System One decisions your code can use

Large language models produce text for humans to read. You need judgments you can branch on in code: route this ticket, score this lead. Forcing a text generator into that slot means you generate prose, parse it back, and repair the parse when wording drifts. You skip that round trip with Jev.

Jev is TypeSafe AI's System One model. TypeSafe launched it on September 15, 2026 with a $40M DCVC seed round. The founders are Diogo Almeida, ex-OpenAI RLHF, Erik Gafni, and Sasha Sheng. System One models return fast, structured decisions. You send state plus typed questions. You get typed values and probability distributions back. You branch on the answers in code. See docs.typesafe.ai/introduction.

Training differs from frontier LLMs. Frontier LLMs optimize with RLHF/RLVR for human preference: agreeable writeups and chat. Jev trains with RLCD, Reinforcement Learning for Calibrated Decisions. RLCD tunes the distribution shape itself. A returned 0.88 means correct in about 88 of 100 runs. You run each question in parallel against the same state in one pass. You add questions to the same pass with little added latency and no context rot. See typesafe.ai/blog/introducing-system-one-models-and-jev.

Three question types cover the decisions. Choice picks one of up to 255 options. Include other or none_of_above when the set has gaps. Choice returns choice + probabilities + confidence. Score places the state on an ordered rubric of 2 to 10 levels and returns the probability-weighted mean plus per-level probabilities. Noul returns a single calibrated noul in [0, 1]. The number is the absolute truth probability, with no separate confidence field. 0.5 means equal odds, not medium. You mix all three in one call. You keep composition, weights, and thresholds in your code.

You pay $0.042 per 1M input tokens with output free, at about 100ms typical. Observed latency is 70 to 180ms. The vendor range is 70 to 500ms. Context is 64k total with 32k max for state plus longest question. Throughput is 250k tokens per second and 1,200 requests per minute. A frontier LLM runs about 2,500ms and about $0.50 per 100k tokens. The reference build runs about 25x faster and costs about 125x less per decision. Test the vendor headline multipliers of 193x and 444x against your own labeled workload before you cite them.

You connect through the Vercel AI Gateway with AI SDK experimental_evaluate against typesafe-ai/jev. You authenticate with AI_GATEWAY_API_KEY. Question types map one to one. choice maps to Choice, score maps to Score, and boolean maps to Noul. See vercel.com/i/what-is-jev and vercel.com/kb/guide/typesafe-jev-and-ai-sdk.