analyze-rl-behavior

You can analyze RL behavior changes by running a local pipeline that synthesizes evaluation traces, training logs, and GPT-5 assessments to compare pre-RL baselines with RL checkpoints. This generates a structured behavioral report detailing changes and persistence.

Yes, you can run reinforcement learning evaluation trace analysis locally on a Mac without a GPU. The pipeline is specifically designed for Mac-based environments to orchestrate behavioral analysis and generate structured Q1–Q4 reports.

The best way to automate behavioral analysis reporting is to use an orchestration pipeline that synthesizes model traces and evaluation metrics into structured Q1–Q4 outputs. It automatically writes results to dedicated output directories for reproducibility.

Yes, the behavioral analysis pipeline supports optional GPT-5-based judgement and failure-mode annotation. This feature is available when your evaluation repositories are writable, allowing for deeper model assessment.

Comparing RL checkpoints to baselines generates Q1–Q4 outputs including behavioral_delta, llm_judge_diff, temporal_trace_analysis, and parse_skyrl_metrics. These results are written to dedicated output directories for reproducibility.

Yes, preflight artifact checks are performed automatically when you run the RL behavioral analysis script. This ensures all necessary evaluation traces and training logs are present before orchestrating the local analysis pipeline.