How Jev sorts a build into one of 21 use cases

Every listing in this directory was sorted by Jev, not by the person who submitted it. This page publishes the full criteria — all 21 of them — because a classifier you cannot inspect is just a black box with a label on it.

Updated September 2026 · 6 min read

One call, two questions

A submission is not run through a pipeline. It is one request carrying two questions of different types, answered together:

Primitive	The question	What comes back
`noul`	Is this a Jev build at all?	A calibrated 0–1 probability. Below threshold, the submission is turned away.
`choice`	Which of the 21 use cases does it belong to?	The winning key, a probability per category, and a confidence.

1 call per submission. The two questions are evaluated in parallel.

This is the part an LLM-based reviewer makes painful. Asking for a label as text means writing a format instruction, parsing the reply, validating it, and handling the case where the model explains its reasoning instead of answering. Here the answer set is declared up front, so the reply is a key from that set or it is not a reply at all.

The 21 criteria, in full

These are the definitions handed to the model, verbatim. They are also what you should read before submitting, if you want to know where your build will land.

Use case	Criterion	Listed
Agents & Browsers	Autonomous agents, browser automation, multi-step tool use on live sites	6
Context & Memory	Context engineering, memory layers, retrieval that persists across sessions	0
Coding & Code Review	Coding agents, code review, refactoring, PR and repo automation	0
Routing & Model Choice	Model routing, cost or latency arbitrage, choosing which model serves a request	0
SDKs & Integrations	Client SDKs, framework adapters, protocol or provider integrations	0
Open Source	Open-source repos, self-hosted releases, community forks	0
Benchmarks & Evals	Benchmarks, eval harnesses, scoring and model comparison suites	0
Search	Search engines, retrieval, deep research over the open web	0
Documents & OCR	Document parsing, OCR, PDF, form and scanned-page understanding	0
Security & Abuse	Security analysis, abuse and fraud detection, prompt-injection defense	0
Inbox & Support	Inbox triage, ticket routing, customer support automation	0
Social Feeds	Social posting, feed curation, creator and community workflows	0
Ads & Marketing	Ad creative, campaign copy, marketing content generation	0
Sales & Leads	Lead qualification, outbound sequences, outreach automation	0
SEO & GEO	SEO and GEO — ranking in search results and being cited by AI assistants	0
Ecommerce	Storefronts, catalog and listing work, orders and returns	0
Games & Real Time	Games, real-time control loops, latency-budgeted decisions	1
Trading & Markets	Trading bots, backtesting, market monitoring	0
Robotics & Devices	Robot arms, drones, embodied control, on-device inference	1
UI	UI generation, design-to-code, frontend QA	0
Tools & Apps	Standalone developer tools, CLI companions, desktop and web apps	11

Aligned one-to-one with madewithjev.com's taxonomy so the two directories can be read side by side.

When a build could fit two categories

This happens constantly. A coding agent that also routes between models fits both Coding & Code Review and Routing & Model Choice. A browser agent that plays games fits Agents & Browsers and Games & Real Time.

One category per listing. A build appears in exactly one place, so the counts add up and a filtered view never shows duplicates.
The model picks by primary purpose, not by the most impressive incidental capability. A tool whose point is review gets filed under review, even if it also routes.
Probabilities expose the close calls. When the top two categories are near-tied, that is visible in the response — useful signal for tightening a criterion that is pulling too much weight.

Misclassification is possible

A single-category taxonomy has to be wrong sometimes. If a listing is filed somewhere that looks wrong to you, it is probably a criterion that needs sharpening rather than a build that needs re-reading — and since categories are assigned by the model rather than chosen by the submitter, nobody can self-promote into a category.

What the reviewer actually sees

Jev never opens your page. It judges the metadata the site publishes — the title, description and Open Graph tags that any crawler would read. That has a practical consequence worth knowing:

Your OG description is your submission. If it does not mention Jev or TypeSafe, the reviewer has little to go on, and the Jev-ness probability comes back low.
Describe the build, not the marketing. A page whose metadata reads like a landing page is harder to place than one that says plainly what it does.
A GitHub repo description is enough. Many listings here are repos whose one-line description was sufficient to be judged.

Why we publish the criteria

Because the alternative is a directory whose ordering nobody can audit. If you disagree with where a build landed, the criteria above are the thing to argue with — and they are stable enough to argue with.

The same reasoning is why the ranking rules are published in full on how it works: community upvotes decide directory order, the paid spotlight is labeled, and neither is allowed to quietly influence the other.

Common questions

How does Jev decide which use case a build belongs to?

One call asks two questions: whether the submission is a Jev build (a calibrated probability) and which of the 21 use cases it belongs to (a single choice with per-category probabilities).

Can a build be in two categories?

No. Each listing gets exactly one category, chosen by primary purpose rather than by incidental capability, so counts never double up.

Who assigns the category — me or Jev?

Jev does. Submitters provide only a link. That keeps categories consistent and means nobody can file themselves into a flattering category.

Why was my submission rejected?

Most rejections are the Jev-ness check: the page's metadata did not indicate a Jev build. Dead links and pure marketing pages are also turned away. Making the repository or page description say what the build is and that it uses Jev is usually the fix.

BuiltOnJev is an independent community project, not affiliated with TypeSafe AI. Specs and eval numbers come from TypeSafe's published materials; directory figures come from this site's own submissions and are updated as builds arrive.