TypeSafe AI launches Jev, a no-text decision model

Jev's name comes from William Stanley Jevons, whose 1865 book The Coal Question described what became known as the Jevons paradox: as steam engines got more efficient, coal consumption rose rather than fell. TypeSafe applies that idea to AI, expecting that cheaper decisions will widen the range of uses. The underlying architecture and weights are not public. CEO Diogo Almeida said on X that training data was 100% synthetic and not simply LLM-generated output. Within days of release, several Jev-like implementations appeared, but what got reproduced was the typed one-shot probability interface — the internal computation was guessed at by each implementer.

TypeSafe describes its training method as RLCD (Reinforcement Learning for Calibrated Decisions), contrasted with RLHF and RLVR in its own comparison chart. The company says Jev's outputs are always calibrated — a judgment given 90% confidence should be right about 90% of the time — though no figure showing the degree of calibration appears in the official blog. Tests of the same request sent five times on October 3, 2026 produced different results each time: the technical probability ranged from 0.78 to 0.86 and confidence from 0.68 to 0.79, and usage also differed from the documented example. Sending the same state translated into Japanese three times produced probabilities of 0.83 to 0.86, the same range as English.

The company also published its own caveats. The 70–500ms latency and the 40–200x speed advantage over frontier models on System One tasks were measured mostly from a laptop on the US West Coast, where the service currently runs. The "193.6x faster, 444.6x cheaper" figure on TypeSafe's homepage comes from its own workflow evals, and the company itself expects real-world gains to be on the high side. In its comparison demo against GPT-5.6 Terra, the states were short and dense paragraphs favoring Jev, and the single disagreement in the recorded round was on an item TypeSafe itself considers ambiguous. The 0% structured-output error rate shown in a chart was not measured but derived from the type guarantee, and even TypeSafe acknowledges LLMs can match output shape through constrained decoding like OpenAI's Structured Outputs.