TypeSafe AI pitches Jev as non-hallucinating decision model

TypeSafe AI, a lab run by ex-OpenAI researcher Diogo Almeida, has come out of two years in stealth with Jev, the first of what it calls System One Models: small, schema-bound models that return a typed decision plus a calibrated confidence score in 70 to 500 milliseconds, priced at $0.042 per million input tokens with output free. The trick is that Jev never generates free text. Every possible output is declared upfront, so there is nothing to hallucinate, and the company leans on a reinforcement learning variant it calls RLCD to keep those probabilities honest. The pitch lands on the same problem Anthropic, OpenAI and a crowd of agent-framework vendors have been circling: general LLMs are too slow and too erratic to sit inside conditional logic. If the 40x speed and cost claims hold up outside TypeSafe's own benchmarks, the more interesting shift is teams routing chat models through a cheap classifier for guardrails rather than trusting the big model alone.