is no in-between, use a Choice or several Nouls instead of a Score.
• Several Scores in one call: official, divide each score by `len(criteria) - 1` so every scale is 0–1, then apply weights you own in code. Different-length scales are not comparable until you do.
• Start levels as strings. Official: if the model keeps scoring between two neighbouring levels on inputs you think are clear, give each level an object — what it covers plus a few example situations that look like your real inputs. Same field names on every level. An unrelated example leaves the split unchanged. Higher confidence is not proof the answer is correct.

Official score.md Structured level descriptions: the Safari export-crash report scored 1.43 at confidence 0.35 with plain strings. Adding a matching example (“export fails in one browser but works in another”) moved it to 1.03 at 0.96. An unrelated workaround example left the original split. Official: pick examples with known expected levels, then test the revised descriptions on separate inputs before keeping them.

{
  "criteria": [
    { "what": "Cosmetic; no impact to functionality", "examples": ["typo in a label"] },
    { "what": "Broken or degraded feature, but workaround exists", "examples": ["export fails in one browser but works in another"] },
    { "what": "Blocking issue; no workaround exists", "examples": ["cannot log in"] }
  ]
}

Writing a Noul

Official noul.md: the number is the answer and the certainty in one. Python builds optional clarifications with NoulCriteria(true=…, false=…). Official JS noul() defaults instructions to null;criteria is { true?, false? } or null.

• Ask one yes/no per Noul. Official: “Is the customer angry and asking for a refund?” hides two judgments — ask two Nouls and combine them in code.
• Phrase so a high value means yes. Official: “Does the message contain personal data?” is clear. “Is the message free of personal data?” inverts the meaning, and later code will read it backwards.
• A statement works as well as a question. Official: “The customer is requesting a refund” — near 1 means the statement is true. Try both phrasings on your documents.
• The number is the answer and the certainty in one. Official: there is no separate confidence field. 0.5 means yes and no are about equally likely — not a degree of the thing you asked about.
• Make the yes/no boundary unambiguous. Official: “Does this candidate have any Python experience?” works because “any” leaves no middle.
• When the boundary is subtle, add criteria.true and criteria.false (Python: NoulCriteria). Official advanced.md: a structured true / false object can carry a definition plus examples on each side. Try with and without; keep whichever is better on your documents.
• Do not treat a Noul as a Score. Official primitives page: a Noul of 0.5 means equal probability of yes and no — not a medium rating.
• Noul and a yes/no Choice on the same idea are different numbers. Do not copy a threshold from one to the other (official jaggedness).

Third-party example to watch: jev-agent.com asks Is this free of personal data? as a Noul self-check. Official noul.md says that invert phrasing makes later code read the number backwards. AskDoes the message contain personal data? instead.EDU

When a string is not enough

OfficialAdvanced: structure and JS EntryType: instructions and criteria may be a string, object, array, or null. Start with a string. Add JSON when the question has parts, or when the data is already an object. Official Python JSONContent is a string, mapping, or sequence — top-level None is not in that alias. JSON on a field is still a Jev question, not a JSON Schema document. Communityjev_jsonschema is a compiler that maps a schema onto Choice / Score / Noul and refuses free-form strings. Jev itself has no JSON Schema dialect.PRESS

• Clarity: a question with several parts is easier to read when the keys are labeled.
• Supporting data: a schema, taxonomy, or database row is already JSON — pass the object instead of stuffing it into a string template.
• Loops: one field object can drive a Noul, a Choice, and Scores. Build those questions per field or record in code and send them in one call (official SDE cascade does this).
• Compare lists: use an array — or an object with a compare list of paths — when the instruction is several things to check against each other.

Official API example: put a database row in one field and the question in another, then point at the row with backticks — same path style as state.

{
  "instructions": {
    "potential_duplicate": {
      "name": "John Smith",
      "location": "Oakland, California",
      "last_employer": "Google"
    },
    "question": "Is the resume for the same person as `potential_duplicate`?"
  }
}

Official advanced.md: one field object can drive mixed types — a Noul that verifies a value, a Choice that picks from candidates, and Scores that place the value on a scale. Loop over records in code and send every question in one call; the SDE cascade cookbook is the same pattern.

Arrays work when the instruction is a list of things to check or compare:

{
  "instructions": {
    "question": "Does the claimed sender identity conflict with the sending domain?",
    "compare": ["ticket.sender.display_name", "ticket.sender.email"],
    "focus": "Compare the named organization with the email domain."
  }
}

Choice options can be objects that say what an option covers, what it does not, and a few examples. Use the same field names on every option. Score levels can be objects too. Noul true / false can carry a definition plus examples when the yes/no boundary is subtle.

For a deep taxonomy, walk one Choice per level in code. Official: the options are the children of the current node, and each option’s value is that child’s subtree — so the model sees what lives under a branch before committing. That matters when a bottle could sit under bothSporting Goods > Cycling andHome & Kitchen > Drinkware. Closeprobabilities can justify exploring both branches. The hierarchical classification cookbook includes beam search for that case. Trim huge subtrees to direct children plus a sample of leaves.

Ask many questions in one call

Official: every question sees the same state, is evaluated independently, and returns a typed answer under the ID you chose. Adding questions barely changes response time and costs only the extra question tokens. Ask speculative ones and ignore them in code — that is fan-out, not waste.

Questions in one request are independent. Make a second request only when code cannot build it until it has the first answer: fetch more data, decide what the state is made of, or pick the next options. Official examples: skill suggestion (rank, then re-read top three), structure recovery (merge lines, then classify blocks), hierarchical classification (each Choice narrows the next options).

Official how-to-build: keep deterministic work in code (days overdue is a date subtraction, not a Jev question). Combine independent answers with rules or weighted sums you own. For learned composition, feed the probabilities as features into a downstream classical model. If you do not have labels, official: use an ensemble of expensive reasoning models to generate them; the AutoResearch cookbook shows the training loop. Two requests are the exception.

# One state, three question types — official primitives shape
state = { ticket_message, refund_policy }

refund_requested: Noul("Does `ticket_message` request a refund?")
request_type:     Choice(refund | rebooking | information)
frustration:      Score([calm, concerned, very angry])

# Then ordinary code:
if refund_requested.noul > 0.85 and request_type.choice == "refund":
    start_refund()