Explore the catalogue ↗ First call Adapt a project Independent reports 中文
Counts describe saved link and evidence records, not current CI passes or runtime tests. About these counts
On this page · full reading map
Patterns · Compatibility · Vetting
- 01 What this is
- 02 What Jev returns
- 03 Start here
- 04 Coverage
- 05 Measured, not claimed
- 06 By decision pattern
- 07 By resource kind
- 08 Also in this repo
- 09 What is verified, and what is not
- 10 Machine-readable data
- 11 Contributing and licence
What this is
- Jev is a decision model from TypeSafe AI. It does not write text — you hand it state plus typed questions and it returns typed answers with calibrated confidence, fast and cheap enough to sit in an agent's inner loop.
- This repo indexes public examples of using it, organised by the decision being made. The resource you read this week is disposable; the decision pattern is not.
- How to assess it: every row names its source. Call-site citations, primitive claims and caveats are recorded where available, so you can inspect what was read and what remains untested.
[!NOTE]
Not the product, not an SDK, not affiliated with TypeSafe AI, and not a recommendation. Inclusion is a source record, not a runtime or performance endorsement. See what is verified.
What Jev returns
Three primitives. Every pattern below is built out of them, and the asymmetry in the last row is the single most common source of bugs.
Under each primitive, two counts that are never added together: rows whose question_types records a person reading the code call it, and rows where only a text signal shows it — the weekly refresh found its request or answer shape in the one file the row cites (primitives_seen), which does not show that the code calls it. See what is verified.
Input is text only — string, JSON object, or array of text. Context is 64k tokens per request, 32k for the state plus the longest question. Output tokens are free. There are no published weights, so it cannot be run locally. Full cross-platform differences: docs/compatibility.md.
Which primitive fits a decision? The figure below arranges TypeSafe's own guidance as a decision list: read down and stop at the first yes. The order and the wording are this repository's; every step cites the page it follows, listed under the figure. It is the vendor's design guidance with its sources, not a recommendation of any row here.
Beside each primitive, the rows filed under the pattern named there, counted as the figure above counts them: read by a person, or text signal only. The steps live in picker.json, which the site's primitives view reads too. Sources, numbered as in the figure: [1] docs.typesafe.ai/model-jaggedness/jev-1.13#generation · [2] docs.typesafe.ai/concepts/how-to-build-with-system-one#use-code-when-you-can · [3] docs.typesafe.ai/primitives#split-a-complex-judgment-into-several-questions · [4] docs.typesafe.ai/primitives/noul#good-practice-ask-more-than-one-question-per-call · [5] docs.typesafe.ai/model-jaggedness/jev-1.13#common-sense-structural-invariants · [6] docs.typesafe.ai/primitives#choose-a-question-type · [7] docs.typesafe.ai/primitives/score#writing-good-levels · [8] docs.typesafe.ai/primitives/noul#writing-a-noul-question · [9] docs.typesafe.ai/primitives#ask-for-one-snap-judgment-per-question.
Start here
Six things in reading order. Hand-picked, because "most starred" is not the same as "read this first".
The canonical first call: one support ticket, one Choice, one Score and one Noul in a single request, in Python, JS and cURL. The most useful page in the docs and the least linked. It explains, among other things, that a Choice over options and one Noul per option answer different questions. Written from the official API reference and checked field by field against it, but not executed against the live API. Exactly two nouls per tool call: does knowing this call happened still matter, and is the full output still needed verbatim. Despite the word "scored" in its own description, no score primitive is used. The best structured tutorial found. It states plainly that typed output does not guarantee a correct decision, lists the documented weaknesses, and qualifies its own cost illustration rather than selling it. The single most credible row in this catalog. Recall came out below their existing summariser, and at a matched context budget it tied plain recency ordering. Cost was genuinely far lower. Publishing a negative result on a hyped model is rare.Coverage
Every decision pattern, sized by how many entries this catalogue contains. Use the pattern index below the chart to jump to a section. A zero is a research gap, not a rendering bug.
All 18 patterns have at least one catalogue entry. Coverage does not imply runtime testing or equal maturity. See docs/status.md.
Pattern index · jump to the examples
| Decision pattern | Decision pattern | | :--- | :--- | | Tool selection · 231 | Intent routing · 37 | | Context compaction · 35 | Safety gating · 140 | | Output validation · 135 | Retry control · 7 | | Human escalation · 69 | Model routing · 45 | | Speculative fan-out · 32 | Search & ranking · 64 | | Structured extraction · 17 | Classification · 120 | | ML feature extraction · 8 | Document triage · 20 | | Support triage · 9 | Content scoring · 166 | | Recommendation · 1 | Overview · 453 |
Measured, not claimed
Independent measurement reports in the catalogue, including negative results that help explain where an approach fails. These are the original authors' measurements; this repository has not independently reproduced them. Check each report's dataset, method and model version before comparing results.
Author's conclusion is the direction a benchmark's own author states for Jev on the task they measured (measurement.direction: favourable, mixed, unfavourable or inconclusive), indexed from the author's report: author-stated, not reproduced here, and absent where the author states none in words. docs/benchmarks.md sets every benchmark's measurement side by side.
Negative results first
Rows whose own author measured Jev for the use and concluded against it: a benchmark whose measurement's direction is unfavourable, or another row flagged measured, not adopted. Author-stated, not reproduced here. Read them before the positive examples; the site lists them.
Ported the Jev compaction approach, measured it against their shipping summariser, and published the conclusion not to adopt it.Benchmark · ★100k+ · Py · noul · call site, read 2026-09-22 · author's conclusion: unfavourable (author-stated, not reproduced here)
> The single most credible row in this catalog. Recall came out below their existing summariser, and at a matched context budget it tied plain recency ordering. Cost was genuinely far lower. Publishing a negative result on a hyped model is rare.
Two Choice questions over threat level and category, held in shadow mode after a blind evaluation found Jev merely tied the incumbent model.Benchmark · ★10k+ · TS · choice · call site, read 2026-09-22 · author's conclusion: unfavourable (author-stated, not reproduced here)
Caveats: shadow mode
> Wired in but deliberately inert: by their own statement nothing Jev returns reaches a label, a cache row or an alert. Ships a golden fixture. A model to copy for how to trial a new model without betting production on it.
Nine agent skills plus a CLI covering model routing, memory filtering, turn retention, one-of-many skill selection and next-action choice.Plugin · ★1k+ · Py · choice · score · noul · call site, read 2026-09-22
Caveats: measured, not adopted
> Notable for publishing a use it dropped: Jev-summarised handoffs had worse recall than raw transcripts.
One Score per candidate file to pre-brief code review — measured twice, then removed: more billed input for essentially no wall-clock gain, and offline replay showed the candidate list could not reach where review findings land.Benchmark · ★1k+ · Go · score · author's conclusion: unfavourable (author-stated, not reproduced here)
> Removed in PR #1165 (2026-09-22). Their offline measurement found the candidate generator excluded changed files by construction while nearly all review findings sit in changed files, and that per-file excerpts made the list less precise at higher token cost. The code is gone from the default branch, so this row cites the change that removed it.
Claude Code plugin: asks TypeSafe Jev which installed skill fits each prompt and logs the answer (shadow-first). A working reference for the skill-suggestion cookbook on Claude Code — the README records why it is unlikely to help a strong model as a router. (earlier upstream description)Plugin · ★10+ · shimo4228 · Py · call site, read 2026-09-22
Caveats: measured, not adopted
> Its author ran it on 2026-09-21 and concluded that, as a router, it is unlikely to help a strong model, which already sees every skill's description. Its README: 6 of 6 scripted requests handled sensibly on 0.2.0, 3 of 6 real-session prompts wrong on 0.1.0, which the author calls an anecdote, not a rate. It stays in shadow mode. https://dev.to/shimo4228/i-added-jevs-skill-router-to-claude-code-and-turned-back-just-before-rewriting-the-skill-listing-34in A model wrote this row's Chinese note.
Other independent reports
The best independent test found: 24 Norwegian documents on one pinned model version, opening with a case the model got wrong while correctly reporting low confidence.Benchmark · Lindfors
> Methodology is stated cleanly and scoped honestly as a single-day snapshot. Leading with a failure case is what makes it a real calibration test rather than a testimonial.
The only three-way head-to-head found, with each model's prompt tuned separately and the scope limited to one task rather than a general ranking.Benchmark · Near Here
> Self-limits correctly: a use-case study, not a model leaderboard. That restraint is rarer than the numbers.
Biologically-inspired memory for AI agents. Decay, retrieval strengthening, consolidation. Zero runtime deps, SQLite, MCP. Benchmarked retrieval with an opt-in TypeSafe Jev reranker.Benchmark · ★100+ · kitfunso · TS · call site, read 2026-09-22 · author's conclusion: mixed (author-stated, not reproduced here)
An introduction to Jev with hands-on tests: Choice, Score and Noul turn natural language into typed judgements for classification, scoring and routing, compared with DeepSeek on comment labelling, speed and results, with CSV import, replay and offline reports.Benchmark · ★100+ · nanmicoder · JS · call site, read 2026-09-24
JevBench v1 - a benchmark for Jev-class typed decision models: smart, cheap, fast, reliable, open. (earlier upstream description)Benchmark · ★100+ · fstandhartinger · Py · call site, read 2026-09-22
10 of 74 shown: the negative results first, then the picks of the curated independent reports path, in its order, then the first of the others in list order · all 74 on one page, with every note → · filter on the site
By decision pattern
The primary index. Each heading is a decision an agent has to make; the rows are examples of making it. Caveats appear as short tags — the full note for each row is in catalog.json and on the site.
★ gives a repository's GitHub stars as a band — ★10+, ★100+, ★1k+, ★10k+ and ★100k+; rows with no repository or under 10 stars show no band. Rows run official first, then with code, then by band, then by title. A band is a popularity signal, not a quality verdict; the exact count, as last read from GitHub, is in catalog.json and on the site.
A call site link opens the one file a row cites (evidence.path) at HEAD of the repository's default branch; the date after it is the day a person last read that file (evidence.read_on): a reading, not a run of the code. A cited file link is the same for a file that shows the project speaking Jev's request shape rather than building on Jev, or only an example it ships (evidence.kind). Neither is pinned to a commit, so it opens the file as it is now, which may differ from what was read, and stops resolving once the file moves; the weekly claims check reports that.
Tool selection
_Which tool or action the agent should call next._
Maps natural-language trading requests onto ordinary typed functions by turning function names and closed-set arguments into confidence-aware questions.Official docs · Py · choice
Picks at most one skill out of 182 for an agent turn: one request ranks every skill and asks whether the turn needs one at all, a second reads the top three.Official docs · Py · choice · noul
Runnable demo code for a smart home assistant that evaluates user requests with typed decisions.Official docs · Py
An AI Hedge Fund Team (upstream description)Integration · ★10k+ · virattt · Py · call site, read 2026-09-24
Three independently installable Claude Code plugins — guardrails, model router and skill suggestion — each with its own hooks and tests.Plugin · ★10k+ · Py · TS · choice · score · noul · call site, read 2026-09-22
Compiles a tool catalogue into questions and reconstructs tool calls from the answers, with typed errors for abstention and confirmation-required cases.Project · ★10k+ · Py · choice · call site, read 2026-09-22
Computer-use action selection in Python and TypeScript: Jev picks the next browser action from an immutable candidate set, with reobserve and abstain as reserved options.Project · ★10k+ · Py · TS · choice · call site, read 2026-09-22
Two-stage MCP tool search: a wide Choice coarse-ranks the whole catalogue, then a shortlist gets full descriptions plus one Noul each to decide whether it does the job at all.Project · ★10k+ · Py · choice · noul · call site, read 2026-09-22
A high-speed browser agent from Browser Use: Jev decides the operation and which element to act on, and a small LLM is called only when text must be typed.Project · ★10k+ · Browser Use · Py · choice · call site, read 2026-09-22
Caveats: vendor numbers
Project · ★10k+ · Vercel Labs · TS · choice · call site, read 2026-09-22
10 of 231 shown · all 231 on one page → · filter on the site
Intent routing
_Classify what the user wants and send the request down the right branch._
Runnable demo code for a smart home assistant that evaluates user requests with typed decisions.Official docs · Py
Treat confidence as a second axis: the answer tells you what, the confidence tells you whether to act on it.Official docs · Py
Classify an incoming request and route it to the cheapest adequate handler: deterministic code, a specialist LLM, or a person.Official docs · Py · choice
Seven production blocks — choice, score, yes/no, ask-many, route, pick-best, filter — with a UTF-8 byte budget, verbatim wire capture and eleven test files.Project · ★100k+ · Py · choice · score · noul · call site, read 2026-09-22
Turns downstream task ids into a choice option set, with a minimum-confidence gate that routes uncertain runs to a human.Integration · ★10k+ · Py · choice
Seven distinct email decisions, each with its own separately chosen threshold, falling back to the normal LLM on any error.Project · ★10k+ · TS · choice · noul · call site, read 2026-09-22
A graded course from a first call through each primitive, state shapes and criteria, to ticket triage and a multi-step workflow, mirroring all four official patterns.Tutorial · ★1k+ · Py · choice · score · noul · call site, read 2026-09-22
An Android reply co-pilot that judges intent, timing and risk from on-screen text, while separate models handle OCR and drafting.Project · ★1k+ · Java · choice · score · noul · call site, read 2026-09-22
A teaching example with a deliberate control group: the same station-enquiry task written in plain Python that only accepts Y/N, next to a Noul that reads intent.Tutorial · ★1k+ · Real Python · Py · noul · call site, read 2026-09-22
A software-factory foreman that uses Jev to decide what an agent pipeline should do next.Project · ★100+ · thruwire · Py · call site, read 2026-09-22
10 of 37 shown · all 37 on one page → · filter on the site
Context compaction
_Decide which tool calls and results still matter so stale context can be dropped._
Ported the Jev compaction approach, measured it against their shipping summariser, and published the conclusion not to adopt it.Benchmark · ★100k+ · Py · noul · call site, read 2026-09-22 · author's conclusion: unfavourable (author-stated, not reproduced here)
Replaces the whole retrieval stack for memory recall — no embeddings, no BM25, no reranker — with one batched Noul per candidate memory.Project · ★10k+ · Rs · noul · call site, read 2026-09-22
A Claude Code plugin that replaces the compaction summary with per-item decisions: stale tool calls are dropped or truncated, everything kept stays verbatim.Plugin · ★1k+ · tamaratran · TS · noul · call site, read 2026-09-22
Nine agent skills plus a CLI covering model routing, memory filtering, turn retention, one-of-many skill selection and next-action choice.Plugin · ★1k+ · Py · choice · score · noul · call site, read 2026-09-22
Caveats: measured, not adopted
Project · ★100+ · kunchenguid · TS · call site, read 2026-09-22
Trims long shell output before the model sees it, asking one Noul per chunk.Plugin · ★100+ · tamaratran · TS · noul · call site, read 2026-09-22
Coding agent and desktop app built on pi that asks Jev at 38 decision points which tool-output chunks enter the context, which stale tool results to drop, whether a rule-flagged command was asked for and whether fetched pages or MCP output carry injected instructions.Project · ★100+ · qybaihe · TS · choice · noul · score · call site
Caveats: self-submitted
Plugin · ★100+ · Py · noul · call site, read 2026-09-22
Claude Code plugin: Jev for rule checks, verbatim compaction, and prompt routing (upstream description)Plugin · ★10+ · 0x7067 · Py · call site, read 2026-09-22
Jev judgment, not generation: prune long tool output, screen fetched pages for injected instructions, and gate completion claims inside DeepSeek Harness. (upstream description)Plugin · ★10+ · horusjiang · TS · call site, read 2026-09-24
10 of 35 shown · all 35 on one page → · filter on the site
Safety gating
_Decide whether an action is safe to run. Defence in depth, never a security boundary._
Scores each retrieved passage, then decides in code which reach the answering model — keeping contradictory ones flagged and dropping ones carrying prompt injection.Official docs · Py
Screens every message in and out of an LLM app in one request, naming hazards and scoring how much harm complying would do.Official docs · Py · noul · score
The JavaScript counterpart of the LangChain integration, with the same classifier and middleware shapes.Integration · ★10k+ · TS · choice · score · noul · call site, read 2026-09-22
Three independently installable Claude Code plugins — guardrails, model router and skill suggestion — each with its own hooks and tests.Plugin · ★10k+ · Py · TS · choice · score · noul · call site, read 2026-09-22
Drops in as a moderation API by asking many parallel Noul questions in one request, one per hazard category, with an anti-injection prefix on every instruction.Project · ★10k+ · Go · noul · call site, read 2026-09-22
Three Score questions on a shared severity scale, blocking the request when two or more cross the line, and failing closed.Project · ★1k+ · Rs · score · call site, read 2026-09-22
Reviews each tool call on three axes — risk level, whether the user authorised it, and an explicit prompt-injection pressure check.Project · ★1k+ · TS · choice · noul · call site, read 2026-09-22
The verifiable coding agent runtime. Define your coding agent's process in natural language with stages, checks, and approval gates instead of hoping it follows your instructions.Project · ★100+ · bastani-inc · TS · call site, read 2026-09-24
Connect TypeSafe Jev to Grok Bot as a cheap decision layer - usage gates, skill template, examples (upstream description)Plugin · ★100+ · bodila51 · Py · call site, read 2026-09-22
A computer-use agent that asks which accessibility-tree element to act on, plus a separate noul for whether the action needs explicit user confirmation.Project · ★100+ · JS · choice · noul · call site, read 2026-09-22
10 of 140 shown · all 140 on one page → · filter on the site
Output validation
_Check a model's output against a rubric before it reaches a user._
Catches wrong or invented citations against the source document with one Choice, using its confidence to flag borderline cases for review.Official docs · Py · choice
Screens every message in and out of an LLM app in one request, naming hazards and scoring how much harm complying would do.Official docs · Py · noul · score
Open-source observability for AI agents. Find where your agents fail, dispatch your coding agent to fix it, and verify the fix against real traces. (upstream description)Project · ★1k+ · latitude-dev · TS · call site, read 2026-09-22
AI agents can generate code, but still struggle to understand what they build. Reticle brings Jev-style machine-native runtime perception to web & desktop applications. (upstream description)Project · ★1k+ · reticlehq · TS · call site, read 2026-09-24
Make your coding agent abide by all your project rules (upstream description)Plugin · ★100+ · coldteadotai · TS · call site, read 2026-09-24
The verifiable coding agent runtime. Define your coding agent's process in natural language with stages, checks, and approval gates instead of hoping it follows your instructions.Project · ★100+ · bastani-inc · TS · call site, read 2026-09-24
Guards against a coding agent claiming it finished: reads tool output, the diff and test results, then judges whether the completion claim holds.Project · ★100+ · TS · noul · score · call site, read 2026-09-22
A fast browser agent: Jev picks each action from what is on the page, an LLM reads and plans, and every claim in an answer cites a quote from the page. (upstream description)Project · ★100+ · agent-labs-dev · Py · call site, read 2026-09-22
Submit Forma Plugin · ★100+ · timrogers · Rs · call site, read 2026-09-22
An agent that routes evaluation work through typed decisions.Project · ★100+ · vinilana · TS · call site, read 2026-09-22
Caveats: no licence
10 of 135 shown · all 135 on one page → · filter on the site
Retry control
_Decide whether a failed step is worth retrying._
Zero-dependency System One decision harness: 5 semantic gates saving frontier AI agent tokens on trivial errors & doom loops. Python + TypeScript + Rust. MCP-compatible. (upstream description)Plugin · ★10+ · ismaelsoilet · Py · call site, read 2026-09-24
Judge agent steps — ok / retry / escalate / stop via TypeSafe Jev (upstream description)Project · ndolinschi · TS · call site, read 2026-09-22
Caveats: one commit · no licence
Project · Really Artificial · JS · choice · score · noul · call site, read 2026-09-22
Caveats: AI-written
Project · andreuvm · Py · call site, read 2026-09-24
Caveats: no licence
Plugin · vicente-md · Java · call site, read 2026-09-22
Caveats: no licence
SDK · nsstudent · Swift · call site, read 2026-09-22
A local decision layer in Jev's shape: yes/no, choice and rubric questions read off one token's logprobs from a local model, with a Claude Code permission gate measured on held-out command sets.Jev-like alternative · Xinyu Liu · TS · noul · choice · score · cited file, read 2026-09-25
Caveats: not Jev itself · AI-written
All 7 shown · on its own page · filter on the site
Human escalation
_Use calibrated confidence to decide what a person must see._
Classifies annual reports into 75 industry groups, then reads the answer's own confidence to decide whether to report that group or the broader division above it.Official docs · Py · choice
Catches wrong or invented citations against the source document with one Choice, using its confidence to flag borderline cases for review.Official docs · Py · choice
Decides which of 450 candidate pairs from two product catalogues describe the same thing, with one Score whose three levels are the three available actions.Official docs · Py · score
Adds an explicit "uncertain" outcome to moderation decisions and measures label agreement against the share of actions taken automatically.Official docs · Py · choice
Routes uncertain probabilities to human review while keeping the underlying noul values visible rather than collapsing them to a label.Official docs · Py · noul
Treat confidence as a second axis: the answer tells you what, the confidence tells you whether to act on it.Official docs · Py
How confidence is derived from the probability distribution, and why a threshold tuned on one question type does not transfer to another.Official docs
Turns downstream task ids into a choice option set, with a minimum-confidence gate that routes uncertain runs to a human.Integration · ★10k+ · Py · choice
Compiles a tool catalogue into questions and reconstructs tool calls from the answers, with typed errors for abstention and confirmation-required cases.Project · ★10k+ · Py · choice · call site, read 2026-09-22
Seven distinct email decisions, each with its own separately chosen threshold, falling back to the normal LLM on any error.Project · ★10k+ · TS · choice · noul · call site, read 2026-09-22
10 of 69 shown · all 69 on one page → · filter on the site
Model routing
_Pick which downstream model or tier should handle a request._
A two-stage mini-then-verify-then-reasoning cascade that reaches most of a big reasoning model's quality at a fraction of the cost.Official docs · Py
Classify an incoming request and route it to the cheapest adequate handler: deterministic code, a specialist LLM, or a person.Official docs · Py · choice
The JavaScript counterpart of the LangChain integration, with the same classifier and middleware shapes.Integration · ★10k+ · TS · choice · score · noul · call site, read 2026-09-22
Three independently installable Claude Code plugins — guardrails, model router and skill suggestion — each with its own hooks and tests.Plugin · ★10k+ · Py · TS · choice · score · noul · call site, read 2026-09-22
Nine agent skills plus a CLI covering model routing, memory filtering, turn retention, one-of-many skill selection and next-action choice.Plugin · ★1k+ · Py · choice · score · noul · call site, read 2026-09-22
Caveats: measured, not adopted
Plugin · ★100+ · miuuyy · JS · call site, read 2026-09-24
Judges how hard a coding turn is, then picks the model tier, reasoning depth and speed mode to match.Plugin · ★100+ · JS · choice · score · call site