# Guardrails can use typed judgments as their check engine

> LLM judge and moderation guardrails can run on calibrated probabilities from TypeSafe judgments, batching every check on a stage into one request.

Source: <https://everruns.com/docs/changelog/2026-09-17-guardrails-typed-judgment-engine/>

Date: 2026-09-17
Products: Platform
Category: Feature
Available: In 0.29.0
Pull request: everruns/everruns#3644

The `llm_judge` and `moderation` guardrails take an `engine` setting: `utility_llm` (the default, unchanged) or `jev`. With `jev`, every check on a stage is answered in one request, and the verdict is the configured `threshold` applied to a calibrated probability rather than a JSON verdict parsed back from text. A check can no longer pass because a response failed to parse.

Operators enable it with the `UTILITY_TYPESAFE_API_KEY` deployment credential. When it is unset, jev checks are skipped with a log line and the turn proceeds. An experimental `jev` capability with a `jev_evaluate` tool is available in development registries.

See [Guardrails](https://everruns.com/docs/capabilities/guardrails/) and [TypeSafe](https://everruns.com/docs/integrations/typesafe/). [#3644](https://github.com/everruns/everruns/pull/3644)