evaluation/query_family/conditional/analysis_report.md

Conditional Breakdown Report

Scope

• Source analysis run: 20260526_v2_official20_plus_c2patch_plus_subgroupconditionalrepair_merged49
• Family analyzed: conditional_dependency_structure
• Excluded models: cdtd, codi, goggle
• Included models: 11 from the frozen README roster
• Deduplicated dataset-model panels: 510
• Conditional query rows used: 25108

Canonical and derived views

• Canonical score: mean(dependency_strength_similarity, direction_consistency, slice_level_consistency)
• Derived subgroup score: mean(direction_consistency, slice_level_consistency)
• The derived score is not a replacement for the frozen contract; it isolates the two subgroup-sensitive conditional branches for paper analysis.

Main findings

1. RealTabFormer is the strongest model on the subgroup-facing conditional view with mean derived subgroup score 0.748.
2. Canonically, RealTabFormer leads the full conditional family with mean conditional score 0.671.
3. TVAE is the most direction-heavy model (direction minus slice = 0.167), while RealTabFormer is the most slice-heavy (0.070).
4. TVAE shows the largest strength-to-subgroup drop risk: its dependency-strength mean exceeds its subgroup-facing mean by -0.197.
5. Dataset difficulty is uneven: n9 is hardest on the subgroup-facing conditional view (0.000 mean across models), while n11 is easiest (1.000).

Files to use first

• figures/conditional_subgroup_tradeoff_scatter_main.pdf
• figures/conditional_strength_vs_subgroup_bridge.pdf
• figures/conditional_branch_dumbbell_main.pdf
• tables/conditional_model_summary_generated.tex
• data/model_summary.csv

README compliance note

• All plotted models are restricted to the frozen README roster with fixed colors.
• Scatter plots now use legends instead of point-side model labels, matching the README figure annotation rule.
• Model order is fixed globally instead of being re-sorted by score.

Prefix note

• Prefix coverage summary rows: 33
• Prefix-level figures are exported for c / m / n slice checks, but the paper-facing core keeps the full deduplicated panel.