Class PlanAdherenceEvaluator

java.lang.Object
dev.dokimos.core.BaseEvaluator
dev.dokimos.core.evaluators.agents.PlanAdherenceEvaluator
All Implemented Interfaces:
Evaluator

public class PlanAdherenceEvaluator extends BaseEvaluator
Evaluates whether an agent's executed tool calls followed its stated plan using a rule-based check and an optional LLM check.

The plan is read from actualOutputs[reasoningStepsKey] (default "reasoningSteps") as a List<String> and the tool calls from actualOutputs[toolCallsKey] (default "toolCalls"), matching AgentTrace.toOutputMap().

Checks performed:

  • plan_present (rule): a plan is present
  • plan_followed (LLM): the executed tool calls followed the stated plan

Without a judge LLM, only the rule-based check runs. The score is the fraction of checks that passed. No tool calls short-circuits to a score of 1.0, and a missing tool calls key throws.

  • Method Details