git:20261029.cae8289 to git:20261030.31f479d

# awesome-jev

Jev (https://docs.typesafe.ai/) returns typed decisions rather than text: a `choice` among up to 255 options, a `score` on an ordered 2-to-10 level scale, or a `noul` yes-no probability. This repository is an index of how people use it. It is not the product, not an SDK, and not affiliated with TypeSafe AI.

## Machine-readable data

Served by the repository's GitHub Pages site, rebuilt from `main` whenever the data changes (Pages sends `Access-Control-Allow-Origin: *`). The catalogue is too large to read whole, so start from the index and read one decision pattern at a time:

Index of the per-pattern files, with each pattern's row count and its file's address and size: https://kydlikebtc.github.io/awesome-jev/api/v1/index.json

One decision pattern's rows: https://kydlikebtc.github.io/awesome-jev/api/v1/patterns/<pattern>.json (for example safety-gating.json; the keys are listed under "Two axes"). Rows are shaped and ordered as the MCP server's `search_examples` returns them, and every row filed under the pattern is there, caveat flags included: a row flagged `not-jev` or `shadow-mode-only` is not an example of deciding with Jev, and the server leaves those out by default. A row can sit under several patterns.

Retired entries (dead links): https://kydlikebtc.github.io/awesome-jev/retired.json

JSON Schema for one entry: https://kydlikebtc.github.io/awesome-jev/schema/entry.schema.json

Which primitive fits a decision, as TypeSafe's own documentation guides it, arranged as a decision list whose every step cites the docs.typesafe.ai page it follows (design guidance, not a recommendation of any row; the README draws it under "What Jev returns"): https://kydlikebtc.github.io/awesome-jev/picker.json

This file: https://kydlikebtc.github.io/awesome-jev/llms.txt

If the site is unreachable, every file above except those under `api/` is also at the same path under https://raw.githubusercontent.com/kydlikebtc/awesome-jev/main/ (as committed on `main`), for example https://raw.githubusercontent.com/kydlikebtc/awesome-jev/main/catalog.json.

Each entry carries the fields defined in the JSON Schema: <!--n:fields-->slug, title, summary, summary_source, summary_zh, zh_machine, url, kind, patterns, patterns_reviewed, patterns_model_read_on, question_types, languages, platforms, author, repo, stars, repo_license, repo_created_at, repo_pushed_at, repo_commits, package, evidence, primitives_seen, observed_thresholds, evidence_none, wire, official, has_code, published, measurement, first_seen, checked, link_status, flags, notes, notes_zh, sources, license<!--/n-->. `summary` is English and `summary_zh` Chinese; `summary_source` says whose words `summary` is (`upstream-description`: the project's own GitHub description; `upstream-description-stale`: taken from that description, which no longer reads the same; `curated`: written for this catalogue by a person; `model-written`: written for it by a model after reading the project, since 2026-10-24; absent: not recorded); `evidence` names the file a claim was read in, and `evidence_none` says why no file can be cited.

## Two axes

`kind` is the form of the resource: <!--n:kinds-->official-docs, integration, project, plugin, sdk, snippet, tutorial, article, video, case-study, benchmark, discussion, alternative<!--/n-->.

`patterns` is what decision the example makes, and is the primary index: <!--n:pattern_keys-->tool-selection, intent-routing, context-compaction, safety-gating, output-validation, retry-control, human-escalation, model-routing, fan-out, search-ranking, data-extraction, classification, feature-extraction, document-triage, support-triage, content-scoring, recommendation, overview<!--/n-->.

`overview` marks a row that surveys the model or the space. It is also where the keyword rules that suggest patterns put a description they could not place, so <!--n:overview_unindexed-->251<!--/n--> overview rows that are projects or plugins with code count as not yet indexed by pattern: nobody has recorded reading them against the patterns. `patterns_reviewed` is the date a person did, and `patterns_model_read_on` (since 2026-10-24) the date a model did, shown as a model's; their absence means no reading was recorded, not that none happened. A row whose `patterns` equal the rules' suggestion for its summary (<!--n:patterns_rule_identical-->689<!--/n--> rows) shows agreement with the rules and nothing more.

## Human pages

English index: https://github.com/kydlikebtc/awesome-jev/blob/main/README.md

Chinese index: https://github.com/kydlikebtc/awesome-jev/blob/main/README.zh-CN.md

Decision patterns, with when NOT to use each: https://github.com/kydlikebtc/awesome-jev/blob/main/docs/patterns.md

Every example of one decision, on its own page: https://github.com/kydlikebtc/awesome-jev/blob/main/docs/by-pattern/<pattern>.md (for example safety-gating.md; append .zh-CN before .md for Chinese). The README shows the first ten of each.

Every independent measurement report and negative result on one page, with every note and caveat tag, the negative results first: https://github.com/kydlikebtc/awesome-jev/blob/main/docs/measured.md (docs/measured.zh-CN.md for Chinese). The README's "Measured, not claimed" shows the negative results first, then the picks of the curated independent-reports path in collections.json, then the first few others.

Benchmark measurements side by side: https://github.com/kydlikebtc/awesome-jev/blob/main/docs/benchmarks.md (docs/benchmarks.zh-CN.md for Chinese). <!--n:measured_rows-->27<!--/n--> `kind: benchmark` rows carry a `measurement`: the task, named datasets, comparators, kinds of metric, n, the model string, when, and the author's own `direction` (favourable, mixed, unfavourable or inconclusive), indexed from the author's report. Every direction is author-stated, not reproduced here, and a measurement without `read_on` was filled in by a script or a model, not read by a person.

Negative results: <!--n:negative_results-->5<!--/n--> rows record that their own author measured Jev for the use and concluded against it: a `kind: benchmark` row through `measurement.direction` (unfavourable), any other row through the `negative-result` flag ("measured, not adopted"). Author-stated, not reproduced here; read them before the positive examples. docs/status.md lists them, and the README's "Measured, not claimed" and docs/measured.md put them first.

Cross-platform compatibility table: https://github.com/kydlikebtc/awesome-jev/blob/main/docs/compatibility.md

The catalogue's shape as a dataset (languages, how rows reach Jev, star bands by kind, languages by pattern, patterns filed together, how many authors, no author named): https://github.com/kydlikebtc/awesome-jev/blob/main/docs/shape.md (docs/shape.zh-CN.md for Chinese). It describes what the catalogue holds, not the ecosystem at large, and its `typesafe-api` count is largely a default: discovery does not tag `platforms`.

What to check before trusting an entry: https://github.com/kydlikebtc/awesome-jev/blob/main/docs/vetting.md

How the catalog was built: https://github.com/kydlikebtc/awesome-jev/blob/main/docs/method.md

Sources and licences: https://github.com/kydlikebtc/awesome-jev/blob/main/docs/sources.md

Runnable examples: https://github.com/kydlikebtc/awesome-jev/blob/main/examples/README.md

## For agents

MCP server: in Claude Code, `/plugin marketplace add kydlikebtc/awesome-jev` then `/plugin install awesome-jev@awesome-jev` (installs this and the agent skill together); elsewhere, `pip install awesome-jev-mcp` (if pip finds no such package, `pip install git+https://github.com/kydlikebtc/awesome-jev`) and run `awesome-jev-mcp`. Source at https://github.com/kydlikebtc/awesome-jev/tree/main/src/awesome_jev_mcp — query the catalogue by decision pattern, language and primitive rather than reading it. Tools: search_examples, get_example, list_patterns, compatibility, check_model_string. search_examples also filters benchmark rows by what their author measured (comparator, dataset, direction), and each such row carries its `measurement`, the direction marked author-stated and not reproduced here; search_examples(outcome="negative") returns the negative results, shadow-mode rows included, each marked `negative_result`, and list_patterns counts them per pattern, with an `evidence` object per pattern (TypeSafe AI's documentation pages, cited files by evidence.kind, independent reports, negative results, rows citing no file: reports counted, not a verdict) that docs/shape.md and each docs/by-pattern/ page print too. compatibility() also counts, per surface, the catalogued rows whose `platforms` record a value that surface lists in compat.json's `catalog_platforms`, and search_examples(platform=…) takes that surface's id or a value rows record, matched whole; a surface marked coarse shares its value with another route, so a row matched through it may use that route. Resources: awesome-jev://flags (what each caveat flag means), awesome-jev://collections and awesome-jev://collections/{id} (curated entry points, each pick with a reason and a caution), awesome-jev://patterns/{key} (every row under one decision pattern). Prompt: wire_pattern(pattern, language, surface) — the pattern and a link to its section of docs/patterns.md, the surface's model strings and request shape, up to five cited rows, and the repository's own skeleton for that pattern and language when it ships one, marked as never executed. Every result carries the row's caveat flags, and an answer holding flagged rows ends with a caveat_glossary saying what those flags mean; rows that do not call Jev at all are excluded from search by default.

Agent skill: https://github.com/kydlikebtc/awesome-jev/blob/main/skills/awesome-jev/SKILL.md — the facts generated Jev code most often gets wrong, how to reach the catalogue's examples, caveats, platform differences and independent results, and the few design rules the catalogue adds. It complements TypeSafe's own agent skill, https://github.com/typesafe-ai/skills (catalogued as typesafe-skills-repo), which, with the live docs it reads (https://docs.typesafe.ai/llms.txt), is the reference for the API contract and for designing questions.

Decision taxonomy: https://kydlikebtc.github.io/awesome-jev/patterns.json

Kind and flag labels (EN/中文): https://kydlikebtc.github.io/awesome-jev/taxonomy.json

Platform matrix: https://kydlikebtc.github.io/awesome-jev/compat.json

(Each also under https://raw.githubusercontent.com/kydlikebtc/awesome-jev/main/, as above.)

## What is verified

Link records: `link_status` and `checked` together record a dated HTTP response. A success establishes availability at that time, not code execution or access to protected page content. Repository stars and licences are metadata snapshots. So are `repo_created_at` and `repo_pushed_at` (when GitHub says the repository was created and last pushed to, UTC) and `repo_commits` (commits on its default branch), re-read weekly; wherever `repo_commits` is recorded, the `single-commit` flag is set exactly when it is 1. They are GitHub's figures, not a verdict on whether a project is maintained.

Thresholds: <!--n:threshold_rows-->35<!--/n--> rows so far record in `observed_thresholds` constants their cited file compares a Jev answer with (<!--n:thresholds_recorded-->56<!--/n--> in all; a first reading, not every row whose file has one): the primitive, which quantity of its answer, the constant, what the code does on which side of it, and the text in the file it stands on, which is one of `evidence.matched` and re-read weekly. They are what each project chose for what being wrong costs it, not recommendations: a `noul` probability is not a `choice` confidence, and a threshold tuned on one model version does not carry to the next. An item without `read_on` was read by a model, not a person, and since 2026-10-24 a model dates its reading in `model_read_on`.

A row without a dated link-success record may have a host that refuses scripted requests. A 403 from Forbes or Medium is not evidence that a page is gone; such rows are left unstamped instead of inventing a successful check.

Runtime and performance: treat all catalogue code as untested by this repository, including entries without `code-untested`. Linked benchmarks are their authors' measurements and have not been independently reproduced here. Repository build checks and package smoke tests do not test catalogued integrations or the live Jev API. Inclusion is not a security review, maintenance guarantee or recommendation.

## Caveats for anyone summarising this data

1. **Identify who measured each result.** Rows flagged `vendor-reported` repeat the vendor's own benchmarks. Entries with `kind: benchmark` can contain independent reports, but the label alone is not a methodological endorsement or a reproduction by this catalogue. Attribute results to their authors, preserve limitations and distinguish vendor measurements from independent ones.

2. **`kind: alternative` rows are NOT Jev.** They are independent reimplementations with a compatible wire format. Several are among the most-starred repositories mentioning Jev, and are frequently miscatalogued elsewhere as Jev usage examples. A compatible API does not imply compatible calibration, so thresholds do not transfer. Where someone has read an alternative's files, its row carries `wire` (its route, how it spells the yes/no type, whether its answers come from open weights, a closed model or a proxy in front of another provider's model, whether it calls Jev to compare), each value backed by a string in a cited file; section 7 of https://github.com/kydlikebtc/awesome-jev/blob/main/docs/compatibility.md tabulates them.

3. **Jev cannot be run locally.** No weights are published. Any content titled "run Jev locally" is describing a substitute; one such article is catalogued with the `not-jev` flag specifically as a correction.

4. **jevai.org is not official.** It is an unaffiliated community site running a separate API with a different endpoint and request shape. Its `/jev-api` page documents a shape matching no primary source.

5. **The yes-no primitive is named `noul`, not "binary" or "boolean".** Vercel's AI SDK evaluation API spells it `boolean` and returns `.probability`; every other surface uses `noul` and `.noul`. `noul` answers carry no confidence field at all, unlike `choice` and `score`.

6. **There is no portable model string.** `jev-latest` on the native API, `typesafe-ai/jev` on Vercel, `typesafe/jev` on Cloudflare, `typesafe/jev-1.13` on OpenRouter. `typesafe/jev-1` does not exist anywhere and is a common fabrication.

7. **Star counts are a popularity signal, not a quality verdict.** Many entries were created within days of the model's launch; some carry `single-commit`, `no-license`, `archived` or `shadow-mode-only` flags precisely because popularity and substance diverge sharply here. The same holds for sibling-directory citations: after the sources that say where a row was found, `sources` lists each sibling directory whose README links the row's repository, as `{"catalog": "owner/name", "url": "https://github.com/owner/name"}`, re-read weekly (<!--n:cited_rows-->1180<!--/n--> rows carry at least one). These lists copy from each other, so the number says how widely a project is mentioned, not that anyone checked it.

8. **English is the model's primary training language.** The official docs state CJK scripts are handled but not equally well. Test on your own content before relying on it for non-English workloads.