rlxp-define-metrics

To define reinforcement learning metrics beyond training reward, establish a primary outcome metric, add component guardrails, and create a scorecard for evaluating task progress and comparing experiments fairly.

Guardrails and diagnostics in RL evaluation are component metrics, worst-bin checks, lower-quantile checks, and failure-mode tracking used to measure specific performance boundaries and identify issues beyond the primary training reward.

You need a scorecard for evaluating RL improvements when success criteria are ambiguous, baseline metrics are incomplete, or you require normalized component and diagnostic measures to make defensible accept or reject decisions.

To ensure RL experiment comparisons remain fair and stable, maintain a stable evaluation protocol with metric provenance, evaluation outputs, and failure modes so metric specs, candidate validation, and result audits stay comparable.

Yes, this defines task-scoped reinforcement learning metrics, guardrails, and scorecards that apply to one specific task and one study to measure progress without changing the protocol mid-study.