What 5,314 Jev Cases Show: A Summary of Jinqiu Research's Study
Published: 2026-09-25
Jinqiu Research sorted 5,314 public Jev case records into 1,912 applications. Gates, action control, and routing make up over half, and only about 11% have reached product or framework integration. The numbers, examples, and caveats, summarized.
Part 1
What was counted
Jinqiu Research collected 5,314 case records from nine public Jev directories as of 2026-09-23 (JEVLAB, Awesome TypeSafe Jev, AwesomeJev.com, Made with Jev, Jev Finder, Jevfast, Jev-xyz, Jev Atlas, and TypeSafeai.org Cases). Deduplicating by GitHub repository, X post, YouTube video, product page, project name, and author left 3,075 entities. Of those, 1,912 are applications, tools, product integrations, or public experiments (SDKs, tutorials, news, reviews, and directory mirrors were not recorded), and 1,531 can be traced to public code or the creator's own post. Not every codebase was run to verify it.
Part 2
By function: gates, action control, and routing make up over half
TypeSafe defines Jev as a System One model that only makes the judgment; workflow, permissions, computation, and actions stay in code. By the job Jev does inside a program, three uses lead: verification and gates (20.6%, 393 applications), action selection and real-time control (19.0%, 364), and routing and triage (12.1%). Together they are 51.7%. What they share: the judgment sits at a decision point in code, it runs often, it must be fast, and code can fall back when it is wrong. Examples from the study: Vercel founder Guillermo Rauch used Jev as a command safety reviewer for the fx coding agent, and LangChain ships Jev as a TypeSafeClassifier, model-routing middleware, and a dangerous-tool-call interceptor. The study sees gates as strongest on cases that are semantically suspicious but hard to enumerate as if-else rules.
{ "model": "jev-latest", "state": { "agent_goal": "Clean the build output before a release", "proposed_command": "rm -rf ./dist && git push --force origin main", "workspace": "production repository" }, "questions": { "gate": { "type": "choice", "instructions": "Should the agent run proposed_command without asking a human?", "criteria": { "allow": "Local and easy to undo", "review": "Changes shared state; a human should confirm", "block": "Destroys work or history" } }, "on_goal": { "type": "noul", "instructions": "Does proposed_command serve agent_goal?" } } }Part 3
By scenario: agents, games, browsers, and content first
Agent infrastructure (397, 20.8%), games and real-time interaction (278), browser and computer operation (209), and content, marketing, and creative tools (197, 10.3%) together make up 1,081 applications, or 56.5%. The authors note these counts reflect developer interest and what is easy to demo, not market size. Another 23.8% did not fit a category: mostly internal tools, small local apps, generic judgment CLIs, and mixed cases. Enterprise operations are only 3.6% and security, moderation, and verification 3.8%, which the study reads as an engineering barrier (permissions, data, approvals) rather than a lack of demand; these buyers are tied to human queues, risk, and real money, and are probably more willing to pay. Search, RAG, and knowledge are just 14 applications (0.7%), or 98 (5.1%) when counted by function, well below the weight TypeSafe gives to reranking and citation checks. That may not mean low demand: RAG is hard to turn into a viral clip of a few dozen seconds, or existing embeddings and rerankers may already be mature enough that Jev adds little there for now. The study suggests the opening is judging evidence after retrieval: does this passage support the claim, conflict with another, or carry a prompt injection. Simon Willison has tested Jev for reranking and warns about black-box bias and evaluation.
Part 4
How the action demos work, and where they stop
Browser, computer, and game demos follow one recipe: compress the environment into a set of legal candidate actions using the DOM, the accessibility tree, OCR, sensors, or game state; let Jev pick the next one; let code execute and check it. Browserbase put Jev into the Act step of Stagehand, choosing the action type and then a DOM element, and falling back to an LLM when confidence is low; its early self-tests cut Act's median latency from 1.97 s to 0.46 s, and Browserbase calls Jev a component of an agent, not a production agent. The study draws the line clearly: Jev is good at a single choice among given candidates and not at open-ended planning, long-term memory, or reflecting on failure. A fast game demo does not prove long-task success, and a trading demo that makes a decision every 300 ms does not prove it makes money. Steve Krouse's Typewriter refreshes 16 judgments on every keystroke, and the study adds that faster feedback is not more accurate feedback: typed output guarantees the format, not that the judgment is right.
Part 5
Timeline and maturity
By the date each directory indexed a case (not an exact launch date), there were 9 on September 15, 299 on the 16th and 17th, a peak of 883 from the 18th to the 20th, and 292 after the 21st. The first week was dominated by action selection, games, and browser demos; the second shifted to agent infrastructure, gates, and routing. The study does not see these as contradictory: the Jev on social media (fast enough to play games and click buttons) did the market education, and the Jev in GitHub directories (agent routing, gates, screening, verification) is the engineering adoption. Of the 1,912 applications, 809 are runnable open-source projects (42.3%), 729 are demos or prototypes (38.1%), 209 are product or framework integrations (10.9%), and 165 are only directory entries (8.6%). In the study's words, the wave is driven by demos that run and example repositories, and it is still far from production at scale.
Part 6
Where Jinqiu expects the value to go (their view)
The study argues that when a semantic judgment costs a few percent of what it did, work that could only be sampled can run on every record; public cases already batch 500 or about 1,700 emails at a time. The judgments did not get smarter, only cheap enough to run everywhere, and that only holds while low latency and low cost hold together. It expects the gains to go to companies that already own high-frequency workflows, such as SaaS vendors, agent harnesses, browser automation, and vertical software, rather than to new companies built around Jev. Thin model routers, generic classifiers, Jev-compatible endpoints, directories, and one-off demos will commoditize, because they own neither outcome data nor the customer's workflow. Simple call-an-LLM-and-parse-JSON steps will be squeezed too: once a task can be written as Choice, Score, or yes/no, the cost and latency case for an LLM in that step gets recalculated. Compatible interfaces, open plugins, and alternative implementations already exist, so calling Jev is not a moat in itself; the model can be replaced. The moats it names are domain labels and error cases, deep integration with permissions, logs, and review queues, and a design for accountability and compliance. JEV Hub note: the gate request above is our example, not code from the study; try it in the Jev Playground or wire it into an agent through the Jev MCP server.
Related projects
- @typesafe-ai/sdk — Official TypeScript/JavaScript client for POST https://api.typesafe.ai/v1/systemone.
- jev-sentinel — Pi, Claude Code, and Codex CLI guard that asks Jev whether a tool call is on-task, risky, or injected before it runs.