TypeSafe's Jev returns typed, calibrated decisions instead of text
A new 'System One' model class: Choice/Score/Noul primitives, 70-500ms replies, output tokens free, on OpenRouter and Vercel without a waitlist.
Copy markdown
Decisions, not prose
Jev takes a state plus typed questions and hands back a structured value your code branches on directly - no JSON-mode parsing, no retry loops. Outputs conform to your schema and can't hallucinate or throw a type error.
Three primitives: Choice, Score, Noul
Choice picks from up to 255 options, Score rates against ordered levels, Noul returns a calibrated true/false probability. Every answer ships with a confidence number, so you can gate on certainty instead of trusting a raw label.
70-500ms, and output is free
Replies land in 70-500ms - TypeSafe clocks a demo workflow at 0.114s against 8.57s for a frontier chat model. Input is $0.042/MTok and output is 'too cheap to meter.' The 190x-cheaper, 440x-faster claims are vendor-run, so treat them as a ceiling.
Skip the waitlist via OpenRouter or Vercel
The direct API is behind a waitlist, but Jev 1.13 is already live on OpenRouter (32K context, 99.99% uptime, no waitlist) and the Vercel AI Gateway. One catch: it runs on a Decisions API, not chat completions, so standard chat SDKs won't call it.
What to build this weekend
Early users shipped command-safety gates, email triage, browser- and phone-agent guardrails, and content scoring within days. Anywhere you're bolting brittle if/else onto an LLM's free text, a typed decision with a confidence score can replace it.