Optional Jev assessor
AgentPlat now includes an optional source integration for TypeSafe Jev. It lets an application ask typed questions about bounded state and use the structured answers as assessment signals in an AgentPlat Inference Control path.
What this is for
Jev is designed for software-facing decisions such as classification, scoring and routing. Its Choice, Score and Noul questions return structured values instead of free-form text. See TypeSafe's question primitives (opens in a new tab) and confidence semantics (opens in a new tab) for the provider contract.
In AgentPlat, the optional @agentplat/assessor-typesafe adapter implements the existing ControlledAgentAssessorV1 port. The application chooses which state to send, defines the questions, and maps the answers to an existing assessment disposition. A Jev result does not grant a capability, authorize an external effect or replace a human approval. The Inference Control guide describes the controls that remain in force around the integrated call.
Example use
An Agent Room proposal workflow can evaluate a particular proposal version against requirements such as “does this include a weekly human review?” and “does it define a stop decision after four weeks?” The result can help a person focus review. The reference proposal example (opens in a new tab) keeps the automated review advisory and records a separate human decision.
The proposal demo uses deterministic local rules by default and makes no external model call. Applications can opt into Jev on the server, where they can configure a pinned model, explicit API key and a spend cap. The adapter bounds request and response size, concurrent calls, duration and retries. Its optional evidence sink records content-free metadata such as the resolved model, usage, target digest, disposition and safe failure code.
When Jev is enabled, the selected proposal content and rubric are sent to TypeSafe. Only send data your application is authorized to share. A spend reservation uses the configured token bound and price, and the application must verify the provider's current pricing and context limits.
Availability and evidence
@agentplat/assessor-typesafe is currently source-only and is not published on npm. Review the adapter source (opens in a new tab) and integration plan (opens in a new tab) for local installation and status.
Adapter fixtures and the deterministic Room demo verify local behavior and the provider response contract. No live Jev quality or cost comparison has been collected, so AgentPlat makes no accuracy, latency or savings claim for this integration. The component maturity matrix separates implemented source from package distribution and operational evidence.