Jev use cases
People hand Jev the decisions a system makes over and over: sort this pile, route this ticket, score this draft, pick the next click, allow or block. This page groups the 86 items we have catalogued by which of those decisions they show.
In short
- 86 items catalogued, 60 of them tagged with at least one job; the rest are explainers and announcements, which is why there is no "getting started" facet.
- 18 of the 18 projects have a page of their own on this site — what it does, how it works, what we read in it, and what we did not check.
- Nine jobs, and the two that repeat most are sorting a corpus nobody will read and picking the next action in an agent loop.
- Our own measurements, not a vendor's: median wall time 783 ms per call (647–1030), 420 input tokens, $0.0000176 a call.
- 13 adversarial calls × 4 runs: 0 answers outside the option set. The full numbers are here.
The nine jobs, at a glance
Each row links to the section below it, and names one item from our catalogue that shows the job with its own numbers attached.
| The job | One item that shows it |
|---|---|
| Sort a pile that is too big to read | Jev - The Ultimate Classification Model? — 323,066 views · 16 min |
| Pick the next step in an agent loop | browser-use/jev-ultrafast — 17,034 ★ · last push 2026-09-18 |
| Shrink what the model has to read | tamaratran/fast-jev-compaction — 6,135 ★ · last push 2026-09-18 |
| Judge every row where the data already is | Show HN: jevals – replacing LLM judges with typed Jev decisions — 42 points · 3 comments · 2026-09-20 |
| Check the work before it ships | reticlehq/reticle — 811 ★ · last push 2026-09-22 |
| Block the step you should not take | Sac-Y/Jev-cu — 564 ★ · last push 2026-09-22 |
| Decide inside a frame | wfzyx/von — 399 ★ · last push 2026-09-21 |
| Decide with money on a clock | jarrodwatts/jev-trader — 1,929 ★ · last push 2026-09-17 |
| Choose what to show someone | superagents-lab/jev-search — 394 ★ · last push 2026-09-20 |
Sort a pile that is too big to read
Classification, at a volume where reading each item was never on the table.
This is the job the model was built for and the one most of this topic's rows land in: a corpus, a queue or an inbox where a wrong answer is recoverable and a slow answer is not. The pattern that keeps working is small questions with named answers, decided in parallel, with the rule for combining them kept in code you can read.
- YouTube
YouTube · video
Jev - The Ultimate Classification Model?
Reads Jev as a classifier, which is the closest everyday description of a typed-decision model. Worth watching for that framing alone.
- GitHub
GitHub · project
TianyuCodings/NanoJev
A small replica of the mechanics: parallel decisions, dynamic routing. Useful as the readable version of the architecture if you are not going to work through the paper.
- X
X · video
一个只会做选择题的小模型，能长出多少玩法？ TypeSafe 的 Jev 就是这么个东西：给它最多 255 个选项，几百毫秒挑一个，再附一个校准过的置信度，输出 token 免费 结果开…
A Chinese-language walkthrough of what a multiple-choice-only model grows into: up to 255 options, a few hundred milliseconds, and a calibrated confidence attached.
Pick the next step in an agent loop
The decision that used to be a chat completion, once per click.
An agent loop asks the same question all day: of everything on this screen, what is the next thing to do? Answering it with a chat model means paying for prose nobody reads. Answering it with a decision model means the loop gets an operation and a target, and the program decides what to do when it is unsure.
- GitHub
GitHub · project
browser-use/jev-ultrafast
The browser agent we read line by line for the resources column: Jev picks the next action from an indexed space instead of emitting free-form clicks. The most-starred build in this list, and the one with the clearest engineering write-up.
- GitHub
GitHub · project
awlevin/typesafe-computer-use
Computer use for about $0.0002 a step, on macOS: OCR the screen, classify the next action with TypeSafe, click. Python 3.12 with an MIT licence and a CI badge, and the price per step is in the repository description rather than buried in a table. We have not run it.
- GitHub
GitHub · project
Sac-Y/Jev-cu
The one repository in this queue with no published description, and the one that states its design constraint most plainly: text goes in and screenshots do not, a local policy gate blocks the sensitive actions, and Codex does the reading and clicking. The README is in Chinese.
Shrink what the model has to read
Deletion instead of paraphrase, so a file path cannot be summarised away.
Context windows get filled by tool calls and their output, most of which stops mattering within minutes. Scoring each item and dropping the stale ones keeps the exact text of everything that stays — which is a different promise from a summary, and a better one when the cost of losing one exact string is a wrong answer twenty turns later.
- GitHub
GitHub · project
tamaratran/fast-jev-compaction
A Claude Code plugin that replaces the compaction step with a Jev call. Small, unglamorous, and among the highest-starred entries here — the pattern tends to show up in tooling before it shows up in products.
- GitHub
GitHub · project
TianyuCodings/NanoJev
A small replica of the mechanics: parallel decisions, dynamic routing. Useful as the readable version of the architecture if you are not going to work through the paper.
Judge every row where the data already is
Evaluation, moderation, labelling: the same decision, run on the whole table.
Judging with a chat model costs a generation per row, which is why teams judge a sample and hope. When the judgement is a handful of bounded choices, the whole table becomes affordable — and the judgements become reviewable, because each one is a label with a confidence rather than a paragraph.
-
Hacker News · thread
Show HN: jevals – replacing LLM judges with typed Jev decisions
Replaces the LLM judge with a typed decision, which is the use case we expected to appear first and the one with the clearest failure mode. The thread's own objection is the one to keep: what happens when the judgement depends on context the typed decision cannot carry.
- X
X · image
Build a Jev Judge
Takes the same argument to evaluation: if the judgement is a handful of bounded decisions, does it need another round of text generation to make it? The most concrete Jev-as-judge case we have found on X, and the one we would test before relying on it.
- GitHub
GitHub · project
devagrawal09/jev-review
Read the README for the pipeline rather than the dashboard: a Null risk matrix, then Choice and Score file profiles, evidence selection, mechanism classification, severity, and a conditional Choice that routes which reviewer sees it. Orchestration stays in code and Jev only answers the bounded judgements — which is the pattern to copy.
Check the work before it ships
A decision model above the agent, asking whether the agent is actually finished.
The most consistent pattern in the field notes: the team that builds the thing is not the team that checks it. A decision model is a cheap second opinion — complete or not, requirements met or not, test sufficient or not — and it can run on every change rather than at the end of a sprint.
- GitHub
GitHub · project
reticlehq/reticle
File this one under roadmap, which is where Reticle puts Jev: the README says plainly that nothing ships against it yet. The case is still worth reading, because it is the only project here that names the decisions it would route — is this page settled, is this finding worth chasing — and then says why a 70–500ms typed answer beats an LLM there.
-
GitHub · project
thruwire/foreman
Puts the decision model above the coding agent instead of inside it: a Codex or OpenCode worker does the engineering, and Foreman independently asks whether the ticket is finished, the tests exist and the requirements were met. The separation of the worker from the thing that checks the worker is the pattern most of our field notes end up describing.
-
Hacker News · thread
Show HN: Jeff – A read-only CLI for semantic code review using Jev
A read-only review CLI, and the thread argues about the word read-only rather than the model: one commenter points out that in a coding agent an edit is refused unless the file was already read in that conversation. Good thread for the enforcement question.
Block the step you should not take
A gate in front of the action, not a paragraph warning about it.
A typed decision is easy to enforce: if the answer is high risk, the program stops. That is a different kind of safety from asking a model to be careful, because the model is never the thing that decides whether to proceed — it only grades what it was shown.
- GitHub
GitHub · project
Sac-Y/Jev-cu
The one repository in this queue with no published description, and the one that states its design constraint most plainly: text goes in and screenshots do not, a local policy gate blocks the sensitive actions, and Codex does the reading and clicking. The README is in Chinese.
-
X · post
This made me rethink where AI actually fits into security engineering. For purely engineering…
A security engineer's account of where a decision model fits in that work — a question the vendor's launch post never answers.
Decide inside a frame
Real-time loops, where the model has tens of milliseconds and one action.
Games and robotics are the honest stress test of the whole idea: the decision is small, the clock is merciless, and the result is either a working run or a recording of one. The open models in this column give up accuracy against the hosted one and win on latency and cost, which is the trade to make explicitly rather than by accident.
- GitHub
GitHub · project
wfzyx/von
The open substitute, published with its own comparison against the model it substitutes: 72.0% macro on a 49-task suite against TypeSafe's 96.6%, and 9.00 kills against 5.62 in ViZDoom — a clear loss on accuracy and a win on the real-time loop. Its GitHub description says sub-15ms and its README headline says sub-25ms; both numbers are the publisher's.
- GitHub
GitHub · project
rmalde/minecraft-agent
A planner picks the goal and Jev picks the action, and then the repository publishes the run: nether-final-08 finished in 8 minutes 43.300 seconds against 14 minutes for the previous attempt, with the routes and a read-only sensor kept separate from the game. End-to-end times for a Jev loop are rare enough to be worth the row.
Decide with money on a clock
Buy or sell, every block, with the consequence attached.
Trading is where a decision model earns its keep if it earns it anywhere: the cadence is fixed, the answer is binary, and nobody wants a paragraph in the loop. The repositories here are careful about the distinction that matters — the model chooses, the program prices and sends, and the default is a dry run.
- GitHub
GitHub · project
jarrodwatts/jev-trader
One trade decision per Monad block, made by Jev — the most literal reading of the model in this list, a decision every few hundred milliseconds with real money behind it.
- GitHub
GitHub · project
OpenByteInc/QuantDinger
An open-source trading OS. It is here because its own description puts the decision loop at the centre rather than a chat window; the strategy numbers are the author's, and we have not audited them.
Choose what to show someone
Ranking and selection: which sources, which rows, which result.
Search and recommendation are the same decision as routing — pick from a candidate set, show your reasoning as a score — and they are the version a reader can check by using it, because a search result is right or wrong in front of them.
- GitHub
GitHub · project
superagents-lab/jev-search
Search where Jev chooses the sources, the time range and the terms, and then ranks what comes back — links and snippets with visible relevance scores, no generated answers. There is a live instance, so this is the one row in the column you can check in a browser without cloning anything.
- GitHub
GitHub · project
featherless-ai/simple-jev
A different route to the same output contract: read the next-token logits for each question and assemble choices, rubric scores or truth values, with no classifier head and no training. Served as a playground, which makes it the quickest way to see the shape of a Jev answer without a key.
By the work you do, rather than the decision
The same rows, grouped by the kind of work instead of the kind of decision — the way somebody looking for "trading" rather than "classification" would search for them.
-
19 items
Triage and routing
Picking a route, a label or a next model — classification where a wrong answer is recoverable.
-
18 items
Coding and context
Inside a coding agent: compaction, context windows, the harness around the model.
-
20 items
Research and data
Benchmarks people ran, architectures people reverse-engineered, results that did not work.
How to tell whether your job is one of these
Three questions, in order. Is the answer one of a fixed set of options you could write down in advance? Can you recover if it is wrong — retry, escalate, or let a human see it? And is the question asked over and over, so that a slow or expensive answer costs you every time? Three yeses and the decision belongs in this list. A no on the first one means you want an LLM, and a no on the second means you want a person.