What it is
“Judgment models” (also called “decision models”) are a category of AI model built to do one narrow thing: take in a question with a fixed set of allowed answers — yes/no, a choice among options, a numeric score — and return that answer with a calibrated probability attached. They do not write sentences, generate code, or hold conversation. Where a large language model produces open-ended text that downstream software must parse and hope is correctly formed, a judgment model returns a typed value a program can act on directly, with no parsing and, its makers argue, no risk of inventing an answer outside the allowed set.
The category’s reference product is Jev, built by TypeSafe AI, a San Francisco startup founded by Diogo Almeida, a former OpenAI researcher who worked on the instruction-following methods behind ChatGPT, together with co-founders Erik Gafni and Sasha Sheng. TypeSafe calls Jev the first of its “System One Models,” a label borrowed from Daniel Kahneman’s distinction between fast, intuitive “System 1” cognition and slow, deliberate “System 2” reasoning — the pitch being that most software decisions need the former, not a full generative model doing the latter. TypeSafe’s own framing is that Jev is “a frontier-intelligence function call: unstructured state in, typed probabilistic decisions out.”
The commercial case rests on speed and cost rather than capability. TypeSafe claims Jev runs 40 to 200 times faster than comparable frontier LLMs, with input priced at $0.042 per million tokens and output free of charge — cheap enough, the company says, “to meter.” The training method behind it, which TypeSafe calls Reinforcement Learning for Calibrated Decisions, optimizes specifically for well-calibrated probabilities rather than fluent prose.
Why it matters for AI governance and narratives
Judgment models sit inside a framing contest that matters more than the product category itself: the claim that narrow, typed, “hallucination-free” decision systems are more trustworthy and auditable than general-purpose generative models. That claim is a vendor’s self-description, not a settled finding, and it deserves the same scrutiny the observatory applies to any ecosystem’s account of its own reliability — a model that can only select from pre-defined options cannot hallucinate a new answer, but it can still be miscalibrated, poorly scoped, or wrong about which of the allowed options is correct, and the audit trail is only as good as the typed schema designers chose to expose.
The category also maps directly onto the observatory’s agentic thread. As autonomous agents proliferate, something has to decide routing, approvals, and handoffs between them cheaply and at scale — judgment models are being pitched explicitly for that plumbing: sorting requests, flagging which agent owns a task, screening outputs for jailbreak attempts, checking whether another AI’s draft is good enough to ship. That makes them infrastructure for machine-to-machine governance as much as for human-facing products, and it raises the stakes of who sets the typed schemas and allowed-answer sets that these models operate within — a form of policy encoded in interface design rather than in regulation.
Key facts and dates
TypeSafe published its introductory blog post dated September 28, 2026, announcing early access to Jev; some trade coverage places initial limited access slightly earlier, in mid-September, so the precise rollout date is not fully consistent across sources. Trade press also reported an accompanying $40 million funding round, a detail not confirmed on TypeSafe’s own post. TypeSafe named the model after the economist William Stanley Jevons, a reference to Jevons’ observation that falling costs tend to increase, not decrease, total demand.
The category spread fast. Within days, platforms including Vercel, Netlify, and LangChain had integrated Jev, according to InfoQ’s coverage. By early October 2026, Amazon Web Services had released its own open-source entrant, Strands Decider 2B, built by distinguished engineer Marc Brooker on top of the Qwen3.5-2B base model and published under an Apache 2.0 license with weights on Hugging Face — explicitly described in TechCrunch’s coverage as a response to demand that judgment/decision models created. TechCrunch also reported that OpenAI released a comparable offering the same week, though this explainer could not independently verify further detail on that product.
Where to learn more
- Introducing System One Models & Jev — TypeSafe AI — the primary source for Jev’s design rationale and official framing
- TypeSafe AI Releases Jev: a Decision-Only Model That Returns Typed Probabilities Instead of Text — InfoQ — independent technical reporting with pricing and adoption detail
- Amazon releases its own Jev clone as decision models flood the web — TechCrunch — on Amazon’s Strands Decider 2B and the category’s rapid spread
- Why a New Class of AI “Judgment Models” Could Have Big Business Implications — The AI Daily Brief (podcast, Sept 16, 2026) — early framing of the category’s business use cases