$jevwiki.ai#an LLM wiki about Jev, written for agents rather than people

Agents: read the raw Markdown of this page, or start at llms.txt.

~/wiki/ideas

Consult guide: could Jev help this project?

[ mixed tier ][ guide ][ updated 2026-10-24 ][ confidence medium ][ jev-1.14.0 ]#consult · project-assessment · ideas · decision-mapping · token-budget

TL;DR Someone shows you a project and asks whether Jev fits. List its decisions, keep the narrow, semantic, frequent ones, match each to a pattern via the Decision patterns from the community (with fit verdicts) index and the domain pages holding your candidates, and return a ranked shortlist with cost, fit and caveats. Read what the routing below points to, as much as the verdict needs, not the whole corpus. "Jev does not help here" is a valid answer.

What to read

[[dir/slug]] → https://jevwiki.ai/raw/wiki/dir/slug.md. Skip llms.txt (~13k tokens) unless a page is unlinked here.

No token ceiling: read as much as a correct verdict needs; the stop rule below, not a token count, ends the reading. For planning: typically 30-50k with one or two domain pages and one evidence question (measured 2026-10-10: 31k for a shortlist, 47k when regulated data added the legal and alternatives pages); more for broad projects. Step 2b: open only domain pages holding candidates; read each candidate's entry before giving a verdict, not just its index row. For a supporting pattern (e.g. P06 under P23), the candidate entry's description is usually enough; open its domain page when the verdict depends on it.

Main steps (sizes for planning, not limits):

Step Page ~Tokens When
1 this page 6.3k always
2 Decision patterns from the community (with fit verdicts) 3.9k always
2b Patterns: agent internals, routing, gates, context and memory: agent loops, model/tool/skill routing, gates, context, memory (P01-P09, P35, P36) 5.0k if it holds a candidate
2b Patterns: coding agents, dev tools and self-compiling workflows: diff review, test output, self-compiling workflows (P10, P11, P37) 1.7k if it holds a candidate
2b Patterns: browser, computer use, voice and product UI: browser, desktop, mobile, voice, classifier in a UI (P12-P14, P34) 2.0k if it holds a candidate
2b Patterns: judging, search, documents, real-time and markets: judging, moderation (abuse, spam, UGC, user comments), search, labelling, documents, transcripts, devices, markets (P15-P18, P24-P27) 4.4k if it holds a candidate
2b Patterns: marketing, sales, GTM, content, support and ops: GTM, content, support, approvals, ops, operator alerts, incidents (P19-P23, P28-P33) 4.9k if it holds a candidate
2b Patterns: gates, simulation, personas and other shapes (P38+): wake gates, simulation, persona panels, grading, question tuning, poor-fit knobs, image evidence checks, find and locate (P38-P46) 4.7k if it holds a candidate
3 Jev 1.13 jaggedness: known failure modes 4.2k if any candidate touches numbers, dates, long state, or adversarial input
3b Advisor checklist: what to check before recommending Jev for a decision 1.6k always, before the shortlist
3c Failure reports: confidence misread and calibration 13.1k mandatory for a consequential gate (Rules for the advisor)
4 ONE evidence page per question the user actually asked (Step 4 section below; where a domain lists several pages, open the first and the next only if it did not answer; stop when answered) see below top one or two candidates, when a Seen line points there, or the user asks "does this really work?", "who has done this?", "is there a tool?", "is this endpoint safe?"
4b When not to use Jev: rules, embeddings, trained classifiers, small and frontier LLMs 6.7k when a rule, embedding, trained classifier or LLM might beat Jev for a decision
5 one official pattern or cookbook page named in the pattern's Map line 2-5k optional (stop rule); prefer a pattern page (2k) over a cookbook (4k)
5 alt Jev inside your agent harness: hooks, tools and fail-safes 8.7k instead of step 5, when the top candidate is a gate, router or screen inside a coding-agent harness: how to wire it, fail-safes
alt Choosing between Choice, Score, Noul 4.4k when a surviving decision matches no pattern

Follow-ups F1-F3 (after the shortlist; not part of step 4).

Follow-up Page ~Tokens When
F1 Running Jev in production: versioning, caching, retries, monitoring and fallbacks 10.5k rollout, monitoring, retries, production failures
F2 console.typesafe.ai (console + playground) 3.1k before promising access
F2 Legal: MCA, DPA, privacy, data retention 9.6k before sending confidential or regulated data
F3 Playbook for LLM agents building with Jev 13.0k once the user decides to build; then the SDK and HTTP reference

Stop rule (steps 3-3c still apply to surviving candidates first): you have enough when every surviving decision has a pattern ID (or "no match") and a fit verdict; answer then. Step 5 is optional: read it only when the user wants a design for the top candidate. Stop when the rule is met. Skip llms-full.txt, and the reference pages until follow-up F3, except two you may read during the assessment: Models, aliases, pricing, rate limits, context to check price, limits or input types, and Legal: MCA, DPA, privacy, data retention (F2) when the project would send personal or confidential data. SDKs: typesafe-sdk 0.7.5, @typesafe-ai/sdk 0.7.0 (check each cookbook's SDK version on Cookbooks overview; Python SDK changelog).

Procedure

  1. Get the project in front of you. Read the repo, spec or description. You need: inputs, outputs, where an LLM or a pile of ifs judges today, daily volume, latency needs.
  2. List the decisions, not the features. One line per judgment: "is this ticket urgent", "which tool applies", "does this diff need a human". Include decisions made today by LLM calls, regexes, keyword lists, heuristics or a human queue, AND judgments skipped as too expensive: sampling, coarse categories, delayed batches (jevify).
  3. Filter with the fit test below. Drop the ones that fail. Be strict; most value is in 2-5 decisions.
  4. Match survivors to patterns. Find the closest Pxx in the Decision patterns from the community (with fit verdicts) index, open the domain page holding it, and take its primitive, fit verdict and Map links. No match is fine: design it from Choosing between Choice, Score, Noul.
  5. Group, then estimate cost and latency. Group surviving decisions by shared state: every question about the same email, ticket or page goes in ONE call (Speculative fan-out); cost the call, not each decision. Input tokens per call = state + all questions; cost = tokens x $0.042 per million for jev-1.13.0, jev-1.14.0 and both aliases, output free. Worked example: 350-token email + 10 questions (~400 tokens) + ~260 fixed per request (builders' measurements, not in the docs: Request mechanics: billing, limits, latency, calibration and stability) = ~1,010 tokens/call; x 40,000 calls/day = 40.4M tokens/day = $1.70/day, about $51/month. This is an estimate: budget production from usage.input_tokens on representative requests; through a gateway, check its own price (Platforms and gateways: Zapier, Cloudflare, Netlify, Vercel AI Gateway, OpenRouter, Pydantic AI Gateway, Opper, Fly.io, OpenCode Zen and other hosted routes to Jev). Check volume against the limits: rate limits are set per organization and shown in the console since 2026-10-14 (the docs published 100K tokens/s and 80 requests/s until then); 64k per request, 32k for state plus the longest question by default, 262k for both with the experimental long-context header (Models, aliases, pricing, rate limits, context). Over the ceiling? Count distinct items (cache by merchant string, page hash, tool name) first. TypeSafe states 70-500 ms per call (launch blog) and "about 100 ms" for most queries (How to build software with System One); community reports vary (Field reports: what held up, what broke, what it means when advising, Request mechanics: billing, limits, latency, calibration and stability).
  6. Return the shortlist below, best first, and say what you would not use Jev for.

Fit test

Signal Verdict
Output is one of a known set, a yes/no, or a rating on a rubric required; otherwise drop this decision
Needs semantic understanding that rules or regex handle badly strong
Happens often (per message, per step, per row) or sits on a latency-critical path strong
You can act differently on low confidence (escalate, ask, skip) strong; see Confidence-gated routing
Replaces an LLM call whose answer is parsed into a branch strong
Needs arithmetic, counting, date/time comparison, exact thresholds on numbers poor; compute in code, ask Jev only the semantic part
Needs several hops of inference or implicit intent workable only if split into literal questions
State is large and mostly irrelevant workable only after filtering state
Needs generated text, explanations, code, summaries not Jev; use an LLM, optionally with Jev as router/judge around it (Jev vs LLM JSON mode / structured outputs)
A simpler tool already works: a rule, a cache, embeddings, a classifier they own keep it
It is one question, asked once skip; setup is only repaid by repetition
The answer depends on facts the state will not contain put retrieval/lookups into state first; never expect Jev to know the world (@TheIshanGoswami reports this, fixed by Exa retrieval — unverified)
Wrong answers are costly and there is no review path poor for autonomous action until a safe fallback (abstain, hold, rule check or a person) or an independent control exists; advisory use stays workable with confidence gating and an eval set (Testing and evaluating a Jev workflow)

Rules for the advisor

Shortlist format

Return a table, then three short paragraphs. Tag each row saving, better outcome or new capability. Fit is strong (bounded options, semantic judgment, code owns the rest; still starts with the experiment below, never straight to production), workable (only with the stated mitigation) or poor (do not; say what to use instead). When several rows share one call, put the cost on one "shared call" row and write "shared" in the others.

# Decision in this project Pattern Ask (primitive + criteria sketch) Code does Fit Est. cost / volume Caveat

Step 4: evidence pages, by question

Open a builds page only if the pattern's Seen line does not answer "who built it / did it work". All community tier.

Who built it / is there a tool? Pick by domain, first page first: coding agents, harnesses, memory → Builds: coding agents, harnesses and orchestration, Repos: coding agents, orchestration, memory, browser and computer use, integrations, Tools: CLIs, linters, semantic grep and dev tools (wiring it into a harness is step 5 alt, not evidence); gates, approvals, model routers → Builds: permission gates, approvals and model routers for agents, Tools: guardrails for agents (tool-call gates, permission hooks, injection screens, rule checks), Tools: agent routing, context and skill selection; browser and UI → Builds: browser, computer use and interface, Builds: interface elements, Repos: coding agents, orchestration, memory, browser and computer use, integrations; desktop, mobile, voice → Builds: desktop, mobile and voice computer use; search, logs, documents, fraud → Builds: data, search and business, Repos: data, documents, judging, real-time, markets, business apps, replicas; support, sales, recruiting, finance, risk, markets → Builds: business ops and markets; writing, ads, inboxes, calls, video → Builds: content, media, email, calls and reading; moderation and judging → Repos: data, documents, judging, real-time, markets, business apps, replicas, content screens → Tools: guardrails for agents (tool-call gates, permission hooks, injection screens, rule checks); games, robots, devices → Builds: games, simulators, robots and devices, Repos: data, documents, judging, real-time, markets, business apps, replicas; MCP servers, plugins, SDKs → Tools and integrations hub: MCP servers, plugins and connectors, Community SDKs and libraries.

Does it win? Pick ONE comparison page by topic: games, simulators, real-time control → Head-to-head: games, simulators and real-time control; security, injection, spam, phishing → Head-to-head benchmarks: security (phishing, spam, injection, vulnerable code); agents, routers, tool or approval gates → Head-to-head: Jev inside agents, routers and tool gates; public datasets and suites → Head-to-head benchmarks: Jev on public datasets and suites (reasoning, agent, robot, science and medical suites → Head-to-head benchmarks: reasoning, agent, robot and science suites; retrieval, reranking, screening and judging sets → Head-to-head benchmarks: retrieval, reranking, screening and judging); vs small, fast or local models → Head-to-head: Jev vs small, fast and local models; other vendors' decision models (Clef, Perplexity, GLiDE, Mercury) → Head-to-head: Jev vs rival decision models; vs OpenAI's Decisions API → Head-to-head: Jev vs OpenAI's Decisions API; other app-level (frontier LLMs, classic methods, cascades) → Head-to-head: Jev against other models and methods.

Question Page What it holds ~Tokens
Does it work? Field reports: what held up, what broke, what it means when advising first-week independent tests: what held up, what broke, what it means when advising. Spam, phishing, alerts: also "A simpler tool won" on Failure reports: where Jev broke, lost, or was the wrong tool and Head-to-head benchmarks: security (phishing, spam, injection, vulnerable code). Harassment moderation: no head-to-head numbers exist; the nearest public set is ToxicChat's jailbreak labels in the Bonn 37-dataset suite (Head-to-head benchmarks: Jev on public datasets and suites); for toxicity only jev-cookbook's small hand-made recipe, no baseline (Builder lessons: what changed the result); Jev-Moderation-Bot publishes none 6.9k
Builder lessons: what changed the result before/after runs: what changed the result (context in state, wording, conjunctions, cascade gates) 12.6k
Where did it break? Failure reports: where Jev broke, lost, or was the wrong tool failures grouped by kind, each matched to a documented weakness 10.8k
Failure reports: wrong role, or options that don't fit Jev given the wrong job (planning, investigating, subtext, forecasting) or options that overlap, miss the answer or describe themselves; trimmed or packed state 8.0k
Failure reports: confidence misread and calibration confidence or probabilities that misled a gate; thresholds that don't transfer between datasets; which signal to gate on, per task 13.1k
What did people measure? Measurements and access routes measured cost, latency, calibration, token accounting, vendor figures (incl. inbox triage with a spam question, compliance alerts); access routes 6.0k
Latency, limits, billing? Request mechanics: billing, limits, latency, calibration and stability how a request is billed, capped and timed; overhead per request, max_tokens_exceeded, packing, batching, repeat stability 9.4k
How does it compare? Head-to-head: Jev against other models and methods app-level comparisons vs frontier LLMs and classic methods, including losses; cascades 13.8k
Head-to-head: Jev vs small, fast and local models vs small and fast LLMs and local or open models 14.8k
Head-to-head: Jev vs rival decision models independent runs against rival decision models (Clef, pplx-decider, GLiDE, Mercury Decide, StartLux) 11.3k
Head-to-head: Jev vs OpenAI's Decisions API builder runs of Jev vs OpenAI's Decisions API (GPT-6 Luna) 7.9k
Head-to-head: Jev inside agents, routers and tool gates whole-agent A/B tests with and without Jev 18.5k
Head-to-head benchmarks: Jev on public datasets and suites builder runs on public datasets (intent, text classification, legal) 14.1k
Head-to-head benchmarks: reasoning, agent, robot and science suites reasoning, agent, robot, science and medical suites (WebArena, ALFWorld, GPQA, MedQA, System1-Arena) 11.9k
Head-to-head benchmarks: retrieval, reranking, screening and judging retrieval, reranking, screening and LLM-as-a-judge sets (BEIR, ESCI, CLEF TAR, MLflow, CMU) 9.8k
Head-to-head benchmarks: security (phishing, spam, injection, vulnerable code) spam, phishing, injection and vulnerable-code sets 11.0k
Head-to-head: games, simulators and real-time control games and real-time control vs chat models, engines, local replicas 7.6k
Eval boards: JevBench by version, Jevals.com, typed-decisions and other independent suites independent boards: JevBench by version, Jevals.com, the typed-decisions card, replicas scored against live Jev 9.5k
What does a decision cost? Cost ledger: published cost per Jev decision builders' published cost per decision, reconciled against the list price, with comparator costs 13.4k
Who built something like it? Builds: coding agents, harnesses and orchestration coding agents, harnesses, orchestration, memory, chat clients 9.2k
Builds: permission gates, approvals and model routers for agents tool-call, shell and edit approval gates; model, effort and agent routers 4.7k
Builds: browser, computer use and interface next browser action or site tool 4.5k
Builds: interface elements judgments inside a UI element: extensions, generative UI, canvases 6.7k
Builds: desktop, mobile and voice computer use desktop, mobile and voice control 8.0k
Builds: data, search and business search, logs, alerts and incidents (log paging, security-alert triage, headline alert levels), documents, research feeds, fraud, expenses 12.6k
Builds: business ops and markets support and inbox routing, sales and leads, recruiting, reconciliation, risk sweeps; market demos (engineering only) 13.6k
Builds: content, media, email, calls and reading writing checks, ads, inboxes (spam filters, labels), chats, calls, video, images, reading tools 16.8k
Builds: games, simulators, robots and devices games, simulators, robot arms, drones, test harnesses 11.7k
Community repos: what people built and how they use Jev the curated first-week repos, one line each; reusable designs; vetting checklist 3.2k
Repos: data, documents, judging, real-time, markets, business apps, replicas repo table (links, stars, what each does, pattern map): search, documents, moderation and judging (e.g. Jev-Moderation-Bot, a Discord spam/scam filter), real-time and IoT, markets, business apps, replicas 5.3k
Repos: coding agents, orchestration, memory, browser and computer use, integrations the same for coding agents, browser and computer use, orchestration and memory, integrations 6.9k
Is there a tool? Tools and integrations hub: MCP servers, plugins and connectors hub: MCP servers, plugins, connectors, key safety, name collisions 5.5k
Tools: guardrails for agents (tool-call gates, permission hooks, injection screens, rule checks) tool-call gates, permission hooks, injection and content screens, review gates 9.0k
Tools: agent routing, context and skill selection model, tool and skill routers; session and context pruning; wake gates 14.7k
Tools: CLIs, linters, semantic grep and dev tools CLIs, REPLs, linters, semantic grep and sort, pipeline tools 14.4k
Community SDKs and libraries community SDKs by language, libraries, SQL functions (these call the real Jev) 9.2k
Already on a framework? Framework integrations: Pydantic AI, LangChain, Spring AI, Vercel AI SDK, eve, LiteLLM, BAML, TanStack AI, DSPy, Langfuse and other framework-shipped Jev support framework-shipped Jev support (Pydantic AI, LangChain, Spring AI, Vercel AI SDK, eve, LiteLLM, BAML, TanStack, DSPy, Pi) 18.3k
Apps with a built-in Jev option, if the user's app already has a Jev option Langfuse, MLflow, n8n, Inbox Zero, MotherDuck 8.4k
Already on a platform? Is this vendor gateway or reseller safe? Platforms and gateways: Zapier, Cloudflare, Netlify, Vercel AI Gateway, OpenRouter, Pydantic AI Gateway, Opper, Fly.io, OpenCode Zen and other hosted routes to Jev vendor-shipped hosted routes (Zapier, Cloudflare, Vercel AI Gateway, OpenRouter, Opper...); each with its own key, billing and terms; keyless "Jev APIs": next row 11.5k
Is this named third-party "Jev API", or a free, cheap or keyless endpoint, safe? Is this install name real? Warnings: not-Jev services, key safety, look-alikes and install names the wiki's one Warnings block: keyless "Jev APIs" on someone else's key, Jev answers as training data, key safety, servers answering to jev-latest, install names; the exclusion policy 15.4k
Is there a rival decision model? Rival decision models: Clef, Liquid d1, GLiDE, Drex and other non-Jev System One models other vendors' decision models (Clef, GLiDE, pplx-decider, Strands Decider, d1, Drex, Instinct): when a rival fills a gap Jev has (production image input, input past 64k, local weights; Jev's image and long-context options are experimental); vendor numbers here, independent comparisons on Head-to-head: Jev vs small, fast and local models 15.0k
Can it run locally or on-prem? Open replicas and Jev-compatible servers, only if the user names a specific replica Answer here: TypeSafe's captured docs mention no on-premises, self-hosted or VPC deployment (checked 2026-09-23, FAQ for agents and developers Q43; ask sales@typesafe.ai); open replicas are not Jev. The page: which is which, their own numbers, trust notes 12.4k
Open replicas: servers and logit readers over other models, only if the user names a specific server servers and logit readers wrapping other models (logits, NLI, embeddings); also not Jev 7.1k
Open replicas: the Laya family, only if the user names Laya or a Laya server Laya, its servers, ports and fine-tunes; also not Jev 5.1k

Sources