Jev Wiki

An unofficial, agent-maintained knowledge base about Jev, TypeSafe AI's System One model. Give it to your AI agent (Claude, Codex, Cursor, Kimi, DeepSeek, anything that can fetch a URL) so it can (1) tell you whether and where Jev fits your project, and (2) build with it: the exact HTTP contract, models and prices, SDK signatures, the three question primitives, confidence semantics, known failure modes, architectural patterns, all official cookbooks, and what the community has built.

If you are not technical, three things to know. Jev does not write text; it reads text (not images or audio) and picks an answer from options you define, with a probability. Your agent will read this wiki and hand you a ranked shortlist of places in your project where that helps, and where it does not. Nothing is installed unless you paste the install prompt below; the build prompt writes code, the consult prompt only reads.

It follows the Karpathy "LLM Wiki" pattern as used by agentwikis.com: refreshable raw sources in raw/ (never hand-edited), LLM-written pages in wiki/, a schema in CLAUDE.md, and every page served as plain Markdown with YAML frontmatter.

Ask your agent

Paste one of these into Claude Code, Codex, Cursor, or any agent that can fetch URLs.

"Could Jev help my project?"

Read https://jevwiki.ai/raw/wiki/ideas/consult.md and follow it exactly
(it tells you which further pages to open). Then look at my project (this repo / the description
below) and give me the ranked shortlist it asks for: which decisions Jev fits, the primitive and criteria
sketch for each, estimated cost, caveats, and what you would NOT use Jev for.
Read only the pages it routes you to (typically 30-50k tokens). Do not load llms.txt or llms-full.txt for this.

"Audit my existing code for Jev opportunities"

This one INSTALLS TypeSafe's agent skill into your project first (npx skills add ...; source: github.com/typesafe-ai/skills); the community prompt is via @k2sbhai (skill wording adapted so it works outside Claude Code).

npx skills add typesafe-ai/skills --skill typesafe-ai

Use the typesafe-ai skill (/typesafe-ai in Claude Code) to audit this project. Find every place where we make a slow or expensive LLM call
that is really a yes or no decision, a ranking, or a classification. For each one, tell me what Jev would
replace, what it would cost, and what would break if the answer is wrong. Then list 3 new features this
project could add if judgments were instant and nearly free. Before showing the list, check every item against
the fit test in https://jevwiki.ai/raw/wiki/ideas/consult.md and mark the ones that fail. Don't change any code yet.

Then always send this second prompt. The skill's list is a brainstorm; this step applies the wiki's fit test and adds what not to use Jev for:

Check each item against the fit test in https://jevwiki.ai/raw/wiki/ideas/consult.md and return its
shortlist format, including what you would NOT use Jev for.

"Build this with Jev"

Read https://jevwiki.ai/raw/wiki/guides/agent-integration-playbook.md, then one SDK page
(reference/python-sdk or reference/javascript-sdk). Open the HTTP API page only for raw HTTP, the
jaggedness page only when a decision touches numbers, dates, long state or adversarial input, and a
cookbook only if one matches. Then implement: <task>.
Use exact field names from the reference pages and gate actions on confidence.

For agents: start here

Assessing a project or building with Jev? Use the two entry pages in the note at the top instead. For anything else, start from the index:

curl https://jevwiki.ai/llms.txt

Recommended reading order for a coding task:

1. guides/agent-integration-playbook — decision tree, checklist, code templates (always)
2. reference/python-sdk or reference/javascript-sdk — the quick contract you will code against (always); internals pages only when you need headers, gateways or logging
3. reference/http-api — only if you call the API without an SDK
4. concepts/jaggedness-jev-1-13 — only when a decision touches numbers, dates, long state or adversarial input
5. the closest cookbooks/* page — only if one matches your task
6. guides/production-operations — before real users: versioning, caching, retries, monitoring, fallbacks

Assessing whether Jev fits a project at all? Start at ideas/consult instead (community tier: a pattern index plus six domain pattern pages, then evidence pages only when needed: builds, tools and integrations, head-to-head comparisons, field and failure reports). Not sure Jev is the right tool at all: syntheses/when-not-to-use-jev.

Wikilinks in raw pages look like [[concepts/confidence]] and resolve to /raw/wiki/concepts/confidence.md.

Trust semantics

• Scope is declared in llms.txt (Covers / Not covered / Current as of). Current as of gives two dates: the last complete refresh (every upstream source fetched) and, when newer, the latest refresh with the sources it could not fetch (usually only console.typesafe.ai, which answers 403). Anything newer than the latest refresh date: read the live docs at https://docs.typesafe.ai/llms.txt.
• Provenance: every page lists its sources: in frontmatter, pointing at files in raw/ (which map to upstream URLs via raw/MANIFEST.json). The full list for every page, ideas/ included, is in /index.json (our unpublished working notes under docs/ are left out).
• Tiers: pages under ideas/ are community-sourced (source_tier: community, confidence capped at medium). reference/ and concepts/ are TypeSafe's official documentation (docs.typesafe.ai, the OpenAPI spec and the SDK repos), distilled with a citation per fact: check claims there, and go to the live docs only for anything newer than the latest refresh date in llms.txt. Official pages win any conflict.
• Confidence: each page carries confidence: high|medium|low. Numbers, field names, and code come from the sources; inferences are marked "(inferred)"; marketing claims are attributed as claims.

Layout

CLAUDE.md / AGENTS.md   schema, page inventory, workflows (ingest / query / lint / refresh)
raw/                    immutable sources: docs pages (.md), OpenAPI, SDK repos, site, blog, evals, press
wiki/                   the knowledge base (concepts, reference, patterns, cookbooks, guides, ideas, entities, syntheses)
scripts/build.mjs       lint + build dist/ (HTML, raw copies, llms.txt, llms-full.txt, index.json, sitemap)
scripts/refresh.mjs     re-fetch sources, diff, list wiki pages that need re-ingestion
site/                   Cloudflare Worker (routing + content negotiation), stylesheet (theconsigliere.ai Brand Standards v2.0), mark and favicon (generated by scripts/make_brand_assets.mjs)

Maintaining it

npm install
npm run check      # lint: frontmatter, dangling wikilinks, orphans, required sections
npm run build      # produce dist/
npm run refresh    # re-fetch upstream, print a change report (raw/LAST_REFRESH.md)
npm run deploy     # build + wrangler deploy (Cloudflare Workers, static assets)

For maintainers with a checkout of the repo (not published on the site): point any agent at MAINTAINING.md and name the procedure: "run the health check", "add this post", "refresh upstream", "cold-test the consult route". CLAUDE.md is the schema it reads first. The permission scale in MAINTAINING.md §0 says what an agent may do unasked; npm run deploy refuses to ship a page that fails lint.

Credits

• Nate B. Jones — the four-placement routing on the pattern index, the UI-element pattern's design rules, the 50-case and shadow-mode check, and several first-hand measurements are digested, with his permission, from his member guide Find the Jev-shaped problems in your software. His audit prompt and recipe-card prompts live on his page, not here.
• Community patterns and repos are credited inline to the people who posted them. Official material is TypeSafe AI's.

Status

Unofficial. Not affiliated with TypeSafe AI, Inc. Source material © TypeSafe AI and the cited authors; wiki text is derived from it for reference use. Report mistakes by opening an issue.