Typed, Calibrated Judgments for Routing, Guardrails and LLM Evaluation, from Probe Harness to Production

"Most of the model calls in a production AI system are not asking for prose. They are asking for decisions."

They ask whether an inbound customer ticket is about billing. Whether a financial transaction exhibits fraud patterns. Whether a retrieved document chunk supports the claim. Whether an LLM reply violates compliance policy. Whether an autonomous agent should click, scroll, or escalate.

Each of these is fundamentally a System One decision: a small, bounded, deterministic judgment whose outcome space is one of a handful of predefined values your code already knows how to process.

Yet for three years, the entire software industry has answered these bounded questions with massive, stochastic generative models. We have taken multi-billion-parameter LLMs designed to write Shakespearean prose, asked them to return a single JSON string, parsed that string back out of free text with regex, retried on schema hallucinations, waited seconds per token, and paid exorbitant frontier prices for every output token of the detour.

In September 2026, TypeSafe AI launched Jev: a breakthrough System One model that fundamentally changes this paradigm.

Jev does not generate text. It accepts an unstructured state and a set of typed questions, and produces—for every question simultaneously in a single forward pass—a calibrated probability distribution over answer spaces defined before the call:

• Off-schema outputs are mathematically impossible by construction.
• Pricing is strictly input-based at $0.042 per million tokens.
• Latency sits at tens to hundreds of milliseconds, almost flat regardless of question count.
• Trained via RLCD (Reinforcement Learning from Calibrated Decisions) to produce true probabilities rather than sycophantic human-preference text.