ChatGPT pioneer launches Jev for programmatic decision logic

TypeSafe has emerged from two years in stealth with Jev, a model designed to execute structured decisions inside production code. The company was founded by Diogo Almeida, an OpenAI veteran and early contributor to ChatGPT.

Unlike conversational language models that generate text token by token, Jev accepts an unstructured state and returns typed, structured values through a parallel query. TypeSafe says this design can reduce syntax failures, output-parsing pipelines, and some invalid outputs in automated workflows. Those claims currently rely largely on the company’s descriptions and evaluations.

Parallel sampling and RLCD

According to AI News, Jev is trained with a method TypeSafe calls Reinforcement Learning for Calibrated Decisions, or RLCD. The aim is for confidence scores to track the actual accuracy of decisions. Its hardware-aware sampler evaluates structured options simultaneously and, according to the company, supports selections among as many as 255 discrete choices.

Jev’s outputs are constrained to predefined schemas. It is therefore aimed not at general-purpose prose but at tasks such as automated branching, option selection, output verification, and real-time feature extraction—areas where hand-written rules can become brittle and conversational models may add latency and uncertainty.

Performance figures need context

TypeSafe reports end-to-end latency of 70–500 milliseconds from its West Coast servers and says some fixed workflow evaluations ran up to 193.6 times faster than conversational baselines. It also lists input pricing of $0.042 per million tokens and no separate metered charge for structured output tokens. These are company-published or internal measurements and should not be generalised across workloads without independent testing and full benchmark details.

Demonstrations and early access

Technical demonstrations include a reactive Doom agent operating at ten queries per second and a Wikiracing test that selected links through dense navigation paths. The company also identifies large-scale data processing, verification layers, and automated branching logic as target applications.

TypeSafe has opened early developer access and begun onboarding engineering teams from its waitlist. Jev’s practical value will become clearer when independent results, calibration quality, and performance across varied real-world workloads are available.