Lesson 09 · Jev 101

Your first Jev call (mindset + official path)

This page is a checklist—not a frozen SDK tutorial. APIs move; pin versions from official docs when you implement.

1. Official @typesafeai, Sep 20: “Jev is now available to everyone. No waitlist.” Start at https://console.typesafe.ai/. Same thread: $5 starter credit (~120 million tokens) — official X, not models.md. The homepage FAQ still says “Join the waitlist!” and /join-the-waitlist still exists. Pin the console the day you sign up. Do not confuse that $5 with Vercel AI Gateway’s monthly $5 free-tier credit — different bills, different keys.
2. Create an API key in the TypeSafe console. Official quickstart: get the key from the dashboard. Env var: TYPESAFE_API_KEY.
3. Send one state + one question (start with Noul or Choice).
4. Print probabilities. Write one threshold in code.
5. Only then compose with ChatGPT in a cascade.

Official first hour: the Playground

Officialquickstart: log into theTypeSafe console, open the Playground, paste any text as state, add a Noul, then mix types in one call.

1. Paste a ticket (official sample: Stripe integration failing, “losing sales,” “ASAP”).
2. Add urgency as a Noul: “Does this message express urgency?”
3. Add a Choice for department and a Score for frustration — see all answers at once.

Docs also generate Playground share links underconsole.typesafe.ai/decode#share/…. This 101 does not freeze one.

Official sample curl

Copied from the official quickstart. Pin jev-1.13.0 in production if you tune thresholds; the sample uses the moving alias jev-latest.

curl -X POST https://api.typesafe.ai/v1/systemone \
  -H "Authorization: Bearer $TYPESAFE_API_KEY" \
  -H "Content-Type: application/json" \
  -d @- <<'EOF'
  {
    "state": "Hi, I've been trying to connect my Stripe account for 3 days and the integration keeps failing. I'm losing sales. Please help ASAP.",
    "model": "jev-latest",
    "questions": {
      "urgency": {
        "type": "noul",
        "instructions": "Does this message express urgency?"
      }
    }
  }
EOF

Official HTTP errors

Fromapi.md. SDKs retry 429 and 529 with backoff by default.

Official Python SDK exceptions.md also types 400 / 403 / 404 / 5xx, plus connection, timeout, and response-validation errors. Those are SDK classes, not extra rows on api.md.

Third-party live notes (jevaiguide.com, Sep 19) also reproduced 400Unknown model, 400 max_tokens_exceeded, and 403 when the Authorization header was missing. Their table, not TypeSafe’s.PRESS

Where you can actually call it

Waitlist is still the official TypeSafe path. The rows below areplatforms. Their price, free windows, and options are not TypeSafe defaults.

Official SDK snippets

FromPython SDK (Python ≥ 3.10) andJavaScript SDK (Node.js 20+). Both read TYPESAFE_API_KEY and retry 429 / 529 by default.

Official Python gotcha: the SDK also groups answers asresponse.nouls, response.choices, andresponse.scores — not only a single answers map. Pin the current page when you type.

# docs.typesafe.ai/sdk/python.md
from typesafe_sdk import Choice, Noul, Score, TypeSafeClient

with TypeSafeClient() as client:
    response = client.system_one(
        state={"document": "I was charged twice. Please fix this ASAP."},
        questions={
            "billing": Noul(instructions="Is this ticket about billing?"),
            "tone": Choice(
                instructions="What is the customer's tone?",
                criteria={"calm": None, "frustrated": None, "angry": None},
            ),
            "urgency": Score(
                instructions="How urgent is this ticket?",
                criteria=["can wait", "this week", "today"],
            ),
        },
    )

print(response.nouls["billing"].noul)
print(response.choices["tone"].choice)
print(response.scores["urgency"].score)

// docs.typesafe.ai/sdk/javascript.md + noul/score helpers
import { choice, noul, score, TypeSafeClient } from "@typesafe-ai/sdk";

const client = new TypeSafeClient();
const response = await client.systemOne({
  state: { document: "I was charged twice. Please fix this ASAP." },
  questions: {
    billing: noul("Is this ticket about billing?"),
    category: choice("What is this ticket about?", {
      billing: null,
      technical: null,
      other: null,
    }),
    urgency: score("How urgent is this ticket?", [
      "can wait",
      "this week",
      "today",
    ]),
  },
});

console.log(response.answers.billing.noul);
console.log(response.answers.category.choice);
console.log(response.answers.urgency.score);

Official Python ops notes

From Python usage. SDKs retry 429 / 529 by default; you can pass a RetryPolicy on the client or per call.

• TYPESAFE_API_KEY — required. Official SDKs read it automatically.
• TYPESAFE_BASE_URL — default https://api.typesafe.ai.
• TYPESAFE_DEFAULT_MODEL — default jev-latest. Pin jev-1.13.0 when you tune thresholds.
• TYPESAFE_LOG_LEVEL — debug / info / warning / error / off. Official: secret headers are redacted; request and response bodies are not.
• Errors: catch TypeSafeAPIError and read status plus request_id.
• Forward compatibility: extra_body and raw question dicts exist so new API fields can be sent before the SDK grows a typed field. Prefer upgrading the SDK.
• Official Python changelog: v0.5.7 (Sep 14) was the first public release. v0.6.0 (Sep 15) madeScore.criteria an ordered list, acceptedMapping / Sequence annotations, made exceptions and responses picklable, and handles invalidRetryPolicy values. v0.7.0 (Sep 18) switched serialization from msgspec to pydantic, addedresponse_model on system_one(), and serializes str subclasses as strings instead of lists of characters. Pin typesafe-sdk ≥ 0.7.0 if you subclassstr. v0.7.1 (Sep 21) validates the API key early and excludes the value from logged exceptions, and adds AI-gateway examples. Official JS changelog still ends at v0.6.0.
• Official retries.md defaults: max_retries=2 after the first attempt; backoff 0.5s doubling to 5s with 0.25 jitter; retry HTTP 408, 429, and 500–599 plus connection and timeout errors; honor Retry-After andretry-after-ms; timeout=30 is the total retry budget. max_retries=0 disables retries.
• Official constants.md: DEFAULT_TIMEOUT = 10.0 seconds for each HTTP attempt. That is not the retry budget. Pass a largertimeout= on the client if one attempt needs more than 10s.
• Official sync client: empty questions, or a Score whose criteria list is empty, raise TypeSafeError locally — no HTTP. The same check exists on the JavaScript client.
• Official responses.md: SystemOneResponse has model, usage,answers, plus the grouped nouls /choices / scores maps.usage.input_tokens and output_tokens may benull. ScoreAnswer.legend is required — Python keys are integers; values may be a string, object, or array.request_id is x-typesafe-request-id.raw_http_response is the underlyinghttpx2.Response. Response models areextra="ignore", frozen, and strict.NoulAnswer still has no confidence field.ListModelsResponse.models is a tuple ofModelMetadata.
• Official questions.md: state cannot be None, but values inside an object may be. You can mix Noul / Choice / Scoreobjects and dictionaries with a type key in the same request.
• Third-party live log (jevaiguide.com, Sep 19): some pip mirrors still listed only 0.6.0 after 0.7.0 shipped. Their workaround waspip install -i https://pypi.org/simple typesafe-sdk. Their log, not TypeSafe. jevaiguide Python ↗PRESS

from typesafe_sdk import RetryPolicy, TypeSafeAPIError, TypeSafeClient

client = TypeSafeClient(
    model="jev-1.13.0",  # pin; usage.md also shows model="jev"
    retry=RetryPolicy(max_retries=3, backoff_max=0.2, timeout=1.0),
)

try:
    response = client.system_one(state, questions)
except TypeSafeAPIError as error:
    print(error.status, error.request_id)

Official async client: use AsyncTypeSafeClient so a server does not block the event loop.extra_body is last-write-wins and can override state,model, or questions. PassRetryPolicy(max_retries=0) to disable retries.transport and http_client are mutually exclusive.

# docs.typesafe.ai/sdk/python/api/clients/async.md
import asyncio
from typesafe_sdk import AsyncTypeSafeClient, Choice, Noul

async def main() -> None:
    async with AsyncTypeSafeClient() as client:
        result = await client.system_one(
            state="I was charged twice. Please help.",
            questions={
                "billing": Noul(instructions="Is this about billing?"),
                "tone": Choice(
                    instructions="What is the tone?",
                    criteria={"calm": None, "angry": None},
                ),
            },
        )
        print(result.nouls["billing"].noul)
        print(result.choices["tone"].choice)

asyncio.run(main())

Official JavaScript HTTP errors

From APIError and the leftover connection classes. Same header as Python:requestId is x-typesafe-request-id on HTTP failures. Official APIConnectionError covers DNS, TLS, or a closed body — parent of APITimeoutError (timeoutMs is the configured timeout). Official APIUserAbortError is an AbortSignal cancel.

Official JavaScript ops notes

From the official JavaScript SDK, TypeSafeClient, and RetryPolicy. Helpers: noul(), choice(), score().

• Official ENV names match Python: TYPESAFE_API_KEY,TYPESAFE_BASE_URL, TYPESAFE_DEFAULT_MODEL,TYPESAFE_LOG_LEVEL. JS log default is warn.
• Explicit constructor options beat env vars, which beat SDK defaults. Empty or whitespace env values are ignored.
• systemOne(request, options?) — second argument is per-call timeout, retry, headers, and abort.
• Official RetryPolicy defaults: maxRetries: 2 after the first attempt (0 disables retries); retry HTTP 408, 429, and 500–599; first backoff 500 ms, cap 5 s; backoffJitter 0.25 is a fraction randomly subtracted from each delay. HonorRetry-After and retry-after-ms up tomaxRetryAfterMs 60000 — a longer server delay uses backoff instead. apiConnectionError andapiTimeoutError default true (interrupted response bodies count). Partial overrides inherit unset fields from the client or SDK defaults. Official RateLimitError retryAfterMs is the server delay when present and valid. body is parsed JSON, response text, orundefined for an empty body. requestIdstill comes from x-typesafe-request-id.
• Official TypeSafeClient and RequestOptions: timeout is milliseconds per attempt. There isno total retry budget — unlike Python’sRetryPolicy.timeout=30. systemOne throws locally if questions are empty or Score criteria are not a list of at least two entries. Official Usage: input_tokens and output_tokens are required numbers. Python may return null.
• Official TypeSafeClientConfig: dangerouslyAllowBrowser defaults to false — turning it on exposes the API key to page users. Defaulttimeout is 10000 milliseconds per attempt.debug logs redact credential headers, not bodies. Official Fetch is (input: string, init?: RequestInit) — a string URL, not a Request object. Official EnvVar is just the union of ENV key names.
• Official SystemOneRequest: extra properties on the request are forwarded, includingnull. The JS type allows state: null. Official Python still raises if state isNone. Official EntryType covers state, instructions, and criteria. Official Python JSONContent is str | Mapping | Sequence — top-levelNone is not in that alias. Pin the page you type against.
• Official noul() takes optional instructions (defaultnull) and optional criteria. Official choice() and score() require both arguments. Score criteria are an ordered list of at least two; entries may be null. Official VERSION is 0.6.0.
• Official Models and ModelCard: await client.models.list() returns a ModelCard[]. Official Python returns a ListModelsResponse — iterateclient.models.list().models, a tuple ofModelMetadata with name,description, and release_date asYYYY-MM-DD. Official models.md: the list currently shows aliases; versioned IDs such asjev-1.13.0 are accepted whether or not they appear. Each card has name, description, andrelease_date. The JS constructor throws if the API key is missing, the config is invalid, or the runtime is unsupported. Official Python also raises TypeSafeError if the timeout is invalid.
• Official SystemOneResult: JS returns answers, model, andusage. It does not add the groupednouls / choices / scores maps that official Python responses.md documents. Readresponse.answers.billing.noul in JS. Official Logger is console-shaped: debug / info /warn / error. Official LOG_LEVELS lists those levels from most to least verbose. Official LogLevel is debug | info | warn | error | off —off disables logging.
• Official APIPromise is what systemOne() returns. Call withResponse() for WithResponse — parsed data, requestId, andresponse.requestId is x-typesafe-request-id orundefined. The HTTP response body is already consumed. asResponse() hands the raw Response; do not also await the parsed result on the same promise. Non-2xx reject as APIError, including throughasResponse(). map() transforms the parsed result while sharing one body parse. There is nowithResponse method on the TypeSafeClient class.
• Official JsonValue is string | number | boolean | null | JsonValue[] | object. Nested null is allowed. Official SystemOneRequestPayload is the POST /v1/systemone body after the model is resolved: nonempty questions, state asEntryType (including null).
• Official ScoreCriteria is readonly [EntryType, EntryType, ...EntryType[]] — at least two levels, indexed from zero; null leaves a level undescribed. Official ScoreOf infers those indices from a fixed-length tuple, otherwisenumber. Official ScoreLegend keys rubric text by ScoreOf. Official ChoiceCriteria maps labels to EntryType; Official Description is an alias for EntryType —null leaves the label undescribed.
• Official ResultFor keeps the criteria keys: Noul → NoulResponse (noul + type only), Score → ScoreResponse, Choice → ChoiceResponse (choice is keyof T & string). Official JS ScoreResponse.probabilities is keyed bynumber | `${number}` — both shapes type-check. Python still keys legend and probabilities by integers. JSON still uses string keys.
• Throws if Score criteria are not a list of at least two entries. Official changelog v0.6.0 (Sep 15): that list replaced a dictionary keyed by integers.
• Vercel’s evaluate tutorial maps Noul to type: "boolean" — a Gateway field name, not TypeSafe’s noul. Same docs: Gateway evaluation needs the AI SDK; compatibility endpoints do not support it.rotom exposes that same evaluate door locally at/v1/evaluations and tells you not to send those models to /v1/responses.Vercel KB ↗ · agent-control ↗PRESS
• Official ChoiceQuestion, ScoreQuestion, and NoulQuestion: instructions may be omitted or null. Noul criteria may be { true?, false? } ornull. Official Questions is a map of names to those objects — the same names come back onanswers. Official Question is the union NoulQuestion | ScoreQuestion | ChoiceQuestion, identified by its type field.
• Vercel AI SDK TypeSafe provider (ai-sdk.dev): package @ai-sdk/typesafe-ai, envTYPESAFE_AI_API_KEY — not the official SDK’sTYPESAFE_API_KEY. Callexperimental_evaluate withtypeSafeAi.evaluationModel('jev-latest'). Noul is type: "boolean"; readanswers.*.probability, not noul. Choice / Score confidence is underresult.providerMetadata.typesafe.confidence. Scores and probabilities are rounded to two decimals and may not sum to exactly 1. The provider has no retry loop of its own.PRESS
• Independent hosted wrapper (jevtypesafeai.com/docs): POST /api/v1/decide, keys jv_live_…, prepaid balance. They list $0.42 / 1M input — ten times TypeSafe’s $0.042 / 1M — and say go to the TypeSafe console for at-cost pricing. Extra HTTP 400 / 402 / 403 / 502. Not the official endpoint. A later money trap:hj01857655/jev-router README prices input at $0.0578/M — also not official $0.042/M.PRESS
• Pydantic AI TypeSafe provider (pydantic.dev): pip install "pydantic-ai-slim[typesafe]", envTYPESAFE_API_KEY, modeltypesafe:jev-latest. Each output_typefield is one question; put the question on the field, the ticket in the prompt. A bare bool with no description and no instructions is a UserError. bool isTrue when probability is at least 0.5. Confidence is a margin in provider_details["confidence"], not P(right) — for yes/no it is distance from 0.5, doubled.typesafe_tool_call_threshold default 0.6 is a starting point from a small internal ticket set. A tool pick is classification of the text, not a safety judgment — a no-arg tool that sends mail can still run. Pintypesafe:jev-1.13.0 after you tune a threshold. Platform docs, not TypeSafe SDK defaults.PRESS
• LiteLLM proxy (pass-through and compaction): replace api.typesafe.ai withLITELLM_PROXY_BASE_URL/typesafe. Clients use a LiteLLM virtual key; the proxy injects TYPESAFE_API_KEY. Streaming is not offered. Compaction guardrailjev-compaction runs pre_call: older tool results below relevance_threshold 0.2 become a short notice. If Jev is down, LiteLLM fail_open sends the original request. Optional Auto Router withclassifier_type: jev. Distinct fromtamaratran/fast-jev-compaction (Claude Code plugin + npm library): Jev scores each tool call and result; kept text stays verbatim; stale text is dropped or truncated, never rewritten;keepThreshold 0.5; a Jev error throws and the hook falls back to Claude Code’s built-in summary. Distinct fromkamchatka’s Jev-assisted compaction (one Score per tool result on oneshot / obsolete / referenced / marked, compared to age Trim). Distinct fromtamaratran/jev-pruner which trims Bash after it runs, before the LLM sees it. Stdout of 10,000 estimated tokens or fewer passes through. One Noul per chunk. keepThreshold 0.5; scores above 0.1 also stay. Errors and whole-document commands are never scored. A Jev or archive failure leaves the original output. Distinct fromcompozy/yoshi — an experimental loopback proxy. Jev via Vercel Gateway. Without a key it forwards unchanged. Both loss Nouls must be ≤ 0.2 to omit. Author: no single savings percentage (Fable −34.39% input but slower; Sonnet mixed session no savings). Distinct fromjoelhooks/pi-fast-jev-compaction — a Pi port of tamaratran. Two Nouls per tool call.keepThreshold 0.5. Fail-open, then Pi’s own summary. Distinct fromKamilPostrozny/pi-fast-jev-compaction (same name, not a GitHub fork; 0.7.4 savings gate; Jev never sees result bodies). Distinct fromleonaaardob/fast-dev-compaction — the Codex wrap of the same engine. npm name is stillfast-jev-compaction. Author: do not use it for real work. Fail, missing key, or reduction below 0.25 leaves Codex’s own summary. Distinct fromilkerulusoy/pi-jev-compact — a manual /jev-compact that writes anew session. Prose is opt-in at 0.3. Pi’s footer % can stay stale because replayed usage is copied. More than six tools now. None is a TypeSafe default.PRESS
• Independent decorator (pip install jev, 0.3.0): @jev.fn compiles a Pydantic return model into questions. Needs Python ≥ 3.14. Author: probabilities and confidence are discarded — bool isp ≥ 0.5 (tune bool_threshold orJEV_BOOL_THRESHOLD), Choice is argmax, Score is the expected value. Need a gate? Use officialtypesafe_sdk, orKiln-AI/jev_jsonschema if you already have a JSON Schema — that bridge keeps probabilities, maps bool → Noul (True at ≥0.5), enum → Choice, integer range ≤10 → Score, and refuses free-form strings, arrays, objects, and anyOf. Already on Zod 4.3+?jomatsu/zod-jev keeps shape checks in Zod and sends everysemantic() rule as a Noul in one request. Fail closed: unavailable is never a pass. Defaultt = 0.95; lowering it widens auto-reject (p ≤ 1 − t). parse() throws — useparseAsync. Choice / Score still need the raw SDK. Not jev_jsonschema. Jev itself does not speak JSON Schema. Name collision with otherjev CLIs. Not TypeSafe.PRESS
• Local clones that speak a Jev-shaped API — not TypeSafe.githubnext/localjev is wire-compatible via DiffusionGemma Chat Completions; the author says probabilities are self-reported, not logit reads.TheoLeeCJ/openjev reads option logits off a frozen Qwen; the author reports 0.845 vs 0.883 modal agreement on a 102-row TypeSafe subset.r-ms/mini-jev serves POST /run on :8765 and calls letter shares a ranking-with-gap, not calibrated probabilities. Apidog’s local-clone roundup also maps jevlike, an MLX constrained-decoding engine, and unmerged vLLM PR 57250.omo-jevlike-router is a local frozen Qwen2.5-0.5B + jevlike head on:8318. It shrinks OmO<available_skills> to top-K. Fail-open (2.5 s). Author held-out top-1 27.3% / recall@24 84.1% — an earlier 46.2% / 95.5% figure was an eval bug. Not hosted TypeSafe Jev and not RLCD. A community thread says Jev was trained on 100% synthetic data — treat that as an unverified rumor. None of these reproduce RLCD.NanoJev is a 0.6B Qwen3 local replica (POST :8765/api/evaluate). RLCD post-training is still on the roadmap. Author game scores vs a Jev run stay with the author.Laya (pip install laya) is a self-hosted clone with a Router. The author claims RLCD; Jev comparison rows are third-party published figures the author did not measure. Base checkpoints sit near chance zero-shot.us/jev-local speaks official /v1/systemone; its default scorer is deterministic and has no intelligence — pointing base_url at it is not TypeSafe Jev.LitJev turns any Qwen into the same schema with no training; probabilities are uncalibrated by default.minojev trains a ~0.8M head on frozen Qwen3-1.7B (author in-domain 95.8% vs 80% generation; OOD 31.7% vs 40%) and names the yes/no primitive boolean, not Noul.Von is a 395M local NAR (README on master). Author v2 71.5% vs published Jev 96.6%; its confidence is P(top)−P(second), not TypeSafe peakedness.JevForge is a local 0.8B scorer; its “browser RLCD” is not TypeSafe RLCD.jevbetter beats jevlike on hard synthetic menus — still not hosted Jev.openJev-verdict-2.0 holds two models; the author’s ~150M Jev row is not a TypeSafe publication.rupeshpoojary9/poorjev is commodity NLI plus temperature scaling; author ECE 0.170 → 0.071 on their small set.jaredpalmer/kev is LoRA + a pointer head on Qwen, served as local/v1/systemone (Hub jaredpalmer/kev-4b). Author: no Jev outputs used for training; a TypeSafe-docs example can disagree with Jev.vagmi/jevlite (Hub vagmi/jev-lite) is QLoRA on Gemma 4 E4B. Author: choiceconfidence is p_max of the selected option — not official TypeSafe peakedness.OmniJev/PlayJev (Hub PlayJev-0.8B) is a pixel game model on Qwen3.5-0.8B — JEV-like moves, not the hosted API.scienthoon/luce trains LoRA plus a decision head on Qwen3-4B-Base and servesPOST /v1/ask — not /v1/systemone. Author numbers vs Jev stay with the author. PointingTYPESAFE_BASE_URL at one is a local experiment, not the official API.qwen-rlcd is M0 prefix-fork on Qwen3.5-0.8B — Choice/Score/Noul heads are still unbuilt M1; OmniJev overstated what ships.LFM2.5-350M-RLCD keeps stock LiquidAI weights; valid JSON is assembled in Python and is not correctness; the name is not TypeSafe RLCD.pngwn’s Qwen scorer is LoRA plus a scalar head — tickets_queue 0.234 is the author’s.harshatheg’s “1B-RLCD” loads 1.5B Instruct-4bit and never defines RLCD.eve-rlcd is the author’s own loop and says TypeSafe published no method.system-one-mini only answers five frozen heads.PRESS
• Ranking caveat (yodablocks/jev-orderby-bench): author pre-registered gates. jev-1.13.0 passed on 20 Newsgroups topic membership and failed four of six on Amazon ESCI graded relevance. Two-decimal probabilities left 53 of 360 rows tied at 0.99 — ORDER BY plusLIMIT k can cut inside a tie. One seed. Author numbers, not a TypeSafe bench.PRESS
• Independent routing measurement (FirasSX914/Janus, pip install janus-decide): shipsno default threshold — it measures one on your labeled set or decision log. Author replay: Banking77ROUTE at 0.67 (80.2%, $0.1033, 302 ms p50) vs Web of Science DO NOT ROUTE. The optimal threshold, the sign of the accuracy gap, and whether routing paid all changed. Pin the resolved model, not an alias. Not TypeSafe.PRESS
• Semantic grep tools — four different CLIs.keltokhy/jgrep (uv tool install jev-grep, commandjgrep) asks one Noul per line; default threshold-p 0.5; default budget $1. Author OpenRouter bench (Jev 1.13, 300 SMS): F1 0.90 / 2.7 s / $0.0039.zaydmulani09/jevgrep shortlists with ripgrep, then Jev ranks about 50 candidates via OpenRouter.nassim-arifette/jevgrep is a third jevgrep: npm@nassim-arifette/jevgrep; the command is stilljevgrep. The unscoped package namejevgrep belongs to a different project. TypeSafe transport is simulated, not live-tested.jev-tools (uv tool install jev-tools, adiel-hub/Jtools 0.1.1) is a suite of ten Unix commands — a freshjgrep on shared jevcore that interoperates with keltokhy’s config and cache, plusjsort / jpick / jgate /jwatch / juniq / jhead /jtag / jroute / jmatch. Default -p 0.5. Most tools fail open (exit 5);jgate fails closed. Author Vercel bench (238 ms p50, $0.000013) stays with the author.jselect (uv tool install jev-select) is a sibling: token-budget evidence, not line grep. Full semantic scan is preflighted and never silently shortlists.jlink (uv tool install jlink) is record linkage: blocking is local, then one Noul per candidate pair. TypeSafe key wins. Default threshold 0.5. macOS ships/usr/bin/jlink — use jev-link orpython -m jlink. Author benches stay with the author. Do not send confidential fields.jev-curate is cargo install jev-curate orpip install jev-curate. It sifts Parquet with a preset (README names reasoning-math). Thin README.keltokhy/jsort (uv tool install jev-sort) is pairwise Noul plus Bradley-Terry — not jev-tools jsort, which is a Score. Author CommonLit r=0.824 on 300 excerpts stays with the author. Choice put 93% of answers at the extremes, so the author preferred Noul.hotchpotch/jev-reranker is a pip library; relevance_rerank default threshold 0.2.shinpr/jev-reranker is an npm stdin/stdout CLI; rerank / filter / compress stay separate; filter default 0.5; 429/529 retry twice. Same name, different install.uehaj/jev-semgrep (npm i -g @uehaj/semgrep, commandsemgrep) asks one Noul per line; 30 lines per request; default -t 0.5. Name collision with Semgrep.dev. Every searched line goes to TypeSafe.nlgrep uses 40-line windows, default threshold 0.8, pinsjev-1.13.0, key JEV_KEY.jegrep (cargo install jegrep) is another meaning-search CLI: live tree, no index; cascade last-t is 0.20. Not jgrep, not uehajsemgrep, not nlgrep.AkashPriyadarshii/jev-scout looks up GitHub + crates.io first, then Jev Score(fit 1–4) / Noul(modern) / Choice(best_match). “Zero hallucination” means verified package names, not TypeSafe’s 0% type-error claim. Author: cold 2.0–2.3 s / warm 0.3 s. README is on master.kierandotai/jev-scout is a different repo: an MCP search/fetch gate. Annotate by default; gate uses a 0–2 scale (default 1.0) and fails open. No key → mock keyword scores. README waitlist note is stale.jev-seo scrapes DuckDuckGo and asks Jev on some commands. audit / robots / schema stay local. “Zero-cost / ₹0” means no Semrush subscription — a TypeSafe key still bills. Distinct from the BTK SEO study. README is onmaster. None of these are TypeSafe.PRESS
• Agent plugins that keep text with the LLM.shitianfang/jev-use (npx -y jev-use install) wires Claude Code, Codex, and pi. jev_judge batches check / pick / rate;jev_gate can only deny or ask. An unreachable backend escalates instead of allowing. Distinct fromsavka777/jev-use (macOS Accessibility tree, no screenshots) and the Saccc_c “Jev Use” computer-use clip.sutro-sh/jev-align (jeva) labels unsure rows and lets GEPA propose; a higher training score never auto-accepts. Registryai-functions.dev is public — do not push labels you cannot share.jev-belay is a Stop hook that fails open.limpet is a different Stop hook: one Noul per plain-language rule; shadow until LIMPET_BLOCK; the second stop of a chain always goes through; no key / error / timeout exits 0. TypeSafe wins over Vercel. That repo is noplan-inc/limpet —jev-superpowers names a different limpet (KSym04/limpet). Its “₹0 Jev API key” still bills TypeSafe; “0.0% package hallucination” is the Akash cargo scout, not official 0% type errors.jev-judgment is a skill (npx skills add HyunjunJeon/jev-judgment): closed judgments before a user question, before a destructive command, and after a failure. Without a key it fails open. jev-commit warns on a mismatched message and blocks only a credential-shaped added line.jev-git is a different hook: command git jev. Thin README. Author ~80 ms claim stays with the author.wakegate asks one Choice (wake / not yet / unrelated) and skips a wakeup only if Jev answers and p(wake) < 0.2 — user messages, bare timers, skip-limit, errors, and unsure all wake.doronp/jevc (npm i -g jev-compiler, commandjevc) scans CLAUDE.md / AGENTS.md — author: a text heuristic — and compiles a decidable rule into typed questions plus a reducer.kamchatka --advise can refuse a tool call the rules were going to allow; it can never permit one. Unreachable leaves the rules. --assisted-shell colors a command and decides nothing.jev-guard asks a risk Score plus approval / from_untrusted Nouls. Unreachable Jev allows unlessJEV_GUARD_FAIL_CLOSED.Sniff Test keeps regex rules local; judgment Nouls wait for a yes.typesafe-adblock is a toy: one Noul per candidate, defaultP ≥ 0.70.unclutter is a different MV3: hide only when selected-choice probability and TypeSafe confidence are both ≥ 0.9. Cookie hide is not a rejection.citation-verifier lets Claude propose a quote, code prove it exists, and Jev score supports / contradicts / says nothing — a human still decides. None of these are TypeSafe.PRESS
• Android agents — three different loops.Friedjof/jev-mobile is a local USB worker: Jev sees only prevalidated candidates and cannot invent coordinates, MCP calls, or code. Payments and account changes stay out of scope. Distinct fromdroidrun/mobile-jev (Mobilerun cloud).antiyro/jevdroid reads the accessibility tree and taps via ADB or UIAutomator2. Author bench: 314 ms median, $0.00003374 average, 78.6% on 56 synthetic decisions — author numbers.JevNoiseGate is a notification/SMS gate: OTP never leaves the device; every uncertain path allows; the API key is stored in plaintext.emreozyoruk/hush is a GitHub Action that labels only above your thresholds and stays quiet otherwise — not davesheffer/hush. None of these are TypeSafe.PRESS
• Model routers — five different jev-router names.gargpratyush/jev-router (npm i -g jev-router) launchesjev-claude / jev-codex. Jev picks a shared tier per fresh user turn; tool loops keep it. Fail-open. Prompt text goes to TypeSafe.prismhq/jev-router is a LiteLLM proxy: model: "jev-router". Without a TypeSafe key the cheapest eligible baseline runs.mejiasd3v/pi-jev-router pins model and effort once, then may suggest a fork — it never auto-switches. Opt-in skills load atp ≥ 0.8 (heuristic). Distinct fromphilippdubach/pi-jev-router (Pareto knee; /router auto can switch).jcm-router always routes subagents; the author measured a $19.53 loss when the main chat was switched (cache rewrite at 2×). Model / effort gated at 0.7.jev-agent-skill-router uses the command jev-router — name collision with the npm package. It chooses a skill and never loads it. Author 68/72 synthetic stays with the author.Jev-Auto-Router is a per-call prototype. Author: not production. Floor 0.55 falls back to Terra/medium.pi-typesafe-router (pi install npm:pi-typesafe-router) startsoff. Four backends: TypeSafejev-1.13.0, Cloudflare typesafe/jev, Vercel typesafe-ai/jev, OpenRoutertypesafe/jev-1.13. Do not enable it alongside another automatic router.hj01857655/jev-router is a sixth jev-router name: one call of six questions per ticket (team, P0–P3, urgent, auto-reply, mood, complexity). README prices input at $0.0578/M — that is not official $0.042/M. None of these are TypeSafe.PRESS
• Review, screening, and page grades that keep the last step in code.DiffJury judges a public PR in onesystemOne call. The verdict is not a GitHub merge ACL. Uses jev-latest. No LLM fallback.cv-screen asks one batch per CV, then scores in code. Changing a weight re-scores stored candidates in about 20 ms and 0 tokens. A Noul in 0.35–0.65 goes to a person.needs-review never overrides the verdict. Author 40-CV folder numbers stay with the author.email workflow is a LangGraph Choice (invoice / general). Handlers do not send mail or pay. API failure stops. Default pinjev-1.12.new-api-plugin-typesafe exposes native /v1/systemone — not/v1/chat/completions. Its README billing example uses $42 per million input tokens; official TypeSafe list price is $0.042 per million.pagegrade is a different Kitze MV3 from unclutter: ten 0–4 Scores, then × 25. A–E is an editorial heuristic, not a ranking prediction. Click-to-analyze.Foreman asks ten Nouls; Python continues, steers, or stops. Author: experiment. Uses jev-latest.Clean Code Review lets Jev judge files and Luna write the review. Public PRs. Advice, never merge authorization. Not DiffJury.JevSlop scores note.com writing quality. It is not an authorship detector. Distinct from slop-grader.jev-bfs has Chromium read Wikipedia, Jev rank titles (no article text), and Python walk. Beam, not shortest-path.hermes-jev-approvals is approvals only — not a sandbox. Six questions in one request; a deterministic reducer maps them (self_advocating ≥ 0.6 escalate; secrets + outbound ≥ 0.7 deny). A truncated command is never auto-approved. Backends: jev-latest, OpenRouter~typesafe/jev-latest, or OpenCode Zenjev-1.13-free anonymous.sift labels X posts (Substance / Humor / Chit-chat / Promo / Junk plus AI-written / Off-topic). Author ~$0.00003 per post. Default hide Junk, Off-topic, and AI-written. Distinct fromkbhuw/jev-sift (screens files and public URLs; does not execute tools).stanley-code is a Jev-first agent: Jev routes the workflow and answers fixed-choice questions; your code still decides. First 150 hunks. TYPESAFE_MODEL. Not affiliated with TypeSafe.nexibeo/jev-cookbook is fifteen Node recipes on OpenRouter/api/alpha/decisions with~typesafe/jev-latest. Distinct from chr-kelly/jev-cookbook. Author 16–36 item samples stay with the author.flue-jev-demo keeps chat on gpt-4o-mini and routes with Cloudflare env.AI.run('typesafe/jev'). Jev is a Flue tool, not the chat model. No application API keys. None of these are TypeSafe.PRESS
• Pi extensions that keep policy in code.pi-heed never lets Jev write a rule — Jev only classifies how your words changed an existing one; uncertain answers change nothing. Shadow by default; fail open on a Jev error or a 2.5 s timeout.Nyarlathoteppppp/pi-jev-context hides only blocks with P < 0.15 that they are useless. Old-context pruning is shadow — author: raw Jev would drop later-needed lines on 73% of real-session items.kevinpita/pi-jev-context (pi install npm:pi-jev-context) is a different repo: off until /jev on; keep only when p > 0.8; hides, does not delete; on error earlier pruning still applies until /jev off.pi-verdict fails closed (allow / ask / deny). Its Jev adapter is experimental and defaults to OpenRouter.pi-warden quotes the broken pi-warden.md rule; default steer still lets the write land.DGP lets Jev assess a frame; the application still guards the commit. The included mock demo is not Jev.Supercov asks Jev yes/no quality questions per file; coverage runs need no TypeSafe key.slop-grader grades lines and the document; only confidence ≥ 0.8 line hits print; an LLM writes the fix plan.typesafe-ai-playground prints four independent PHI Nouls and does not combine them into a legal yes/no. OPENROUTER_API_KEY wins.jev-agent-skill calls OpenCode Zen’s free /v1/systemone — pinjev-1.13-free, not jev-latest.fastbrowse is a pre-alpha browser agent: Jev picks an on-page control; an LLM plans and reads. Secrets are typed by name only. Irreversible actions stop without--authorize. Author uncapped 2026-09-20: 41/42 vs hosted Browser Use 42/42 — the earlier 2.9× pass claim was a $0.25 cap. Distinct from jev-ultrafast.jev-browser (npx -y @jkudish/jev-browser) is another picker: code owns stop gates (goal or stuck > 0.85). Password fill is code-only, never submits, never reaches a model. Typing uses a second model or a keyword heuristic.jev-browser-use is a Codex skill: Jev clicks; Codex types and verifies. Author 5–10× is EZCollegeApp, not a TypeSafe SLA.jev-canvas speaks English or Ukrainian at a tldraw canvas. Eight or nine questions per transcript. Thresholds live inpolicy.js. The key stays on127.0.0.1.jevinci (README: jev-paint) paints from probability grids. The key lives in localStorage. Independent of TypeSafe.Augustus is a design skill, not a TypeSafe product: never launder a Noul as a proof.robo-harness lets Jev pick bounded arm steps through Vercel Gateway. Chat and a decision run never move together. Default spend cap $10.jev-for-chrome is a Manifest V3 port of jev-ultrafast in your own tab: Jev picks the operation and element; a small chat model types; password fields are never read. Its loop kit isjev-dev-kit 0.3.2 — npm install github:chy4pro/jev-dev-kit#v0.3.2, not on npm. Jev only ever chooses;options() supplies candidates;text() writes values. No network code inside.jev-reach runs the same pick-then-act loop inside chrome-devtools-mcp: Jev never writes text — typed strings come from values. done 0.85 and max_steps 15 are the author’s. Not TipTour. Distinct fromYing-Kai-Liao/jev-browser (standalone Playwright; 40/42 is the author’s).pi-typesafe is the shared Pi key store: typesafe_evaluatestays off until /typesafe enable;TYPESAFE_API_KEY wins; it ignoresTYPESAFE_BASE_URL. The README waitlist note is stale vs official X. Underpins pi-warden.pi-quiet-ask (pi install git:github.com/HyunjunJeon/pi-quiet-ask) is packs on Pi hooks. Default shadow → warn. Everything fails open. Without a key it registers nothing. Distinct from HyunjunJeon/jev-judgment.pi-typesafe-jev is a different package (pi install npm:pi-typesafe-jev): five tools (typesafe_noul / choice /score / evaluate / ask), native fetch, no SDK. The key is stored by/typesafe setup and is never read from the environment. Author: not verified against the live API.evaluate MCP is one Go binary (evaluate):setup mcp wires Claude / Codex;setup pi writes a native extension.TYPESAFE_API_KEY wins; OpenRouter still uses/api/alpha/decisions. Retries 429 and 529. Distinct fromblakestone-x/jev-mcp (uvx --from git+…@v0.2.1 jev-mcp) — key stays in the client environment, never in tool arguments;jev_screen is a filter, not a security boundary — from tumf’s stdio jev-mcp, fromjkudish/jev-mcp (npx -y @jkudish/jev-mcp; ten tools;jev_screen is advisory — the server never blocks), fromburnigtm/jev-mcp (nine tools; incomplete context never permitsauto; Jev never invents arguments), fromBYK/jev-mcp (npx -y jev-eval-mcp; ask / map / eval — measure a threshold, then apply it), fromminhgv/jev-mcp (nine tools includingjev_coding_loop; the server does not read the repo or apply patches; ci-shadow reportswould_block and exits 0), fromarunav25/jev-mcp (install @arunav25/jev-mcp from GitHub — the bare npm name jev-mcp is a different project; one evaluate tool; 429/529 retry four times, 401/422 do not), and frompedroknigge/mcp_jev (closed catalog of five tools; key once in~/.mcp_jev/.env; ping.packs ≥ 13; not published to npm), and fromrashedInt32/jev-mcp (npx -y jev-mcp / pluginjev@jev-mcp; key never a tool argument; act 0.8 / review 0.5 are the author’s; plugin idjev also collides with jevwire).building-with-jev is a skill that writes Jev programs (targetsjev-1.13), not officialtypesafe-ai/skills.y0usaf/pi-jev (npm @y0usaf/pi-jev) is a shadow-mode gate plus output judge plus jev_ask. Every error path fails open. Distinct fromTheoOliveira/pi-jev (pi install npm:pi-jev; auto / tools / compact off by default; tools 0.65; gate 0.70; unreachable Jev does not activate unjudged tools) and from pi-heed / pi-verdict / pi-warden. None of these are TypeSafe.PRESS

import { TypeSafeClient } from "@typesafe-ai/sdk";

const client = new TypeSafeClient({
  defaultModel: "jev-1.13.0",
  retry: { maxRetries: 3, backoffMaxMs: 2000 },
});

const response = await client.systemOne(
  { state, questions },
  { timeout: 1000 },
);

Official surfaces (re-check before you paste)

• API: POST https://api.typesafe.ai/v1/systemone with Bearer auth
• Python: pip install typesafe-sdk (Python ≥ 3.10). Exception tree: exceptions.md · async client
• JavaScript: @typesafe-ai/sdk
• Model ID to pin: jev-1.13.0 — aliases jev-latest andjev-preview can move. Third-party live test (jevaiguide.com, Sep 19):jev-1.13 without the patch returned HTTP 400 “Unknown model,” even though one official example uses that spelling. Pin the dotted ID.PRESS
• Official agent skill — see the section below. Prefer one install method so you do not get two copies.
• Alternate platform paths — not TypeSafe defaults: Vercel AI Gateway (typesafe-ai/jev, changelog · KB), Netlify AI Gateway ( changelog), OpenRouter beta, Cloudflare Workers AI (typesafe/jev, docs)
• Docs: quickstart, docs index
• Official cookbooks install helper code with pip install … cooksafe --extra-index-url https://pypi.typesafe.ai/. Many notebooks still say jev-1.12. Pin jev-1.13.0when you ship.
• Official comparison client: system-one-adapter-python is a drop-in TypeSafeClient backed by an LLM. Same questions, not Jev.
• Third-party browser playgrounds such as jevplayground.com are not the official console. Prefer console.typesafe.ai when you have a key.PRESS

Official agent skill

From agent-skill.md. Agents still generate via LLMs. The skill stops them asking one question per call and inventing request fields. Official: put questions and thresholds inone file — agents are not great at writing questions.

• Claude Code: claude plugin marketplace add typesafe-ai/skills then claude plugin install typesafe@typesafe-ai. Invoke /typesafe:typesafe-ai.
• Other agents: npx skills add typesafe-ai/skills --skill typesafe-ai. Add -g only if you want it global.
• Update Claude Code: claude plugin marketplace update typesafe-ai then claude plugin update typesafe@typesafe-ai, then /reload-plugins. Other installs: npx skills update.
• Choose one install method. Two copies confuse the agent.
• Claude Code auto-update: open /plugin, then Marketplaces → typesafe-ai → Enable auto-update.
• Official skill prompt points agents at https://console.typesafe.ai/docs/cookbooks — that is the console cookbook index, not docs.typesafe.ai/cookbooks.
• Official: agents are not great at writing questions. Put questions and thresholds in one file and expect to edit collaboratively.
• Official habit: if you only need the best option, pick the one with the highest confidence. Do not invent a threshold for that. If you have a statistical rule in mind, use the probability distribution.
• If the agent invents request or response fields, the skill is stale — update it.

Community language clients

Official SDKs are Python and JavaScript. The rows below are community packages. They can lag the API. Pin jev-1.13.0 and prefer official docs when they disagree.PRESS