Fast judgment calls belong in typed tools that fail closed, not in frontier models

Agents skip cheap checks because frontier models are too slow and expensive to run on every page, claim, or candidate. This MCP server addresses that directly by wrapping TypeSafe's Jev model — a "System One" model that returns calibrated probability distributions rather than generated text — in ten typed judgment tools, each completing in roughly 150 to 500 milliseconds for a fraction of a cent.

The tool set covers the full lifecycle of an agentic task. Before work begins, jev_screen checks fetched content for prompt injection and flags low-substance pages worth skipping. jev_classify and jev_decide can route the task to the right workflow. During execution, jev_find and jev_rerank handle retrieval without embeddings or an index — the README cites CLERC benchmark numbers where this reranking pattern lifted top-1 retrieval from 5% to 18%. jev_extract pulls structured fields using your regex to find candidates and Jev to pick among them, so the model selects a value but never writes one. At completion, jev_review scores a diff on four rubrics — correctness, spec match, test gap, and blast radius — and jev_gate combines that review with claim verification against evidence you supply.

The failure modes are handled carefully. Every tool has an explicit invalid_response path that fails closed rather than treating a malformed model answer as confident output. jev_extract never hallucinates: if the regex finds nothing, the field returns not_found without touching the model. jev_decide includes escape hatches (ask_user, investigate) for when evidence is genuinely insufficient. The transport layer retries only on a conservative set of status codes and never retries ambiguous network failures, because without an idempotency key a re-send can double-charge a paid call.

Provider flexibility is real: TypeSafe direct, OpenRouter, Cloudflare Workers AI, Vercel AI Gateway, or any compatible System One endpoint. The README is candid that thresholds are starting points from TypeSafe's own cookbooks and need tuning against your data, and that typed output guarantees the interface, not the truth.

This is early software by the author's own description. But the design discipline — typed outputs, explicit invalid paths, policy staying in your code — is the right shape for judgment tooling that agents can actually trust.