Skip to main content
Released in Agent Eval 1.0.0; none of this is in 0.18.0. It applies to the generated EARS and RFC 2119 paths in agent-judge-ai-core 1.0.0.
A generated audit is one call, and its workspace slot is meant for a path the judge reads with its tools, not pasted source. Before calling, every generated EARS and RFC 2119 judge and roster jury measures the whole request and refuses one that is too large.

What is measured, and where

  • What: the whole request: the requirement headers plus the rendered prompt.
  • Which builders: EarsJury, Rfc2119Jury, EarsJudge and Rfc2119Judge, for single judges as well as rosters.
  • The default limit: PromptLimits.DEFAULT_MAX_PROMPT_CHARACTERS, 100,000 characters. It is configurable.

Above the limit

No call is made. The result is ERROR with reason code judge_failed, never FAIL: the requirement was not judged, so it has not failed. The reasoning states the size, the limit and the parts:
A refused roster is an ordinary result: it concludes INCONCLUSIVE, retains no invocation because no call was made, and stores and reads back exactly through the requirement codecs.

Set another limit

Call .maxPromptCharacters(n) right after .runtime(...). The value must be positive; otherwise the builder throws IllegalArgumentException. Here judgeModel is the application’s configured EvalModel:
This block is identical to the one in the agent-judge README, which CompileGrammarTest (agent-judge-assertj) compiles. The behaviour is covered by PromptSizeGuardTests and PromptRefusalStorageTests.

Who must act

Behaviour change. A caller that pastes source or other large evidence into the workspace slot now gets ERROR instead of a call. Pass a path the judge can read with its tools instead, or raise the limit. See also Requirements judges and What’s New.