evaluation/tail/breakdown/v2/analysis_report__v2.md

Tail Breakdown Report

Scope

• Source tail-threshold run: 20260519_server_main_refresh_tail_full
• Excluded models: cdtd, codi, goggle
• Included models: 11
• Deduplicated dataset-model panels: 466
• Threshold count per panel: 10-10

Canonical tail views

• Canonical tail-threshold components reused directly: tail_set_consistency, tail_mass_similarity, tail_concentration_consistency.
• tail_coverage_score = mean(tail_set_consistency, tail_mass_similarity)
• tail_breakdown_score = mean(tail_set_consistency, tail_mass_similarity, tail_concentration_consistency)
• coverage_minus_concentration = tail_coverage_score - tail_concentration_consistency

Main findings

1. ARF is strongest on tail concentration with mean tail concentration score 0.524.
2. BayesNet is strongest on tail coverage core (tail_coverage_score) with mean score 0.311, while ARF leads the three-part tail breakdown overall at 0.376.
3. TabSyn is the most coverage-heavy model (coverage minus concentration = -0.124), while TabPFGen is the most concentration-leaning (-0.237).
4. Dataset difficulty remains uneven: c2 is hardest on tail concentration (0.000 mean across models), while n10 is easiest (0.946).

Files to use first

• figures/tail_coverage_vs_concentration_scatter_main.pdf
• figures/tail_coverage_vs_breakdown_bridge.pdf
• figures/tail_prefix_bars_appendix.pdf
• tables/tail_model_summary_generated.tex
• data/model_summary.csv

Prefix note

• Prefix coverage summary rows: 33
• The c / m / n split is exported explicitly because tail concentration behavior differs by dataset family, not just by overall model average.