jevlisted

# Jev (TypeSafe System One) Jev takes a **state** (text or JSON) and a map of **typed questions**, and returns one typed answer per question with calibrated probabilities. It does not generate text, write code, or reason step by step. Think of it as a programmable `if` for things ordinary code cannot judge: "is this urgent", "which team", "how risky". The default route is **OpenRouter**; TypeSafe direct also works (`--provider typesafe`, `TYPESAFE_API_KEY`): - Endpoint: `POST https://openrouter.ai/api/alpha/decisions`. It's under `/alpha/`, so it may change — a 404 means check OpenRouter's Jev model page before debugging anything else. - Headers: `Authorization: Bearer $OPENROUTER_API_KEY`, `Content-Type: application/json`; optional `HTTP-Referer` (site URL) and `X-Title` (site name) for OpenRouter rankings. - Model: `typesafe/jev-1.13` — pin it. A floating alias exists (`~typesafe/jev-latest`, with the tilde — without it, it doesn't resolve here) but move to it deliberately. TypeSafe's direct route pins the fuller `jev-1.13.0`. Responses name the dated snapshot that answered (`typesafe/jev-1.13-20260917`): log it, and match your pin by prefix. - Context: **32,000 tokens for state + all questions combined** on this route (TypeSafe direct differs: 64k total, 32k for state + the single longest question). The bundled script enforces the limit for whichever route it's actually calling. - Body and answers are the same shape as TypeSafe's native API: `state`, `mode