--- title: Jev, TypeSafe's decision model: fast adoption, 20... description: Jev, a new 'decision model' from TypeSafe AI released in limited form on September 15, 2026, scores typed questions (multiple choice, numeric, yes/no) instead... canonical: https://daily.dev/posts/jev-typesafe-s-decision-model-fast-adoption-20-rivals-and-an-unpatched-injection-risk-5shek6mwx twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Jev, TypeSafe's decision model: fast adoption, 20 rivals, and an unpatched injection risk | daily.dev og:description: Jev, a new 'decision model' from TypeSafe AI released in limited form on September 15, 2026, scores typed questions (multiple choice, numeric, yes/no) instead... og:url: https://daily.dev/posts/jev-typesafe-s-decision-model-fast-adoption-20-rivals-and-an-unpatched-injection-risk-5shek6mwx og:image: https://api.daily.dev/og/posts/5sheK6MwX.png og:image:alt: Jev, TypeSafe's decision model: fast adoption, 20 rivals, and an unpatched injection risk og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Jev, TypeSafe's decision model: fast adoption, 20 rivals, and an unpatched injection risk **[Collections](https://daily.dev/sources/collections)** · 5 min read · 2 upvotes · 0 comments ## Summary Jev, a new 'decision model' from TypeSafe AI released in limited form on September 15, 2026, scores typed questions (multiple choice, numeric, yes/no) instead of generating free text, and has become the fastest-adopted model on Vercel's AI Gateway, reaching 35% of paid teams within two weeks. Within four weeks, 20 rivals emerged from Cloudflare, OpenAI, Perplexity, Liquid AI, AWS Strands Labs and others, with mixed, vendor-reported benchmark results that aren't directly comparable. A worked example pairs Jev with TrueForge (TrueFoundry's agent harness) to cut decision costs roughly 150x versus using Claude Opus for every match. However, HiddenLayer researchers found Jev vulnerable to indirect prompt injection and a tokenbreak-style attack, tricking a browser agent into clicking a malicious button — a reminder that constrained output formats don't make a model trustworthy against untrusted input. ## Content TypeSafe AI's Jev launched in limited release on September 15 and set off a rush of imitators within weeks. It is a different kind of model from a chatbot, and people are still arguing about what to call it and how far to trust it. ## What Jev does Jev is built for bounded semantic judgments on messy text. It doesn't generate prose. It answers typed questions and returns probabilities. There are three question types: - Choice: pick from up to 255 developer-defined options. - Noul: a yes/no probability from 0 to 1. - Score: an ordered scale of 2 to 10 levels. TypeSafe pitched this as a "System One" model. @iScienceLuvr points out that the term never caught on. People call Jev a "decision model" instead. Jev is closed source, with no published weights, so it can't run on vLLM. The service is based on the U.S. West Coast, which adds latency and data concerns for European teams. Red Hat's open-weight DiffusionGemma is offered as a self-hostable vLLM option for structured decision workloads. ## Adoption Vercel's AI Gateway data shows Jev reaching 35% of paid teams within two weeks, nearly double the reach of Sonnet 5.5. Vercel calls it the fastest-adopted model ever. Jev is cheap and fast at classification among preset outputs, and it isn't meant for general reasoning. OpenAI and Amazon have already shipped competing clones. The same gateway data shows open-source models handling over 70% of tokens and nearly 25% of spend, while frontier lab spending has grown ten-fold. That looks more like two complementary markets than a fight. ## The alternatives About four weeks after launch, 20 alternatives had appeared, from Cloudflare (Clef), OpenAI (Decisions API), Perplexity (pplx-decider), Liquid AI (d1), AWS Strands Labs, and independent developers. They range from 287M to 35B-A3B parameters. The numbers are messy. Cloudflare's Clef beats Jev on BANKING77 F1, 94.20 to 79.74. Jev still wins harder reasoning benchmarks like GPQA Diamond and leads Respan's third-party table for overall accuracy and consistency. The cheapest hosted option is Perplexity's flash-v1.2 at $0.01 per 1M input tokens. Most benchmarks are vendor-reported and not comparable across suites like JevBench and Decision Index. Independent testing is thin and contradictory. @marktenenholtz reports a private eval set where pplx-decider showed a 40% increase in recall over Jev, and concludes that decision models have a long way to go. @HamelHusain and a collaborator have also tested several of these models, Jev, d1-ultra, clef and pplx-decider among them, on whether they can be trusted for evals. ## Security and trust HiddenLayer tested Jev, which the vendor claims never hallucinates or makes type errors, and found it vulnerable to indirect prompt injection. In a simulated online store, a browser agent built on Jev was tricked into clicking a malicious "Confirm Price" button instead of following its instructions. Chained attacks could give an attacker full control. A tokenbreak-style attack also got past Jev's prompt verification use case. This shouldn't surprise anyone. Decision models are still LLMs under the hood with a different decoding strategy, so they inherit the same manipulation risks. A separate critique makes the trust problem sharper. Jev says "97% confidence" but doesn't explain itself. With an LLM you can at least read the text. Attackers could manipulate inputs or poison training data to get decisions that are confidently wrong. The recommended response is to treat the model as an untrusted component, log inputs, audit for drift, and keep humans in the loop for high-risk calls: verify, then trust. ## Using it in practice Two practical examples show how people are keeping Jev's output in check. Episode 3 of the Decide with Jev video series builds a PHP class called PreflightPolicy that reads Jev's answers and returns one label: ready for editor, revise, or needs review. Nothing publishes automatically, and a person still reviews each draft. The video explains why a missing answer isn't the same as a "no", and why zero confidence and no confidence take different paths. It uses two thresholds, 0.85 and 1.6, as a starting point and not a universal rule. The code is on the episode-3 branch of the Decide with Jev GitHub repository. The other pairs Jev with TrueForge, TrueFoundry's open-source agent harness, in a weekly client follow-up agent that matches client requests to GitHub issues. Jev scores a fixed set of choices with a calibrated confidence score, cutting decision cost roughly 150x compared with using Claude Opus for every match. On a 40-request test it matched Opus's accuracy, and low-confidence matches go back to Opus. TrueForge loads tools on demand and writes large outputs to files, and it matched Claude Managed Agents' task completion on DevRev's Enterprise Bench with about a third of the tokens. The same write-up shows how to turn an LLM into a decision model: 1. Freeze the backbone and add LoRA adapters. 2. Replace the language modeling head with a decision head. 3. Apply softmax per question and train with cross-entropy loss. 4. Calibrate on held-out data. A local test on Qwen3.5 0.8B raised accuracy from 37% to 65% in 60 steps and 10 minutes on 4GB of VRAM. ## Where this leaves things I find the speed of the clone wave more telling than Jev's own numbers. The idea is cheap to reproduce, the benchmarks are self-reported, and the one outside security test found the usual holes. A model that returns a tidy probability is easy to trust, and that's the risk. The sensible pattern, as the PHP example shows, is to let it score and have ordinary code and a human make the call. ## Questions this post answers ### What is a decision model like Jev and how is it different from a regular LLM chatbot? A decision model scores typed questions, such as a choice among fixed options, a numeric score, or yes/no, and returns an answer with a calibrated confidence instead of generating free text. It is built by freezing an LLM backbone, adding LoRA adapters, replacing the language-modeling head with a decision head, applying softmax per question, training with cross-entropy loss, and calibrating on held-out data. _Teams weighing cheap classification against full LLM calls can track decision-model comparisons on daily.dev._ ### Is Jev vulnerable to prompt injection attacks? Yes. HiddenLayer researchers found Jev vulnerable to indirect prompt injection despite vendor claims that it never hallucinates or makes type errors. In a simulated online store test, a browser agent built on Jev was tricked into clicking a malicious 'Confirm Price' button instead of following its instructions, and a tokenbreak-style attack bypassed Jev's prompt-verification use case. _Anyone wiring an agent to act on untrusted web content should follow security findings like this on daily.dev._ ### How much cheaper is using a decision model like Jev instead of a large LLM for agent decision routing? In a 40-request test pairing Jev with TrueForge, an open-source agent harness from TrueFoundry, routing fixed candidate matches through Jev and sending only low-confidence cases to Claude Opus cut decision cost roughly 150 times compared with using Opus for every match, at the same accuracy. _Engineers optimizing agent cost-to-accuracy tradeoffs can follow real benchmark numbers like this on daily.dev._ --- Tags: [#ai-agents](https://daily.dev/tags/ai-agents), [#ai-security](https://daily.dev/tags/ai-security), [#prompt-injection](https://daily.dev/tags/prompt-injection), [#jev](https://daily.dev/tags/jev) [View this post on daily.dev](https://daily.dev/posts/jev-typesafe-s-decision-model-fast-adoption-20-rivals-and-an-unpatched-injection-risk-5shek6mwx) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Jev, TypeSafe's decision model: fast adoption, 20 rivals, and an unpatched injection risk","url":"https://daily.dev/posts/jev-typesafe-s-decision-model-fast-adoption-20-rivals-and-an-unpatched-injection-risk-5shek6mwx","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/jev-typesafe-s-decision-model-fast-adoption-20-rivals-and-an-unpatched-injection-risk-5shek6mwx"},"datePublished":"2026-10-17T23:05:56.420Z","dateModified":"2026-10-23T13:34:21.678Z","description":"Jev, a new 'decision model' from TypeSafe AI released in limited form on September 15, 2026, scores typed questions (multiple choice, numeric, yes/no) instead...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/4c2b9dc35f10d0cbb02ea2da8e828d61?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/4c2b9dc35f10d0cbb02ea2da8e828d61?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Collections","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Collections","logo":"https://media.daily.dev/image/upload/s--fk_6ycEi--/f_auto,q_auto/v1780996001/logos/collections?_a=BAMAMiWQ0","url":"https://daily.dev/sources/collections"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/jev-typesafe-s-decision-model-fast-adoption-20-rivals-and-an-unpatched-injection-risk-5shek6mwx","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":2},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"ai-agents,ai-security,prompt-injection,jev","timeRequired":"PT5M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Collections","item":"https://daily.dev/sources/collections"},{"@type":"ListItem","position":3,"name":"Jev, TypeSafe's decision model: fast adoption, 20 rivals, and an unpatched injection risk"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/jev-typesafe-s-decision-model-fast-adoption-20-rivals-and-an-unpatched-injection-risk-5shek6mwx#faq","mainEntity":[{"@type":"Question","name":"What is a decision model like Jev and how is it different from a regular LLM chatbot?","acceptedAnswer":{"@type":"Answer","text":"A decision model scores typed questions, such as a choice among fixed options, a numeric score, or yes/no, and returns an answer with a calibrated confidence instead of generating free text. It is built by freezing an LLM backbone, adding LoRA adapters, replacing the language-modeling head with a decision head, applying softmax per question, training with cross-entropy loss, and calibrating on held-out data. Teams weighing cheap classification against full LLM calls can track decision-model comparisons on daily.dev."}},{"@type":"Question","name":"Is Jev vulnerable to prompt injection attacks?","acceptedAnswer":{"@type":"Answer","text":"Yes. HiddenLayer researchers found Jev vulnerable to indirect prompt injection despite vendor claims that it never hallucinates or makes type errors. In a simulated online store test, a browser agent built on Jev was tricked into clicking a malicious 'Confirm Price' button instead of following its instructions, and a tokenbreak-style attack bypassed Jev's prompt-verification use case. Anyone wiring an agent to act on untrusted web content should follow security findings like this on daily.dev."}},{"@type":"Question","name":"How much cheaper is using a decision model like Jev instead of a large LLM for agent decision routing?","acceptedAnswer":{"@type":"Answer","text":"In a 40-request test pairing Jev with TrueForge, an open-source agent harness from TrueFoundry, routing fixed candidate matches through Jev and sending only low-confidence cases to Claude Opus cut decision cost roughly 150 times compared with using Opus for every match, at the same accuracy. Engineers optimizing agent cost-to-accuracy tradeoffs can follow real benchmark numbers like this on daily.dev."}}]} ```