TypeSafe Releases Jev Model for Non-Autoregressive Programmatic Logic

Bypassing Text Generation for Programmatic Decisions

Software pipelines frequently call conversational frontier models merely to select an action or pick an enum value. TypeSafe AI, led by ChatGPT co-inventor Diogo Almeida, emerged from stealth with Jev—a specialized System One model architected specifically to eliminate strings from the decision loop. Rather than predicting token sequences, Jev consumes unstructured application state and evaluates structured schemas in a single parallel query.

Parallel Sampling and RLCD Calibration

Instead of relying on standard Reinforcement Learning with Human Feedback (RLHF), TypeSafe trains Jev using Reinforcement Learning for Calibrated Decisions (RLCD). This approach ensures confidence scores correlate mathematically with logic execution accuracy. The system's hardware-aware parallel sampler handles candidate spaces of up to 255 discrete choices via a two-stage scoring mechanism, constraining outputs to predefined schemas and avoiding parsing wrappers altogether.

Latency, Pricing, and Real-World Throughput

In vendor evaluations, Jev achieved end-to-end response latencies of 70–500 milliseconds, compared to 3–329 seconds recorded for conversational frontier deployments. Across multi-step decision branching benchmarks, execution speeds reached up to 193.6 times faster than frontier baselines.

Economic metrics are equally striking: input processing costs $0.042 per million tokens, compared to typical API rates ranging from $0.20 to $10. Because outputs are resolved through parallel state sampling rather than autoregressive generation, TypeSafe does not meter or charge for output tokens. In dynamic stress testing running a Doom game bot at 10 queries per second, operational compute costs stabilized at approximately $7 per hour. Early developer access is now open via a deployment waitlist.