What Is Jev? Inside TypeSafe's Decision-Only AI Model

The prior art is real, and Firecrawl links skeptical threads pointing straight at encoder-style zero-shot classifiers. What the explainer adds is a hosted decision interface with per-option probabilities; whether that earns the price is the open question.

As an interface, yes. As a calibrated replacement, no. Firecrawl treats these as clones of the call shape rather than reproductions of hosted calibration, so verify against your own labels before swapping anything in.

The explainer separates a waitlist route from Vercel AI Gateway access and also covers the AI SDK evaluate path. It is vendor-adjacent, so confirm current pricing and gateway model names against official docs rather than the post.

Not really. It is an explainer of primitives and access paths rather than a measured eval, and the author flags the blog as vendor-adjacent. For calibration evidence you need independent tests, not this page.

Choice asks which of these options and returns the chosen option, a probability for every option and a confidence score. Score asks where on a rubric and returns a numeric position, a probability per level and a confidence. Noul asks is this true and returns a single probability from 0 to 1. Every question is evaluated in parallel and in isolation against the same state, so adding a question barely changes response time.

A coding-agent guardrail, an MCP server, and an open-weights clone of the interface. The guardrail example is pi-warden, which before each bash, write or edit sends the task, the agent's stated plan and the pending command to Jev as four typed questions: is this irreversible, is it off-task, does it mutate anything, and what scope is it. It answers all four in about 250 ms, held 42 times over 17,000 recorded calls, and roughly 88 percent of those holds were judged right.

Frontier LLMs are quoted at 3 to 329 seconds on reasoning tasks with $0.20 to $10 per million input tokens and output priced at roughly five times input; Jev is 70 to 500 ms at $0.042 per million with free output. TypeSafe opened signups to everyone on September 20 with $5 in free credit, then paused new signups on September 22 under demand. Vercel AI Gateway from September 16 via AI SDK 7's evaluate, and OpenRouter, do not go through that signup.