API reference

Factory names in jes.policies and jes.recipes are stable, and so is the text of every question id: a wording change gets a new question id.

Module	What it holds	Page
jes	Guard, AsyncGuard, Limits, Redactions, Result, the question types, the result types, and the errors	Guard, Results, Redactions, Errors
jes.guard	Guard, AsyncGuard, and History	Guard
jes.limits	Limits	Limits
jes.result	Result, refusal, and finding_name	Results
jes.policies	Core policy factories, judge, and the policy protocols	Policies
jes.questions	YesNo, Choice, Score, their answers, and Threshold	Questions
jes.backend	TypeSafe, ModelSpec, resolve_model, Request, Reply, and the backend protocols	jes.backend
jes.recipes	Recipes: pre-made policy factories that haven't been evaluated yet	Recipes
jes.testing	FakeBackend and Rule, for tests	Testing
jes.types, jes.errors, jes.redactions	The result types, type aliases, exceptions, and TOKEN_RE. jes re-exports most of them.	Results, Errors, Redactions

jes.__version__ is the installed version string, for example "0.0.1".

Conventions

• Every judgment factory requires threshold=, as a float or a Threshold. jes ships no default thresholds.
• stages defaults differ per policy. Each stage is one of "input", "untrusted", "tool_call", "tool_result", or "output". Transforms and judge() take stages=. The built-in judgments and allowed_tools have fixed stages.
• name= must be unique within a guard and match [A-Za-z][A-Za-z0-9_-]{0,63}.
• model= on a policy overrides the guard's model for that policy only. It takes a ModelSpec: a TypeSafe model id such as "jev-latest", a classifier, or a backend.