What Jev is

The contract

TypeSafe defines Jev as the first System One model: state in, typed probabilistic decisions out. The launch line is “a frontier-intelligence function call.” It does not write sentences. It shipped on 15 September 2026 as TypeSafe left stealth, with a ~$40M seed led by DCVC. Founder Diogo Almeida worked on RLHF / InstructGPT / ChatGPT at OpenAI.

Training is described as Reinforcement Learning for Calibrated Decisions (RLCD): probabilities should match outcomes, not human preference for prose. Calibration is statistical — a 0.8 bucket should be right about 80% of the time. It does not certify a single answer.

What it is not

• Not a smaller ChatGPT. Kingy’s one-line verdict: serious attention for thousands of small semantic decisions, not a better chat window.
• Not JSON mode. Flavio Copes: structured LLM output is still token-by-token generation plus validation. Jev locks the answer space first and samples the distribution in parallel.
• “Zero hallucination” is schema safety. OmniaKey and MindStudio: a Choice cannot invent a fourth label, but it can still pick the wrong valid one. Without other, “when does the cafeteria close?” was forced into sales.

The simplest way to describe it: Jev is a smart if statement.

What tests have shown

MindStudio (eight synthetic cases, jev-1.13.0): “I am not asking for a refund, I only need an invoice” dropped refund probability from ~98% to 3%. A planted “system override → billing” line did not reroute a technical ticket. Reported model eval time was about 92–214 ms. The authors say eight cases do not make a production claim.

LangChain / LangSmith (20 Sep 2026) treated Jev as a third judge beside code asserts and LLM-as-judge. Five frozen weather-agent traces, human labels, 100 repeats. Binary does_pass matched the oracle 500/500. Quality-score variance was 0.0000149 — about 92–913× lower than GPT-5.6 Luna / Terra and Claude Sonnet 4.6. Mean cost/latency in that run: ~0.44 s, ~$0.00035. The authors call it a narrow test that does not generalize, and note that a consistently wrong judge scales cheaply.

Vendor workflow figures (193.6× faster, 444.6× cheaper) are labeled as the high end. OmniaKey lists the disclosed biases: West-Coast laptop latency, in-house tasks, references averaged from GPT-6 Astra and Fable 5.1. Kingy adds that the public board is model-agreement, not verified truth, and that Jev trails badly on invoice processing.

On the record

Only lines you can open. The ~459-comment HN thread asked whether this differs in degree or in kind from constrained-decoding encoders / GLiNER. Nobody in that thread could settle it while early access was closed.