Showcase
Every link here was opened and dated before it went in, and the grouping is by the shape of the problem rather than the category of the product. Adding a project is cheap; checking one is the part that makes this worth reading.
Quick answer · verified 2026-09-19
Read across these projects and the pattern is not what the marketing suggests. Almost nobody is using Jev as a text classifier. They are using it as a gate inside an agent — is this shell command safe to run, should this permission be granted, is this draft allowed out, which tool runs next, does this event belong to that session.
That fits what the model is: fast enough to ask on every step, cheap enough that asking constantly is not a budget decision, and typed so the answer can drive a branch without parsing. A classifier is asked once per document. A gate is asked hundreds of times per task.
| Projects listed | 12 — each link opened and dated |
|---|---|
| Most common use | safety and permission gates — a cheap judgement in front of an action an agent is about to take |
| Least common use | plain classification — the thing it is usually described as being for |
| Largest project | Pinecone's Cultivar — grades agent skills run in sandboxes |
The most common thing people actually build: a cheap judgement in front of an action an agent is about to take. Almost nobody is using Jev as a plain classifier.
Static analysis of shell scripts, with Jev judging whether a command is dangerous.
Shell safety is the textbook case for a decision model: the question is bounded, the answer is a judgement rather than a fact, and you want to ask it about every line rather than once. Pairs naturally with confidence gating — flag at a threshold, block above it.
GitHub →checked 2026-09-20
Kubernetes incident triage that keeps deterministic rules in code and sends only the judgement calls to a model.
The architecture this whole site argues for, applied to on-call. Rules that can be written as rules stay as rules; the residue — is this actually an incident, how severe, who owns it — goes to a model. Agent-neutral, so the decision layer is not tied to one framework.
GitHub →checked 2026-09-20
Choosing the next step, the right tool, or where an incoming event belongs.
Describes itself as the first general-purpose System One agent for Jev.
An agent whose control flow is decisions rather than generated text — the strongest test of whether a model that cannot write can still drive a loop. Worth watching for where it has to reach for an LLM anyway.
GitHub →checked 2026-09-20
A multi-agent news pipeline that pairs a generative model with Jev, running live at news.aatf.ai.
Uses GLM-5.3-Flash for the writing and Jev for the decisions — include or drop, which section, how important. The pairing the rest of this site describes, shipped and publishing daily.
Things that improve how you use Jev itself, or fold it into an existing workflow.
A CLI that tunes your criteria against labelled data with GEPA, instead of against your intuition.
The missing half of every other page about Jev. Option descriptions are the entire prompt surface, and this is the only public tool that improves them empirically: it measures uncertainty across your dataset, picks the ambiguous rows for you to label, runs GEPA, and shows you the proposed wording diff with its score. Nothing is accepted automatically — a higher training score still waits for you.
GitHub →checked 2026-09-20
Runs your agent skills and docs in sandboxes to test them, with Jev doing the grading.
From Pinecone, and the most starred project in this list. Evaluation is a decision problem wearing a different hat — did this skill do the right thing — so a model that returns a graded judgement rather than an essay fits the job better than a chat model does.
GitHub →checked 2026-09-20
An intent-driven UI runtime: buttons declare what they mean, and the runtime routes it.
Inverts the usual wiring. Instead of a click handler naming a function, the control declares an intent and a policy layer decides which registered surface should handle it. Published on npm with a live demo.
A terminal kanban shared by people and agents, with WIP limits and a GitHub snapshot.
Interesting less for the board than for the premise: a work surface both a human and an agent read and write, where deciding what moves next is the hard part.
GitHub →checked 2026-09-20
A game is a long run of decisions over a fixed action space, which is exactly Jev's shape — so these are more load-bearing than they look.
Generating level terrain during play, one Choice at a time.
A write-up rather than a repository. Level generation is a sequence of bounded picks — which tile, which direction, which feature — and sub-second decisions are what makes doing it during play rather than at build time possible at all.
Write-up →checked 2026-09-20
A word game where everyday phrases are rewritten into absurd synonyms for you to decode.
Plays against the grain of the model, which is what makes it worth a look: Jev cannot write, so the puzzle has to be assembled from picks over candidate words rather than generated.
Live →checked 2026-09-20
A chatbot built on a model that cannot write, by choosing replies one word at a time.
An experiment in doing the thing Jev is explicitly not for. Worth including precisely because it maps the edge: every token is a Choice over candidates, which works, and demonstrates exactly why you would not ship it.
GitHub →checked 2026-09-20
StarCraft II played by picking from the legal action space each step.
Same shape as TypeSafe's own Doom demo and as the chess demo on this site: a game is a long run of decisions over a bounded action set, so it stress-tests decisions-per-second and cost in a way a benchmark does not.
GitHub →checked 2026-09-20
If you have built something with Jev, open an issue on jev-measured with a link. The bar is only that it exists and can be opened — the list is curated for accuracy, not for taste.