Skip to main content
This guide describes the Agent Eval 0.18.0 API. Maven coordinates and Java packages still use agent-judge and io.github.markpollack.judge.
Choose the least intelligent practical evaluator that can establish the property at the required reliability. An available deterministic check is usually preferable; a property being deterministic in principle does not mean a complete bespoke checker is available or economical. A Requirement is a pure immutable value. Evidence belongs to the caller; tools and permissions belong to the configured runtime. A ready Judge or Jury owns execution, with one judge() or vote() operation. Native execution facts remain visible beside finding/support, without inventing unknown usage or universal calibration guarantees. Producer outcomes describe what happened. Verdict conclusion describes the composed determination. Application policy describes whether to rely on it. These separate decisions let the same retained assessment support different application actions while preserving the original facts. Required coverage, opinion reduction, native roster audits and cascade routing answer different questions. Keep complete originals, refused returns, attempted stages and ownership in durable results, with explicit preservation limits rather than silent truncation. Reading a retained result executes nothing. Learn the public programming model, composition distinctions, and retention contract.