llm11

Jev

Jev is TypeSafe AI’s “System One” decision model: instead of generating text, it takes a state and a set of typed questions and returns typed, calibrated answers in one parallel pass - no autoregression. 70–500ms, $0.042 per million input tokens, output tokens unmetered.

Jev is built as a pluggable triage backend on this deployment but is not yet live - we have early access to TypeSafe AI but a production key isn't wired into this environment yet. Every request is triaged by our own heuristic engine until it is, and we say so on every receipt (see "triage backend").

Why triage matters here

llm11’s core mechanic is that the triage call decides how much verification a request deserves - not just which model answers it. Jev’s calibration is what makes that decision trustworthy: independent testing has found its confidence scores genuinely track real-world accuracy (a claim scored below 0.1 confidence really is wrong the overwhelming majority of the time), which means we can gate the expensive verification rungs on a number that actually means something, rather than an LLM judge’s self-reported vibe.

Why it’s not the whole architecture

Jev is one implementation of a pluggable DecisionBackend interface, not a hard dependency. A single-vendor, recently-launched, early-access API is a real business risk to build an entire product around, so the triage layer is designed to degrade to a conservative built-in heuristic rather than fail if Jev is unavailable, and could take a different backend entirely without changing anything downstream.

What is a “System One model”? →