TypeSafe's Jev returns typed decisions with probabilities, not text — here's where it fits in agent loops

prakersh · reddit · 2026-09-21

TypeSafe shipped Jev, a non-conversational model that takes program state plus typed questions and returns typed answers with probability distributions in one parallel pass — no text generation. Three primitives: Noul (yes/no probability), Choice (one of up to 255 options with distribution and confidence), and Score (ordinal levels). Specs: 64K context (32K for state + longest question), text only, $0.042/M input tokens, output free.

Where it fits: agent-pipeline decision points currently built as small LLM calls with JSON parsing and retries — routing, guardrails, triage, loop control. The real value is per-answer probabilities: set confidence thresholds and escalate low-confidence calls to stronger models or humans.

Limitations (TypeSafe's jaggedness page is unusually candid): unreliable counting, poor on hex/RGB, can't judge numeric proximity, reads dates as text, accuracy degrades with irrelevant state content, and P(statement) + P(not statement) need not sum to 1 — don't build logic assuming otherwise.

Headline numbers caveat: the 193.6x faster / 444.6x cheaper figures come from TypeSafe's own harness, where 'correct' is the average of GPT-6 Astra and Claude Fable 5.1 at high thinking, measured against reasoning-off configs; the funding press release says only up to 100x. Also separate cost per token from cost per decision — a short prompt to a cheap model can win for one small question.

Availability: direct API is waitlisted, but it's testable today via Vercel AI Gateway and Cloudflare Workers AI.

Original post →

TypeSafe's Jev returns typed decisions with probabilities, not text — here's where it fits in agent loops

More from coding & agent