rlxp-reward-reflection

To diagnose a reinforcement learning reward run, you analyze local task metrics and reward-component trends to detect saturated, inactive, dominating, or conflicting reward terms. This evidence-backed reflection grounds the analysis in local artifacts rather than training rewards alone.

You create a reward reflection JSON by analyzing local run artifacts for the active RLXP contract and generating a structured output with the diagnosis and next recommended action. This confirms the run scope and compares reward behavior against approved guardrails.

Reward engineering reflection is the process of explaining why a reward-related run succeeded or failed using local task metrics. You need it after a reward-engineering run to decide whether the next iteration requires scalar reward tuning, reward-code changes, or a different validation path.

Yes, you can use local task metrics to detect dominating or conflicting reward terms. The analysis evaluates reward-component trends from local artifacts to identify inactive or saturated terms and compare them against task metrics and approved guardrails.

Yes, you need a local metrics artifact along with run and contract context and known reward keys. These inputs are required to perform scope-aware analysis and generate a structured reward reflection with a recommended next action.

Diagnosing reward runs without approved guardrails limits your ability to confirm the run belongs to the correct task or study contract. Without these guardrails and a local metrics artifact, the reflection cannot reliably validate reward behavior or recommend the next change.