Decomposed Reward Signals (CONCEPT:AHE-3.1)

Overview

Separates step-level reward (local constraint satisfaction) from trajectory-level reward (goal achievement) for accurate credit assignment. Feeds into ExperienceNode distillation (CONCEPT:AHE-3.1).

Implementation Details

• Source Code: agent_utilities/graph/reward_decomposition.py
• Pillar: AHE

Documentation Coverage

This is an auto-generated dedicated concept page to ensure 100% documentation coverage across the ecosystem.

Multi-Strategy EvalRunner (CONCEPT:AHE-3.1)

Overview

Multi-strategy evaluation runner (exact match, semantic similarity, LLM-as-Judge) with composite scoring and EvaluationMonitor integration. Ported from MATE's eval_runner.py. OWL-enabled degradedPerformance inference across sessions.

Implementation Details

• Source Code: agent_utilities/harness/continuous_evaluation_engine.py
• Pillar: AHE

Documentation Coverage

This is an auto-generated dedicated concept page to ensure 100% documentation coverage across the ecosystem.