Agent Eval 0.18.0

Configure requirements and evidence before running a Judge or Jury, retain the complete
assessment, and let application policy decide whether to rely on it.

• Configured execution: typed builders produce ready Judge.judge() and Jury.vote() operations.
• Executable requirements: native RFC2119/EARS requirements, one generated investigation per roster, and typed Jev evaluation.
• Complete results: original judgments, child verdicts and native execution facts survive composition; policy decides RELY, ABSTAIN or ESCALATE independently. AssertJ stages cache execution.
• Durable storage: explicit Jackson 2 codecs retain Verdict/EvaluationResult data in V6; retained reads and assertions execute no evaluator.

Breaking changes: recompile 0.17 consumers and migrate to configured builders and the
new result packages. V6 readers refuse older typed results; preserve historical records.
See the migration guide and MIGRATION_0.18.md.

Java 21. Maven coordinates remain io.github.markpollack:agent-judge-*.

Try the PetClinic walkthrough,
which demonstrates the APIs using archived responses and offline replay.

Documentation

• What's New
• Getting started
• Requirements judges
• Interpreting verdicts
• Migrating to 0.18

Maven Central

io.github.markpollack:<artifact>:0.18.0, 14 modules plus agent-judge-parent. New in 0.18: agent-judge-json-jackson2, agent-judge-jev, agent-judge-assertions, agent-judge-assertj.

<dependency>
    <groupId>io.github.markpollack</groupId>
    <artifactId>agent-judge-core</artifactId>
    <version>0.18.0</version>
</dependency>

Full Changelog: v0.17.0...v0.18.0