TypeSafe, founded by a ChatGPT co-inventor, has left stealth and is launching its Jev model to automate programmatic decisions with parallel sampling architecture.
Software systems requiring automated deterministic logic can now bypass conversational language models following TypeSafe AI’s release of Jev, a specialised System One Model designed to execute structured probabilistic decisions directly inside production codebases.
Developed during two years in stealth by OpenAI veteran and ChatGPT co-inventor Diogo Almeida, the model abandons text and string generation entirely.
Instead of autoregressively generating sequential text tokens, the system takes an unstructured state as an input and outputs type-safe structured values in a single parallel query. The architecture targets deterministic code integration, preventing syntactic type failures and output hallucinations by design.
Diogo Almeida, Founder at TypeSafe, said: “Think of Jev as a frontier-intelligence function call: unstructured state in, typed probabilistic decisions out.”
Jev model architecture: How hardware-aware parallel sampling works
Engineers built the platform around an alternative training methodology termed Reinforcement Learning for Calibrated Decisions (RLCD).
Conventional frontier models train via Reinforcement Learning with Human Feedback (RLHF) or Reinforcement Learning with Verifiable Rewards (RLVR) to satisfy human conversational preferences or isolated programmatic checks. TypeSafe trains Jev to produce calibrated probabilities on execution logic, ensuring returned confidence scores correlate directly with output accuracy.
Departing from autoregressive token generation, Jev relies on a hardware-aware parallel sampler that evaluates and delivers all structured values simultaneously.
The structural constraint restricts outputs to predefined schemas, eliminating the parsing pipelines and guardrail wrappers required when deploying traditional large language models across business automation flows. The internal sampler handles high-cardinality selections up to 255 discrete options, using a two-stage process that scores options independently prior to selection.
Internal evaluations conducted from the team’s West Coast servers record end-to-end response latencies between 70–500 milliseconds, contrasted against the 3–329 seconds observed across conversational frontier deployments.
Published workflow evaluations run across fixed compute graphs (measuring multi-step decision branching against consensus baselines from GPT-6 Astra and Fable 5.1) recorded execution speeds up to 193.6 times faster.
Input processing costs are priced at $0.042 per million tokens, down from standard conversational rates spanning $0.20 to $10 per million tokens. Because parallel sampling generates structured states without autoregressive token generation passes, TypeSafe provides output tokens without metering charges.
From Doom bots to petabyte data: Real-world Jev deployment testing
Production demonstrations show Jev resolving branching rules across high-speed game states and web traversal trees. In dynamic stress testing, engineers operated a real-time reactive bot across Doom game state structures at 10 queries per second, generating operating costs of approximately $7 per hour.
A secondary Wikiracing test evaluated link selection across dense encyclopaedia directories, finishing traversals in fewer steps than external non-reasoning models by evaluating options without generating hallucinated dead ends.
Field testing confirms primary target applications across real-time feature extraction, petabyte-scale data workflows, output verification layers, and automated branching logic where hand-written rules prove brittle.
TypeSafe opened early developer access today and started onboarding engineering teams from its deployment waitlist.