A benchmark that separates fixtures from measured results
The repository contains a deterministic 30-figure, 60-claim generator with exact ground truth. End-to-end extractor measurements are intentionally marked pending until a recorded run produces a versioned result artifact.
ConditionFigures
Clean linear chart8
JPEG / low resolution6
Dashed lines + markers6
Series crossing4
Log10 Y axis4
Similar series colors2
Measured metrics
series_nrmsefinal_point_relative_errorratio_relative_errorclaim_verdict_accuracynot_supported_f1mean_human_correctionsmean_tool_callsworkflow_success_ratemaximum_overlay_deviation_px