TypeSafe & Jev: Company, Team, and Verified Claims
TypeSafe AI is the developer and commercial operator of Jev. The company position is that general-purpose autoregressive LLMs are structurally ill-suited for high-throughput software decision routing due to hallucination, non-determinism, and slow token generation. Jev was built to replace generative reasoning with calibrated, non-autoregressive decision classification.
Editorial Disclosure: Jev Bureau is an independent technical reference. We are not TypeSafe AI, we do not receive compensation from TypeSafe AI, and this page distinguishes official vendor claims from officially documented operational behavior.
TypeSafe Leadership Team
Verified from official TypeSafe company disclosures (September 2026).
Officially Documented Specifications & Platform Deployments
Status: Verified against primary documentationPOST /v1/systemone accepts shared state string and questions array, returning strict JSON evaluation objects.
TypeSafe documents that Jev outputs only categorical selections, probability distributions, continuous rubric scores, and binary truth probabilities. It does not generate prose, chat, or free-form token sequences.
TypeSafe documents a 64k-token total budget per request across state plus all questions combined, with a 32k-token cap for state plus the single longest question. OpenRouter lists Jev 1.13 with a 32,000-token context window in its model catalog.
Available on OpenRouter via model ID 'typesafe/jev-1.13' (and alias '~typesafe/jev-latest') and on Vercel AI Gateway via 'typesafe-ai/jev'.
Billed at $0.042 per 1,000,000 input tokens; $0.00 for output tokens across both TypeSafe direct and OpenRouter.
Vendor Claims & Evaluation Metrics
Status: TypeSafe reported metrics; independent audits pendingTypeSafe reports that Jev was trained using a custom RLCD framework to align output probabilities directly with empirical ground-truth error frequencies. TypeSafe notes that calibration is measured across groups of predictions and does not guarantee that an individual answer is correct. First-party empirical calibration audits by Jev Bureau have not yet begun.
TypeSafe reports internal cluster response times of 70ms to 500ms on early-access infrastructure. TypeSafe notes that published evals were run from laptops on the West Coast. Independent multi-region latency benchmarks have not been conducted by Jev Bureau.
TypeSafe reports that in internal workflow evaluations comparing decomposed multi-step agent decisions, Jev was up to 193.6× faster and 444.6× cheaper than external LLM baselines wrapped in structured adapters. TypeSafe notes these figures represent the high end of achievable gains on decomposed workloads.
TypeSafe states that schema matching is guaranteed rather than empirical because Jev is constrained to predefined Choice, Score, or Noul structures. However, structural schema adherence eliminates JSON parsing syntax errors but does not guarantee semantic correctness or protect against misclassification.