Output

Every run produces a timestamped directory under <project-dir>/output/. The files it contains serve different purposes — from raw forensic evidence to aggregated summary statistics.

• File overview — directory structure, file types, and log file locations
• manifest.csv columns — full column reference for the per-request output file
• Summary CSVs — summary_counts.csv and summary_pct.csv structure and the “perfect run” definition
• Scoring — how list task metrics (exact_matches, coverage, hallucinations, position_score, min_moves) are computed
• Interpreting results — how to read the numbers as research evidence: comparing deployments, identity groups, and drawing reproducibility conclusions