What problem does it solve? Raw experiment outputs such as logs, CSVs, JSON metrics, and ablation tables are hard to turn into defensible paper claims without auditing data quality, baselines, splits, and statistical support. This Skill structures that analysis so results are traceable to artifacts before they become prose. ## Core Features & Use Cases - Automated Result Scanning: The bundled scan_results.py script recursively parses JSON, CSV, TSV, TXT, and LOG files to extract common metrics (accuracy, F1, precision, recall, loss, AUC, RMSE, MAE, R2) with anomaly and scale warnings. - Registry Candidate Generation: result_scan_to_registry.py converts scan output into review-only EXP-AUTO-* rows in an experiment registry, keeping unreviewed results separate from thesis evidence. - Claim Auditing Workflow: Classifies claims as supported, weak, unsupported, or missing, with EDA gates, statistical test selection guidance, and experiment integrity checks (leakage, metric mismatch, dead metric code). - Use Case: After a training run produces scattered result folders, scan the directory, review candidate metrics, compare against baselines, and produce a claim table plus figure handoff for the paper. ## Quick Start Use the research-results-analysis skill to scan my experiment results folder and summarize which paper claims the metrics actually support.
