Add files using upload-large-folder tool

evaluation/query_family/conditional/locality_support_diagnostics/final/must_do/fig_conditional_support_main.tex

@@ -0,0 +1,53 @@
+\documentclass[tikz,border=4pt]{standalone}
+\usepackage{pgfplots}
+\usepgfplotslibrary{groupplots}
+\usepackage{xcolor}
+\pgfplotsset{compat=1.18}
+
+\definecolor{modelarf}{HTML}{777777}
+\definecolor{modelbayesnet}{HTML}{CCBB44}
+\definecolor{modelctgan}{HTML}{EE6677}
+\definecolor{modelforestdiffusion}{HTML}{228833}
+\definecolor{modelrealtabformer}{HTML}{332288}
+\definecolor{modeltabbyflow}{HTML}{882255}
+\definecolor{modeltabddpm}{HTML}{EE7733}
+\definecolor{modeltabdiff}{HTML}{AA3377}
+\definecolor{modeltabpfgen}{HTML}{009988}
+\definecolor{modeltabsyn}{HTML}{66CCEE}
+\definecolor{modeltvae}{HTML}{4477AA}
+\definecolor{summaryblack}{HTML}{000000}
+\begin{document}
+\begin{tikzpicture}
+\begin{axis}[
+width=13.8cm,
+height=8.4cm,
+ymin=0.0, ymax=1.0,
+title={Conditional support decomposition},
+ylabel={Filtered-local conditional fidelity},
+xtick={1,2,3},
+xticklabels={Dense,Medium,Sparse},
+ymajorgrids,
+grid style={draw=gray!20},
+major grid style={draw=gray!28},
+axis line style={draw=black!70},
+tick style={draw=black!70},
+legend style={draw=none, fill=none, font=\scriptsize, at={(0.98,0.98)}, anchor=north east},
+]
+\addplot+[mark=*, mark size=1.8pt, line width=0.9pt, draw=modelarf, fill=modelarf, opacity=0.82] coordinates {(1,0.659045) (2,0.591366) (3,0.590151)};
+\addplot+[mark=*, mark size=1.8pt, line width=0.9pt, draw=modelbayesnet, fill=modelbayesnet, opacity=0.82] coordinates {(1,0.667834) (2,0.612367) (3,0.547617)};
+\addplot+[mark=*, mark size=1.8pt, line width=0.9pt, draw=modelctgan, fill=modelctgan, opacity=0.82] coordinates {(1,0.585926) (2,0.461564) (3,0.438783)};
+\addplot+[mark=*, mark size=1.8pt, line width=0.9pt, draw=modelforestdiffusion, fill=modelforestdiffusion, opacity=0.82] coordinates {(1,0.420536) (2,0.476072) (3,0.363245)};
+\addplot+[mark=*, mark size=1.8pt, line width=0.9pt, draw=modelrealtabformer, fill=modelrealtabformer, opacity=0.82] coordinates {(1,0.830758) (2,0.745247) (3,0.712539)};
+\addplot+[mark=*, mark size=1.8pt, line width=0.9pt, draw=modeltabbyflow, fill=modeltabbyflow, opacity=0.82] coordinates {(1,0.517681) (2,0.518860) (3,0.459520)};
+\addplot+[mark=*, mark size=1.8pt, line width=0.9pt, draw=modeltabddpm, fill=modeltabddpm, opacity=0.82] coordinates {(1,0.538307) (2,0.498335) (3,0.422132)};
+\addplot+[mark=*, mark size=1.8pt, line width=0.9pt, draw=modeltabdiff, fill=modeltabdiff, opacity=0.82] coordinates {(1,0.538708) (2,0.528432) (3,0.450310)};
+\addplot+[mark=*, mark size=1.8pt, line width=0.9pt, draw=modeltabpfgen, fill=modeltabpfgen, opacity=0.82] coordinates {(1,0.609626) (2,0.613785) (3,0.505940)};
+\addplot+[mark=*, mark size=1.8pt, line width=0.9pt, draw=modeltabsyn, fill=modeltabsyn, opacity=0.82] coordinates {(1,0.593607) (2,0.582095) (3,0.492035)};
+\addplot+[mark=*, mark size=1.8pt, line width=0.9pt, draw=modeltvae, fill=modeltvae, opacity=0.82] coordinates {(1,0.517385) (2,0.392556) (3,0.323651)};
+\addplot+[mark=*, mark size=2.6pt, line width=1.8pt, draw=summaryblack, fill=summaryblack] coordinates {(1,0.590639) (2,0.547348) (3,0.483777)};
+\addplot+[only marks, mark=none, draw=summaryblack, error bars/.cd, y dir=both, y explicit] coordinates { (1,0.590639) +- (0,0.053714) };
+\addplot+[only marks, mark=none, draw=summaryblack, error bars/.cd, y dir=both, y explicit] coordinates { (2,0.547348) +- (0,0.055578) };
+\addplot+[only marks, mark=none, draw=summaryblack, error bars/.cd, y dir=both, y explicit] coordinates { (3,0.483777) +- (0,0.053077) };
+\end{axis}
+\end{tikzpicture}
+\end{document}

evaluation/query_family/conditional/locality_support_diagnostics/runs/20260519_192156_conditional_locality_support/README.md

@@ -0,0 +1,8 @@
+# 20260519_192156_conditional_locality_support
+
+This run contains the full reproducible bundle for the conditional locality/support diagnostic.
+
+- `data/` exports the summary and audit CSVs.
+- `figures/` holds the paper-facing figures plus standalone TeX sources.
+- `tables/` holds LaTeX table snippets.
+- `report/` holds the Markdown narrative, captions, and paper paragraphs.

evaluation/query_family/conditional/locality_support_diagnostics/runs/20260519_192156_conditional_locality_support/figures/fig_conditional_locality_by_model.tex

@@ -0,0 +1,65 @@
+\documentclass[tikz,border=4pt]{standalone}
+\usepackage{pgfplots}
+\usepgfplotslibrary{groupplots}
+\usepackage{xcolor}
+\pgfplotsset{compat=1.18}
+
+\definecolor{modelarf}{HTML}{777777}
+\definecolor{modelbayesnet}{HTML}{CCBB44}
+\definecolor{modelctgan}{HTML}{EE6677}
+\definecolor{modelforestdiffusion}{HTML}{228833}
+\definecolor{modelrealtabformer}{HTML}{332288}
+\definecolor{modeltabbyflow}{HTML}{882255}
+\definecolor{modeltabddpm}{HTML}{EE7733}
+\definecolor{modeltabdiff}{HTML}{AA3377}
+\definecolor{modeltabpfgen}{HTML}{009988}
+\definecolor{modeltabsyn}{HTML}{66CCEE}
+\definecolor{modeltvae}{HTML}{4477AA}
+\definecolor{summaryblack}{HTML}{000000}
+\begin{document}
+\begin{tikzpicture}
+\begin{axis}[
+width=13.8cm,
+height=8.4cm,
+ymin=0.0, ymax=1.0,
+title={Conditional locality decomposition by model},
+ylabel={Conditional fidelity score},
+xtick={1,2,3},
+xticklabels={Grouped / Global,2D Surface,Filtered / Local},
+ymajorgrids,
+grid style={draw=gray!20},
+major grid style={draw=gray!28},
+axis line style={draw=black!70},
+tick style={draw=black!70},
+legend style={draw=none, fill=none, font=\scriptsize, at={(0.98,0.98)}, anchor=north east},
+]
+\addplot+[mark=*, mark size=1.8pt, line width=0.9pt, draw=modelarf, fill=modelarf, opacity=0.82] coordinates {(1,0.514410) (2,0.943241) (3,0.578316)};
+\addlegendentry{ARF}
+\addplot+[mark=*, mark size=1.8pt, line width=0.9pt, draw=modelbayesnet, fill=modelbayesnet, opacity=0.82] coordinates {(1,0.528563) (2,0.943418) (3,0.590767)};
+\addlegendentry{BayesNet}
+\addplot+[mark=*, mark size=1.8pt, line width=0.9pt, draw=modelctgan, fill=modelctgan, opacity=0.82] coordinates {(1,0.518313) (2,0.938868) (3,0.502965)};
+\addlegendentry{CTGAN}
+\addplot+[mark=*, mark size=1.8pt, line width=0.9pt, draw=modelforestdiffusion, fill=modelforestdiffusion, opacity=0.82] coordinates {(1,0.428868) (2,0.928066) (3,0.405873)};
+\addlegendentry{ForestDiffusion}
+\addplot+[mark=*, mark size=1.8pt, line width=0.9pt, draw=modelrealtabformer, fill=modelrealtabformer, opacity=0.82] coordinates {(1,0.642171) (2,0.991771) (3,0.725290)};
+\addlegendentry{RealTabFormer}
+\addplot+[mark=*, mark size=1.8pt, line width=0.9pt, draw=modeltabbyflow, fill=modeltabbyflow, opacity=0.82] coordinates {(1,0.443399) (2,0.938558) (3,0.479184)};
+\addlegendentry{TabbyFlow}
+\addplot+[mark=*, mark size=1.8pt, line width=0.9pt, draw=modeltabddpm, fill=modeltabddpm, opacity=0.82] coordinates {(1,0.435330) (2,0.960794) (3,0.480608)};
+\addlegendentry{TabDDPM}
+\addplot+[mark=*, mark size=1.8pt, line width=0.9pt, draw=modeltabdiff, fill=modeltabdiff, opacity=0.82] coordinates {(1,0.467464) (2,0.966405) (3,0.490968)};
+\addlegendentry{TabDiff}
+\addplot+[mark=*, mark size=1.8pt, line width=0.9pt, draw=modeltabpfgen, fill=modeltabpfgen, opacity=0.82] coordinates {(1,0.500275) (2,0.920881) (3,0.553496)};
+\addlegendentry{TabPFGen}
+\addplot+[mark=*, mark size=1.8pt, line width=0.9pt, draw=modeltabsyn, fill=modeltabsyn, opacity=0.82] coordinates {(1,0.478248) (2,0.946274) (3,0.546588)};
+\addlegendentry{TabSyn}
+\addplot+[mark=*, mark size=1.8pt, line width=0.9pt, draw=modeltvae, fill=modeltvae, opacity=0.82] coordinates {(1,0.486116) (2,0.958343) (3,0.404814)};
+\addlegendentry{TVAE}
+\addplot+[mark=*, mark size=2.6pt, line width=1.8pt, draw=summaryblack, fill=summaryblack] coordinates {(1,0.496910) (2,0.948784) (3,0.524149)};
+\addlegendentry{Panel mean}
+\addplot+[only marks, mark=none, draw=summaryblack, error bars/.cd, y dir=both, y explicit] coordinates { (1,0.496910) +- (0,0.032741) };
+\addplot+[only marks, mark=none, draw=summaryblack, error bars/.cd, y dir=both, y explicit] coordinates { (2,0.948784) +- (0,0.013789) };
+\addplot+[only marks, mark=none, draw=summaryblack, error bars/.cd, y dir=both, y explicit] coordinates { (3,0.524149) +- (0,0.045843) };
+\end{axis}
+\end{tikzpicture}
+\end{document}

evaluation/query_family/conditional/locality_support_diagnostics/runs/20260519_192156_conditional_locality_support/figures/fig_conditional_locality_support_combined.tex

@@ -0,0 +1,66 @@
+\documentclass[tikz,border=4pt]{standalone}
+\usepackage{pgfplots}
+\usepgfplotslibrary{groupplots}
+\usepackage{xcolor}
+\pgfplotsset{compat=1.18}
+
+\definecolor{modelarf}{HTML}{777777}
+\definecolor{modelbayesnet}{HTML}{CCBB44}
+\definecolor{modelctgan}{HTML}{EE6677}
+\definecolor{modelforestdiffusion}{HTML}{228833}
+\definecolor{modelrealtabformer}{HTML}{332288}
+\definecolor{modeltabbyflow}{HTML}{882255}
+\definecolor{modeltabddpm}{HTML}{EE7733}
+\definecolor{modeltabdiff}{HTML}{AA3377}
+\definecolor{modeltabpfgen}{HTML}{009988}
+\definecolor{modeltabsyn}{HTML}{66CCEE}
+\definecolor{modeltvae}{HTML}{4477AA}
+\definecolor{summaryblack}{HTML}{000000}
+\begin{document}
+\begin{tikzpicture}
+\begin{groupplot}[
+group style={group size=2 by 1, horizontal sep=1.3cm},
+width=6.6cm,
+height=7.6cm,
+ymin=0.0, ymax=1.0,
+ymajorgrids,
+grid style={draw=gray!20},
+major grid style={draw=gray!28},
+axis line style={draw=black!70},
+tick style={draw=black!70},
+]
+\nextgroupplot[title={Panel A. Locality decomposition}, ylabel={Conditional fidelity score}, xtick={1,2,3}, xticklabels={Grouped / Global,2D Surface,Filtered / Local}]
+\addplot+[mark=*, mark size=1.6pt, line width=0.85pt, draw=modelarf, fill=modelarf, opacity=0.78] coordinates {(1,0.514410) (2,0.943241) (3,0.578316)};
+\addplot+[mark=*, mark size=1.6pt, line width=0.85pt, draw=modelbayesnet, fill=modelbayesnet, opacity=0.78] coordinates {(1,0.528563) (2,0.943418) (3,0.590767)};
+\addplot+[mark=*, mark size=1.6pt, line width=0.85pt, draw=modelctgan, fill=modelctgan, opacity=0.78] coordinates {(1,0.518313) (2,0.938868) (3,0.502965)};
+\addplot+[mark=*, mark size=1.6pt, line width=0.85pt, draw=modelforestdiffusion, fill=modelforestdiffusion, opacity=0.78] coordinates {(1,0.428868) (2,0.928066) (3,0.405873)};
+\addplot+[mark=*, mark size=1.6pt, line width=0.85pt, draw=modelrealtabformer, fill=modelrealtabformer, opacity=0.78] coordinates {(1,0.642171) (2,0.991771) (3,0.725290)};
+\addplot+[mark=*, mark size=1.6pt, line width=0.85pt, draw=modeltabbyflow, fill=modeltabbyflow, opacity=0.78] coordinates {(1,0.443399) (2,0.938558) (3,0.479184)};
+\addplot+[mark=*, mark size=1.6pt, line width=0.85pt, draw=modeltabddpm, fill=modeltabddpm, opacity=0.78] coordinates {(1,0.435330) (2,0.960794) (3,0.480608)};
+\addplot+[mark=*, mark size=1.6pt, line width=0.85pt, draw=modeltabdiff, fill=modeltabdiff, opacity=0.78] coordinates {(1,0.467464) (2,0.966405) (3,0.490968)};
+\addplot+[mark=*, mark size=1.6pt, line width=0.85pt, draw=modeltabpfgen, fill=modeltabpfgen, opacity=0.78] coordinates {(1,0.500275) (2,0.920881) (3,0.553496)};
+\addplot+[mark=*, mark size=1.6pt, line width=0.85pt, draw=modeltabsyn, fill=modeltabsyn, opacity=0.78] coordinates {(1,0.478248) (2,0.946274) (3,0.546588)};
+\addplot+[mark=*, mark size=1.6pt, line width=0.85pt, draw=modeltvae, fill=modeltvae, opacity=0.78] coordinates {(1,0.486116) (2,0.958343) (3,0.404814)};
+\addplot+[mark=*, mark size=2.5pt, line width=1.8pt, draw=summaryblack, fill=summaryblack] coordinates {(1,0.496910) (2,0.948784) (3,0.524149)};
+\addplot+[only marks, mark=none, draw=summaryblack, error bars/.cd, y dir=both, y explicit] coordinates { (1,0.496910) +- (0,0.032741) };
+\addplot+[only marks, mark=none, draw=summaryblack, error bars/.cd, y dir=both, y explicit] coordinates { (2,0.948784) +- (0,0.013789) };
+\addplot+[only marks, mark=none, draw=summaryblack, error bars/.cd, y dir=both, y explicit] coordinates { (3,0.524149) +- (0,0.045843) };
+\nextgroupplot[title={Panel B. Support decomposition}, xtick={1,2,3}, xticklabels={Dense,Medium,Sparse}]
+\addplot+[mark=*, mark size=1.6pt, line width=0.85pt, draw=modelarf, fill=modelarf, opacity=0.78] coordinates {(1,0.662726) (2,0.583871) (3,0.583843)};
+\addplot+[mark=*, mark size=1.6pt, line width=0.85pt, draw=modelbayesnet, fill=modelbayesnet, opacity=0.78] coordinates {(1,0.696954) (2,0.613269) (3,0.553258)};
+\addplot+[mark=*, mark size=1.6pt, line width=0.85pt, draw=modelctgan, fill=modelctgan, opacity=0.78] coordinates {(1,0.614868) (2,0.480298) (3,0.455327)};
+\addplot+[mark=*, mark size=1.6pt, line width=0.85pt, draw=modelforestdiffusion, fill=modelforestdiffusion, opacity=0.78] coordinates {(1,0.445168) (2,0.482988) (3,0.362554)};
+\addplot+[mark=*, mark size=1.6pt, line width=0.85pt, draw=modelrealtabformer, fill=modelrealtabformer, opacity=0.78] coordinates {(1,0.846226) (2,0.759283) (3,0.708563)};
+\addplot+[mark=*, mark size=1.6pt, line width=0.85pt, draw=modeltabbyflow, fill=modeltabbyflow, opacity=0.78] coordinates {(1,0.550897) (2,0.504824) (3,0.433465)};
+\addplot+[mark=*, mark size=1.6pt, line width=0.85pt, draw=modeltabddpm, fill=modeltabddpm, opacity=0.78] coordinates {(1,0.577591) (2,0.523680) (3,0.435008)};
+\addplot+[mark=*, mark size=1.6pt, line width=0.85pt, draw=modeltabdiff, fill=modeltabdiff, opacity=0.78] coordinates {(1,0.571185) (2,0.527559) (3,0.437426)};
+\addplot+[mark=*, mark size=1.6pt, line width=0.85pt, draw=modeltabpfgen, fill=modeltabpfgen, opacity=0.78] coordinates {(1,0.645873) (2,0.605467) (3,0.505942)};
+\addplot+[mark=*, mark size=1.6pt, line width=0.85pt, draw=modeltabsyn, fill=modeltabsyn, opacity=0.78] coordinates {(1,0.617580) (2,0.572811) (3,0.472952)};
+\addplot+[mark=*, mark size=1.6pt, line width=0.85pt, draw=modeltvae, fill=modeltvae, opacity=0.78] coordinates {(1,0.536632) (2,0.382542) (3,0.311541)};
+\addplot+[mark=*, mark size=2.5pt, line width=1.8pt, draw=summaryblack, fill=summaryblack] coordinates {(1,0.616144) (2,0.548679) (3,0.479482)};
+\addplot+[only marks, mark=none, draw=summaryblack, error bars/.cd, y dir=both, y explicit] coordinates { (1,0.616144) +- (0,0.048622) };
+\addplot+[only marks, mark=none, draw=summaryblack, error bars/.cd, y dir=both, y explicit] coordinates { (2,0.548679) +- (0,0.051566) };
+\addplot+[only marks, mark=none, draw=summaryblack, error bars/.cd, y dir=both, y explicit] coordinates { (3,0.479482) +- (0,0.048690) };
+\end{groupplot}
+\end{tikzpicture}
+\end{document}

evaluation/query_family/conditional/locality_support_diagnostics/runs/20260519_192156_conditional_locality_support/figures/fig_conditional_support_by_model.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:009bbe015304ca3a2740b22df65ff45b98735a5acc0a667ad13b8d51d55deb2b
+size 20940

evaluation/query_family/conditional/locality_support_diagnostics/runs/20260519_192156_conditional_locality_support/figures/fig_conditional_support_main.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:81a005b719f388f1bc607949955fd17b910726bfb90b30430590765e55a8c20a
+size 19867

evaluation/query_family/conditional/locality_support_diagnostics/runs/20260519_192156_conditional_locality_support/manifest.json

@@ -0,0 +1,261 @@
+{
+  "task": "conditional_locality_support_diagnostic",
+  "generated_at_utc": "2026-05-19T19:22:10.062635+00:00",
+  "source_analysis_run": "trainonly_v2_current_success_official_20way_official20_20260519_232817",
+  "source_conditional_root": "/data/jialinzhang/TabQueryBench/code_snapshot/Evaluation/query_fivepart_breakdown/conditional_breakdown",
+  "run_tag": "20260519_192156_conditional_locality_support",
+  "run_dir": "/data/jialinzhang/TabQueryBench/code_snapshot/Evaluation/query_fivepart_breakdown/conditional_breakdown/locality_support_diagnostics/runs/20260519_192156_conditional_locality_support",
+  "primary_support_variant": "all_filtered_local",
+  "primary_support_reason": "primary_scalar_variant_missing",
+  "coverage": {
+    "conditional_query_rows": 17606,
+    "locality_dataset_model_panels": 404,
+    "filtered_local_query_rows": 2317,
+    "support_unique_cases": 223,
+    "support_primary_panel_rows": 744
+  },
+  "support_method_summary": {
+    "case_count": 223,
+    "template_count": 1,
+    "mode_counts": {
+      "exact": 214,
+      "unavailable": 9
+    },
+    "sql_artifact_found_count": 223,
+    "sql_artifact_missing_count": 0,
+    "main_eligible_case_count": 0
+  },
+  "support_variant_summary": {
+    "scalar_filtered_local": {
+      "eligible_case_count": 0,
+      "supported_dataset_count": 0,
+      "unsupported_dataset_count": 0,
+      "dataset_notes": []
+    },
+    "all_filtered_local": {
+      "eligible_case_count": 214,
+      "supported_dataset_count": 24,
+      "unsupported_dataset_count": 2,
+      "dataset_notes": [
+        {
+          "analysis_variant": "all_filtered_local",
+          "dataset_id": "c4",
+          "case_count": 9,
+          "unique_support_values": 9,
+          "bucketing_status": "ok"
+        },
+        {
+          "analysis_variant": "all_filtered_local",
+          "dataset_id": "m1",
+          "case_count": 9,
+          "unique_support_values": 9,
+          "bucketing_status": "ok"
+        },
+        {
+          "analysis_variant": "all_filtered_local",
+          "dataset_id": "c5",
+          "case_count": 2,
+          "unique_support_values": 2,
+          "bucketing_status": "unsupported_degenerate_within_dataset"
+        },
+        {
+          "analysis_variant": "all_filtered_local",
+          "dataset_id": "m4",
+          "case_count": 8,
+          "unique_support_values": 8,
+          "bucketing_status": "ok"
+        },
+        {
+          "analysis_variant": "all_filtered_local",
+          "dataset_id": "m10",
+          "case_count": 9,
+          "unique_support_values": 9,
+          "bucketing_status": "ok"
+        },
+        {
+          "analysis_variant": "all_filtered_local",
+          "dataset_id": "n8",
+          "case_count": 4,
+          "unique_support_values": 4,
+          "bucketing_status": "ok"
+        },
+        {
+          "analysis_variant": "all_filtered_local",
+          "dataset_id": "c7",
+          "case_count": 9,
+          "unique_support_values": 9,
+          "bucketing_status": "ok"
+        },
+        {
+          "analysis_variant": "all_filtered_local",
+          "dataset_id": "c8",
+          "case_count": 9,
+          "unique_support_values": 9,
+          "bucketing_status": "ok"
+        },
+        {
+          "analysis_variant": "all_filtered_local",
+          "dataset_id": "n15",
+          "case_count": 9,
+          "unique_support_values": 7,
+          "bucketing_status": "ok"
+        },
+        {
+          "analysis_variant": "all_filtered_local",
+          "dataset_id": "m7",
+          "case_count": 9,
+          "unique_support_values": 9,
+          "bucketing_status": "ok"
+        },
+        {
+          "analysis_variant": "all_filtered_local",
+          "dataset_id": "c16",
+          "case_count": 9,
+          "unique_support_values": 7,
+          "bucketing_status": "ok"
+        },
+        {
+          "analysis_variant": "all_filtered_local",
+          "dataset_id": "c17",
+          "case_count": 8,
+          "unique_support_values": 5,
+          "bucketing_status": "ok"
+        },
+        {
+          "analysis_variant": "all_filtered_local",
+          "dataset_id": "m6",
+          "case_count": 9,
+          "unique_support_values": 9,
+          "bucketing_status": "ok"
+        },
+        {
+          "analysis_variant": "all_filtered_local",
+          "dataset_id": "n17",
+          "case_count": 9,
+          "unique_support_values": 7,
+          "bucketing_status": "ok"
+        },
+        {
+          "analysis_variant": "all_filtered_local",
+          "dataset_id": "m9",
+          "case_count": 9,
+          "unique_support_values": 9,
+          "bucketing_status": "ok"
+        },
+        {
+          "analysis_variant": "all_filtered_local",
+          "dataset_id": "c11",
+          "case_count": 9,
+          "unique_support_values": 9,
+          "bucketing_status": "ok"
+        },
+        {
+          "analysis_variant": "all_filtered_local",
+          "dataset_id": "n5",
+          "case_count": 9,
+          "unique_support_values": 7,
+          "bucketing_status": "ok"
+        },
+        {
+          "analysis_variant": "all_filtered_local",
+          "dataset_id": "c9",
+          "case_count": 9,
+          "unique_support_values": 6,
+          "bucketing_status": "ok"
+        },
+        {
+          "analysis_variant": "all_filtered_local",
+          "dataset_id": "c19",
+          "case_count": 8,
+          "unique_support_values": 7,
+          "bucketing_status": "ok"
+        },
+        {
+          "analysis_variant": "all_filtered_local",
+          "dataset_id": "c18",
+          "case_count": 8,
+          "unique_support_values": 8,
+          "bucketing_status": "ok"
+        },
+        {
+          "analysis_variant": "all_filtered_local",
+          "dataset_id": "m12",
+          "case_count": 9,
+          "unique_support_values": 7,
+          "bucketing_status": "ok"
+        },
+        {
+          "analysis_variant": "all_filtered_local",
+          "dataset_id": "n19",
+          "case_count": 6,
+          "unique_support_values": 2,
+          "bucketing_status": "unsupported_degenerate_within_dataset"
+        },
+        {
+          "analysis_variant": "all_filtered_local",
+          "dataset_id": "c14",
+          "case_count": 9,
+          "unique_support_values": 9,
+          "bucketing_status": "ok"
+        },
+        {
+          "analysis_variant": "all_filtered_local",
+          "dataset_id": "n12",
+          "case_count": 8,
+          "unique_support_values": 4,
+          "bucketing_status": "ok"
+        },
+        {
+          "analysis_variant": "all_filtered_local",
+          "dataset_id": "c15",
+          "case_count": 9,
+          "unique_support_values": 9,
+          "bucketing_status": "ok"
+        },
+        {
+          "analysis_variant": "all_filtered_local",
+          "dataset_id": "c10",
+          "case_count": 9,
+          "unique_support_values": 9,
+          "bucketing_status": "ok"
+        }
+      ]
+    }
+  },
+  "compile_notes": {
+    "fig_conditional_locality_main": {
+      "success": false,
+      "note": "latexmk not available"
+    },
+    "fig_conditional_locality_by_model": {
+      "success": false,
+      "note": "latexmk not available"
+    },
+    "fig_conditional_support_main": {
+      "success": false,
+      "note": "latexmk not available"
+    },
+    "fig_conditional_support_by_model": {
+      "success": false,
+      "note": "latexmk not available"
+    },
+    "fig_conditional_locality_support_combined": {
+      "success": false,
+      "note": "latexmk not available"
+    },
+    "table_conditional_locality_summary": {
+      "success": false,
+      "note": "latexmk not available"
+    },
+    "table_conditional_support_summary": {
+      "success": false,
+      "note": "latexmk not available"
+    }
+  },
+  "key_findings": {
+    "locality_global": "Across panel means, conditional fidelity declines from grouped/global summaries (0.497) to 2D surfaces (0.949) and then to filtered/local slices (0.524).",
+    "locality_model": "The strongest grouped/global to filtered/local drop appears for TVAE, falling from 0.486 to 0.405.",
+    "support_global": "Within the exact-support filtered-local subset, dense slices score 0.616, medium slices 0.549, and sparse slices 0.479, consistent with a sparse-support penalty.",
+    "support_model": "Model behavior is mixed: 11 models have positive dense-minus-sparse gaps and 0 show the reverse; the largest positive gap is TVAE at 0.225."
+  }
+}

evaluation/query_family/conditional/locality_support_diagnostics/runs/20260519_192156_conditional_locality_support/report/conditional_locality_diagnostic.md

@@ -0,0 +1,30 @@
+# Conditional locality diagnostic
+
+## Classification audit
+
+Template-level semantics, not raw SQL column counts, define the primary locality buckets. The explicit mapping below keeps the two-axis filtered template in `filtered_local` while preserving `axis_arity = 2D` as a secondary annotation.
+
+| template_id                       | template_name                        | structure_type   | axis_arity   |   n_query_rows |   n_datasets |   n_models |
+|:----------------------------------|:-------------------------------------|:-----------------|:-------------|---------------:|-------------:|-----------:|
+| tpl_m4_group_condition_rate       | Grouped Condition Rate               | grouped_global   | 1D           |           3777 |           32 |         11 |
+| tpl_m4_group_ratio_two_conditions | Grouped Ratio of Two Conditions      | grouped_global   | 1D           |           2745 |           35 |         11 |
+| tpl_m4_window_partition_avg       | Window Partition Average             | grouped_global   | 1D           |           1808 |           21 |         11 |
+| tpl_tpcds_within_group_share      | Within-Group Share of Total          | grouped_global   | 1D           |           6343 |           33 |         11 |
+| tpl_c2_two_dim_target_rate        | Two-Axis Target Rate Surface         | surface_2d       | 2D           |            616 |            3 |         11 |
+| tpl_c2_filtered_group_count_2d    | Filtered Two-Dimensional Group Count | filtered_local   | 2D           |           2317 |           27 |         11 |
+
+## Coverage and scores
+
+| structure_type   | bucket_label     |   query_row_count |   dataset_count |   model_count |   panel_count |   template_count |   mean_score |   ci95_radius | coverage_note                        |
+|:-----------------|:-----------------|------------------:|----------------:|--------------:|--------------:|-----------------:|-------------:|--------------:|:-------------------------------------|
+| grouped_global   | Grouped / Global |             14673 |              39 |            11 |           404 |                4 |     0.49691  |      0.032741 | adequate                             |
+| surface_2d       | 2D Surface       |               616 |               3 |            11 |            33 |                1 |     0.948784 |      0.013789 | low_dataset_coverage,single_template |
+| filtered_local   | Filtered / Local |              2317 |              27 |            11 |           280 |                1 |     0.524149 |      0.045843 | single_template                      |
+
+## Diagnostic takeaways
+
+- Panel means decline from `0.497` for grouped/global queries to `0.949` for 2D surfaces and `0.524` for filtered/local slices.
+- The steepest grouped/global to filtered/local decline appears for `TVAE`: `0.486` to `0.405`.
+- `surface_2d` still rests on one template family, so the locality trend should be treated as structured diagnostic evidence rather than a universal law over all possible 2D conditional tasks.
+- The current conditional row export carries heuristic subitem labels. This locality decomposition therefore anchors on template semantics and panel-level aggregation instead of over-interpreting any single heuristic subitem tag.
+

evaluation/query_family/conditional/locality_support_diagnostics/runs/20260519_192156_conditional_locality_support/report/conditional_locality_support_report.md

@@ -0,0 +1,86 @@
+# Conditional locality and support report
+
+## Scope
+
+- Source conditional breakdown: `/data/jialinzhang/TabQueryBench/code_snapshot/Evaluation/query_fivepart_breakdown/conditional_breakdown`
+- Source analysis run: `trainonly_v2_current_success_official_20way_official20_20260519_232817`
+- Primary support variant: `all_filtered_local` (`primary_scalar_variant_missing`)
+- Stable aggregation is panel-first throughout: query rows -> dataset-model-bucket panels -> model/global summaries.
+- This diagnostic does not rerun the benchmark and does not overwrite upstream conditional outputs.
+
+## Main supported findings
+
+- Across panel means, conditional fidelity declines from grouped/global summaries (0.497) to 2D surfaces (0.949) and then to filtered/local slices (0.524).
+- The strongest grouped/global to filtered/local drop appears for TVAE, falling from 0.486 to 0.405.
+- Within the exact-support filtered-local subset, dense slices score 0.616, medium slices 0.549, and sparse slices 0.479, consistent with a sparse-support penalty.
+- Model behavior is mixed: 11 models have positive dense-minus-sparse gaps and 0 show the reverse; the largest positive gap is TVAE at 0.225.
+
+## Locality diagnostic
+
+# Conditional locality diagnostic
+
+## Classification audit
+
+Template-level semantics, not raw SQL column counts, define the primary locality buckets. The explicit mapping below keeps the two-axis filtered template in `filtered_local` while preserving `axis_arity = 2D` as a secondary annotation.
+
+| template_id                       | template_name                        | structure_type   | axis_arity   |   n_query_rows |   n_datasets |   n_models |
+|:----------------------------------|:-------------------------------------|:-----------------|:-------------|---------------:|-------------:|-----------:|
+| tpl_m4_group_condition_rate       | Grouped Condition Rate               | grouped_global   | 1D           |           3777 |           32 |         11 |
+| tpl_m4_group_ratio_two_conditions | Grouped Ratio of Two Conditions      | grouped_global   | 1D           |           2745 |           35 |         11 |
+| tpl_m4_window_partition_avg       | Window Partition Average             | grouped_global   | 1D           |           1808 |           21 |         11 |
+| tpl_tpcds_within_group_share      | Within-Group Share of Total          | grouped_global   | 1D           |           6343 |           33 |         11 |
+| tpl_c2_two_dim_target_rate        | Two-Axis Target Rate Surface         | surface_2d       | 2D           |            616 |            3 |         11 |
+| tpl_c2_filtered_group_count_2d    | Filtered Two-Dimensional Group Count | filtered_local   | 2D           |           2317 |           27 |         11 |
+
+## Coverage and scores
+
+| structure_type   | bucket_label     |   query_row_count |   dataset_count |   model_count |   panel_count |   template_count |   mean_score |   ci95_radius | coverage_note                        |
+|:-----------------|:-----------------|------------------:|----------------:|--------------:|--------------:|-----------------:|-------------:|--------------:|:-------------------------------------|
+| grouped_global   | Grouped / Global |             14673 |              39 |            11 |           404 |                4 |     0.49691  |      0.032741 | adequate                             |
+| surface_2d       | 2D Surface       |               616 |               3 |            11 |            33 |                1 |     0.948784 |      0.013789 | low_dataset_coverage,single_template |
+| filtered_local   | Filtered / Local |              2317 |              27 |            11 |           280 |                1 |     0.524149 |      0.045843 | single_template                      |
+
+## Diagnostic takeaways
+
+- Panel means decline from `0.497` for grouped/global queries to `0.949` for 2D surfaces and `0.524` for filtered/local slices.
+- The steepest grouped/global to filtered/local decline appears for `TVAE`: `0.486` to `0.405`.
+- `surface_2d` still rests on one template family, so the locality trend should be treated as structured diagnostic evidence rather than a universal law over all possible 2D conditional tasks.
+- The current conditional row export carries heuristic subitem labels. This locality decomposition therefore anchors on template semantics and panel-level aggregation instead of over-interpreting any single heuristic subitem tag.
+
+
+## Support diagnostic
+
+# Conditional support diagnostic
+
+## Feasibility and recovery modes
+
+- Source SQL artifact coverage found: `223` recovered cases; missing: `0`.
+- Support recovery modes on unique filtered-local cases: `{"exact": 214, "unavailable": 9}`.
+- Primary dense/medium/sparse variant: `all_filtered_local` (`primary_scalar_variant_missing`).
+- Exact vs proxy row-level coverage in the audit export: exact=`2227`, derived_exact=`0`, proxy=`0`, unavailable=`90`.
+
+## Coverage and scores
+
+| support_bucket   | bucket_label   |   query_row_count |   dataset_count |   model_count |   panel_count |   template_count |   mean_score |   ci95_radius | coverage_note   |
+|:-----------------|:---------------|------------------:|----------------:|--------------:|--------------:|-----------------:|-------------:|--------------:|:----------------|
+| dense            | Dense          |               726 |              24 |            11 |           248 |                1 |     0.616144 |      0.048622 | single_template |
+| medium           | Medium         |               678 |              24 |            11 |           248 |                1 |     0.548679 |      0.051566 | single_template |
+| sparse           | Sparse         |               735 |              24 |            11 |           248 |                1 |     0.479482 |      0.04869  | single_template |
+
+## Diagnostic takeaways
+
+- The global panel mean declines from `0.616` on dense filtered-local slices to `0.479` on sparse slices.
+- Model behavior is mixed: `11` models have positive dense-minus-sparse gaps, `0` show the reverse, and `0` are flat. The largest positive gap appears for `TVAE` at `0.225`.
+- In the broader `all_filtered_local` sensitivity view (`24` datasets), dense=`0.616` and sparse=`0.479`; the sparse-support penalty is clearer once the filtered 2D local template is included.
+- The primary support analysis intentionally keeps only scalar filtered-local templates in the main dense/medium/sparse comparison so that the support unit remains the count of real rows satisfying the local predicate.
+- Exact per-cell support is still recovered and audited for the filtered 2D group-count template, but that template is left as a sensitivity-only support basis because its natural support statistic is a cell-count distribution rather than a scalar slice size.
+- On this main scalar subset, sparse support does not by itself explain the filtered-local weakness. Any support-mediated interpretation should therefore be limited to model-specific behavior or to the broader sensitivity analysis, not promoted as a universal driver.
+- Any unsupported or unavailable support cases remain explicit in the audit CSV and are not silently folded into the main claim.
+
+
+## Caveats
+
+- `surface_2d` is still represented by one template family, so the locality trend should be described as a template-grounded diagnostic pattern rather than a universal statement about dimensionality alone.
+- The support main figure intentionally excludes the filtered 2D count template from the primary dense/medium/sparse claim because its most faithful support signal is a distribution of cell counts, not a single filtered-row count.
+- Existing heuristic subitem labels in the conditional row export do not perfectly align with template-level semantics, so this diagnostic relies on template semantics for bucket assignment and uses query-score panel means as the primary outcome.
+

evaluation/query_family/conditional/locality_support_diagnostics/runs/20260519_192156_conditional_locality_support/report/conditional_support_bucket_diagnostic.md

@@ -0,0 +1,27 @@
+# Conditional support diagnostic
+
+## Feasibility and recovery modes
+
+- Source SQL artifact coverage found: `223` recovered cases; missing: `0`.
+- Support recovery modes on unique filtered-local cases: `{"exact": 214, "unavailable": 9}`.
+- Primary dense/medium/sparse variant: `all_filtered_local` (`primary_scalar_variant_missing`).
+- Exact vs proxy row-level coverage in the audit export: exact=`2227`, derived_exact=`0`, proxy=`0`, unavailable=`90`.
+
+## Coverage and scores
+
+| support_bucket   | bucket_label   |   query_row_count |   dataset_count |   model_count |   panel_count |   template_count |   mean_score |   ci95_radius | coverage_note   |
+|:-----------------|:---------------|------------------:|----------------:|--------------:|--------------:|-----------------:|-------------:|--------------:|:----------------|
+| dense            | Dense          |               726 |              24 |            11 |           248 |                1 |     0.616144 |      0.048622 | single_template |
+| medium           | Medium         |               678 |              24 |            11 |           248 |                1 |     0.548679 |      0.051566 | single_template |
+| sparse           | Sparse         |               735 |              24 |            11 |           248 |                1 |     0.479482 |      0.04869  | single_template |
+
+## Diagnostic takeaways
+
+- The global panel mean declines from `0.616` on dense filtered-local slices to `0.479` on sparse slices.
+- Model behavior is mixed: `11` models have positive dense-minus-sparse gaps, `0` show the reverse, and `0` are flat. The largest positive gap appears for `TVAE` at `0.225`.
+- In the broader `all_filtered_local` sensitivity view (`24` datasets), dense=`0.616` and sparse=`0.479`; the sparse-support penalty is clearer once the filtered 2D local template is included.
+- The primary support analysis intentionally keeps only scalar filtered-local templates in the main dense/medium/sparse comparison so that the support unit remains the count of real rows satisfying the local predicate.
+- Exact per-cell support is still recovered and audited for the filtered 2D group-count template, but that template is left as a sensitivity-only support basis because its natural support statistic is a cell-count distribution rather than a scalar slice size.
+- On this main scalar subset, sparse support does not by itself explain the filtered-local weakness. Any support-mediated interpretation should therefore be limited to model-specific behavior or to the broader sensitivity analysis, not promoted as a universal driver.
+- Any unsupported or unavailable support cases remain explicit in the audit CSV and are not silently folded into the main claim.
+

evaluation/query_family/conditional/locality_support_diagnostics/runs/20260519_192156_conditional_locality_support/report/paper_caption.txt

@@ -0,0 +1,8 @@
+Figure 1. Conditional locality decomposition.
+Across panel means, conditional fidelity declines from grouped/global summaries (0.497) to 2D surfaces (0.949) and then to filtered/local slices (0.524). Points and error bars show panel means with 95% confidence intervals; colored traces show per-model means under the frozen model roster and color convention.
+
+Figure 2. Conditional support decomposition.
+Within the exact-support filtered-local subset, dense slices score 0.616, medium slices 0.549, and sparse slices 0.479, consistent with a sparse-support penalty. The main support figure uses the all filtered-local templates subset so that support is measured on a comparable exact real-row-count scale within each dataset.
+
+Figure 3. Combined conditional locality/support diagnostic.
+Panel A shows the locality decomposition from grouped/global summaries to filtered/local slices. Panel B shows the dense/medium/sparse comparison inside the filtered-local subset. Both panels use panel-level aggregation and expose coverage caveats in the companion audit tables rather than hiding thin buckets.

evaluation/query_family/conditional/locality_support_diagnostics/runs/20260519_192156_conditional_locality_support/report/paper_paragraphs.md

@@ -0,0 +1,5 @@
+Across panel means, conditional fidelity declines from grouped/global summaries (0.497) to 2D surfaces (0.949) and then to filtered/local slices (0.524). This suggests that axis count alone is not the most interpretable explanation for the conditional-family weakness: grouped/global summaries remain comparatively more stable, while narrow filtered slices are harder to preserve.
+
+Within the exact-support filtered-local subset, dense slices score 0.616, medium slices 0.549, and sparse slices 0.479, consistent with a sparse-support penalty. That pattern indicates that sparse real support explains part of the local-slice collapse, consistent with synthetic generators smoothing away rare conditional interactions.
+
+Model behavior is mixed: 11 models have positive dense-minus-sparse gaps and 0 show the reverse; the largest positive gap is TVAE at 0.225. At the same time, the support diagnostic does not fully explain the conditional gap on its own: even dense local slices can remain weak for some models, and the 2D-surface bucket still rests on limited template coverage.

evaluation/query_family/conditional/locality_support_diagnostics/runs/20260519_192156_conditional_locality_support/tables/table_conditional_locality_summary.tex

@@ -0,0 +1,16 @@
+\documentclass{standalone}
+\usepackage[table]{xcolor}
+\usepackage{booktabs}
+\begin{document}
+\scriptsize
+\emph{Panel-level locality summary.}\\[0.4em]
+\begin{tabular}{llllll}
+\toprule
+Bucket & Panels & Datasets & Templates & Mean & 95\% CI \\
+\midrule
+Grouped / Global & 404 & 39 & 4 & 0.497 & 0.033 \\
+2D Surface & 33 & 3 & 1 & 0.949 & 0.014 \\
+Filtered / Local & 280 & 27 & 1 & 0.524 & 0.046 \\
+\bottomrule
+\end{tabular}
+\end{document}

evaluation/query_family/conditional/locality_support_diagnostics/runs/20260519_192156_conditional_locality_support/tables/table_conditional_support_summary.tex

@@ -0,0 +1,16 @@
+\documentclass{standalone}
+\usepackage[table]{xcolor}
+\usepackage{booktabs}
+\begin{document}
+\scriptsize
+\emph{Panel-level support summary.}\\[0.4em]
+\begin{tabular}{llllll}
+\toprule
+Bucket & Panels & Datasets & Templates & Mean & 95\% CI \\
+\midrule
+Dense & 248 & 24 & 1 & 0.616 & 0.049 \\
+Medium & 248 & 24 & 1 & 0.549 & 0.052 \\
+Sparse & 248 & 24 & 1 & 0.479 & 0.049 \\
+\bottomrule
+\end{tabular}
+\end{document}

evaluation/query_family/conditional/locality_support_diagnostics/runs/20260519_192327_conditional_locality_support/README.md

@@ -0,0 +1,8 @@
+# 20260519_192327_conditional_locality_support
+
+This run contains the full reproducible bundle for the conditional locality/support diagnostic.
+
+- `data/` exports the summary and audit CSVs.
+- `figures/` holds the paper-facing figures plus standalone TeX sources.
+- `tables/` holds LaTeX table snippets.
+- `report/` holds the Markdown narrative, captions, and paper paragraphs.

evaluation/query_family/conditional/locality_support_diagnostics/runs/20260519_192327_conditional_locality_support/data/conditional_locality_summary.csv

@@ -0,0 +1,4 @@
+structure_type,bucket_label,panel_count,dataset_count,model_count,template_count,query_row_count,prefix_coverage,prefix_count,mean_score,std_score,se_score,ci95_low,ci95_high,ci95_radius,coverage_note
+grouped_global,Grouped / Global,404,39,11,4,14673,"c,m,n",3,0.49691,0.335757,0.016705,0.46417,0.529651,0.032741,adequate
+surface_2d,2D Surface,33,3,11,1,616,c,1,0.948784,0.040414,0.007035,0.934994,0.962573,0.013789,"low_dataset_coverage,single_template"
+filtered_local,Filtered / Local,280,27,11,1,2317,"c,m,n",3,0.524149,0.39138,0.023389,0.478305,0.569992,0.045843,single_template

evaluation/query_family/conditional/locality_support_diagnostics/runs/20260519_192327_conditional_locality_support/data/conditional_support_bucket_summary.csv

@@ -0,0 +1,4 @@
+support_bucket,analysis_variant,bucket_label,panel_count,dataset_count,model_count,template_count,query_row_count,prefix_coverage,prefix_count,mean_score,std_score,se_score,ci95_low,ci95_high,ci95_radius,coverage_note
+dense,all_filtered_local,Dense,248,24,11,1,726,"c,m,n",3,0.616144,0.390666,0.024807,0.567521,0.664766,0.048622,single_template
+medium,all_filtered_local,Medium,248,24,11,1,678,"c,m,n",3,0.548679,0.41432,0.026309,0.497113,0.600245,0.051566,single_template
+sparse,all_filtered_local,Sparse,248,24,11,1,735,"c,m,n",3,0.479482,0.391209,0.024842,0.430792,0.528172,0.04869,single_template

evaluation/query_family/conditional/locality_support_diagnostics/runs/20260519_192327_conditional_locality_support/data/conditional_support_dense_sparse_drop.csv

@@ -0,0 +1,12 @@
+model_id,model_label,dense_score,sparse_score,dense_minus_sparse
+arf,ARF,0.662726,0.583843,0.078883
+bayesnet,BayesNet,0.696954,0.553258,0.143696
+ctgan,CTGAN,0.614868,0.455327,0.159541
+forestdiffusion,ForestDiffusion,0.445168,0.362554,0.082614
+realtabformer,RealTabFormer,0.846226,0.708563,0.137663
+tabbyflow,TabbyFlow,0.550897,0.433465,0.117432
+tabddpm,TabDDPM,0.577591,0.435008,0.142583
+tabdiff,TabDiff,0.571185,0.437426,0.133759
+tabpfgen,TabPFGen,0.645873,0.505942,0.139931
+tabsyn,TabSyn,0.61758,0.472952,0.144628
+tvae,TVAE,0.536632,0.311541,0.225091

evaluation/query_family/conditional/locality_support_diagnostics/runs/20260519_192327_conditional_locality_support/data/conditional_template_mapping.csv

@@ -0,0 +1,7 @@
+template_id,template_name,structure_type,axis_arity,classification_rationale,n_query_rows,n_datasets,n_models,n_panels,dataset_prefixes,observed_subitems,contract_supported_subitems,contract_allowed_roles,support_bucket_basis,support_main_eligible,support_basis_note
+tpl_m4_group_condition_rate,Grouped Condition Rate,grouped_global,1D,One grouped axis with no local filter; the query compares a global condition rate across groups rather than drilling into a filtered slice.,3777,32,11,334,"c,m,n","dependency_strength_similarity,direction_consistency","dependency_strength_similarity,direction_consistency","within_group_proportion,focused_target_view",not_applicable,False,Not part of the filtered-local support diagnostic.
+tpl_m4_group_ratio_two_conditions,Grouped Ratio of Two Conditions,grouped_global,1D,"One grouped axis with a contrastive ratio view, but still a global grouped summary rather than a local filtered slice.",2745,35,11,365,"c,m,n",direction_consistency,direction_consistency,contrastive_conditional_view,not_applicable,False,Not part of the filtered-local support diagnostic.
+tpl_m4_window_partition_avg,Window Partition Average,grouped_global,1D,Window partitions still summarize one grouping axis at the full-dataset level; they do not define a narrow filtered slice.,1808,21,11,216,"c,m,n","direction_consistency,slice_level_consistency","slice_level_consistency,direction_consistency","filtered_stable_view,ranked_signal_view",not_applicable,False,Not part of the filtered-local support diagnostic.
+tpl_tpcds_within_group_share,Within-Group Share of Total,grouped_global,1D,"The item-level shares are nested inside a grouped global summary, not exposed as a two-axis interaction surface.",6343,33,11,339,"c,m,n",dependency_strength_similarity,dependency_strength_similarity,"within_group_proportion,focused_target_view",not_applicable,False,Not part of the filtered-local support diagnostic.
+tpl_c2_two_dim_target_rate,Two-Axis Target Rate Surface,surface_2d,2D,This template explicitly constructs a two-axis interaction surface over two grouping fields without a local filter.,616,3,11,33,c,"dependency_strength_similarity,direction_consistency","dependency_strength_similarity,direction_consistency","within_group_proportion,ranked_signal_view",not_applicable,False,Not part of the filtered-local support diagnostic.
+tpl_c2_filtered_group_count_2d,Filtered Two-Dimensional Group Count,filtered_local,2D,"Even though the output is two-dimensional, the defining semantics are local because the surface exists only inside a predicate-defined slice.",2317,27,11,280,"c,m,n",slice_level_consistency,slice_level_consistency,count_distribution,median_cell_row_count,False,"Exact per-cell counts are available, but the primary support analysis keeps scalar filtered-slice support units comparable and therefore excludes this template from the main dense/medium/sparse claim."

evaluation/query_family/conditional/locality_support_diagnostics/runs/20260519_192327_conditional_locality_support/figures/fig_conditional_locality_by_model.tex

@@ -0,0 +1,65 @@
+\documentclass[tikz,border=4pt]{standalone}
+\usepackage{pgfplots}
+\usepgfplotslibrary{groupplots}
+\usepackage{xcolor}
+\pgfplotsset{compat=1.18}
+
+\definecolor{modelarf}{HTML}{777777}
+\definecolor{modelbayesnet}{HTML}{CCBB44}
+\definecolor{modelctgan}{HTML}{EE6677}
+\definecolor{modelforestdiffusion}{HTML}{228833}
+\definecolor{modelrealtabformer}{HTML}{332288}
+\definecolor{modeltabbyflow}{HTML}{882255}
+\definecolor{modeltabddpm}{HTML}{EE7733}
+\definecolor{modeltabdiff}{HTML}{AA3377}
+\definecolor{modeltabpfgen}{HTML}{009988}
+\definecolor{modeltabsyn}{HTML}{66CCEE}
+\definecolor{modeltvae}{HTML}{4477AA}
+\definecolor{summaryblack}{HTML}{000000}
+\begin{document}
+\begin{tikzpicture}
+\begin{axis}[
+width=13.8cm,
+height=8.4cm,
+ymin=0.0, ymax=1.0,
+title={Conditional locality decomposition by model},
+ylabel={Conditional fidelity score},
+xtick={1,2,3},
+xticklabels={Grouped / Global,2D Surface,Filtered / Local},
+ymajorgrids,
+grid style={draw=gray!20},
+major grid style={draw=gray!28},
+axis line style={draw=black!70},
+tick style={draw=black!70},
+legend style={draw=none, fill=none, font=\scriptsize, at={(0.98,0.98)}, anchor=north east},
+]
+\addplot+[mark=*, mark size=1.8pt, line width=0.9pt, draw=modelarf, fill=modelarf, opacity=0.82] coordinates {(1,0.514410) (2,0.943241) (3,0.578316)};
+\addlegendentry{ARF}
+\addplot+[mark=*, mark size=1.8pt, line width=0.9pt, draw=modelbayesnet, fill=modelbayesnet, opacity=0.82] coordinates {(1,0.528563) (2,0.943418) (3,0.590767)};
+\addlegendentry{BayesNet}
+\addplot+[mark=*, mark size=1.8pt, line width=0.9pt, draw=modelctgan, fill=modelctgan, opacity=0.82] coordinates {(1,0.518313) (2,0.938868) (3,0.502965)};
+\addlegendentry{CTGAN}
+\addplot+[mark=*, mark size=1.8pt, line width=0.9pt, draw=modelforestdiffusion, fill=modelforestdiffusion, opacity=0.82] coordinates {(1,0.428868) (2,0.928066) (3,0.405873)};
+\addlegendentry{ForestDiffusion}
+\addplot+[mark=*, mark size=1.8pt, line width=0.9pt, draw=modelrealtabformer, fill=modelrealtabformer, opacity=0.82] coordinates {(1,0.642171) (2,0.991771) (3,0.725290)};
+\addlegendentry{RealTabFormer}
+\addplot+[mark=*, mark size=1.8pt, line width=0.9pt, draw=modeltabbyflow, fill=modeltabbyflow, opacity=0.82] coordinates {(1,0.443399) (2,0.938558) (3,0.479184)};
+\addlegendentry{TabbyFlow}
+\addplot+[mark=*, mark size=1.8pt, line width=0.9pt, draw=modeltabddpm, fill=modeltabddpm, opacity=0.82] coordinates {(1,0.435330) (2,0.960794) (3,0.480608)};
+\addlegendentry{TabDDPM}
+\addplot+[mark=*, mark size=1.8pt, line width=0.9pt, draw=modeltabdiff, fill=modeltabdiff, opacity=0.82] coordinates {(1,0.467464) (2,0.966405) (3,0.490968)};
+\addlegendentry{TabDiff}
+\addplot+[mark=*, mark size=1.8pt, line width=0.9pt, draw=modeltabpfgen, fill=modeltabpfgen, opacity=0.82] coordinates {(1,0.500275) (2,0.920881) (3,0.553496)};
+\addlegendentry{TabPFGen}
+\addplot+[mark=*, mark size=1.8pt, line width=0.9pt, draw=modeltabsyn, fill=modeltabsyn, opacity=0.82] coordinates {(1,0.478248) (2,0.946274) (3,0.546588)};
+\addlegendentry{TabSyn}
+\addplot+[mark=*, mark size=1.8pt, line width=0.9pt, draw=modeltvae, fill=modeltvae, opacity=0.82] coordinates {(1,0.486116) (2,0.958343) (3,0.404814)};
+\addlegendentry{TVAE}
+\addplot+[mark=*, mark size=2.6pt, line width=1.8pt, draw=summaryblack, fill=summaryblack] coordinates {(1,0.496910) (2,0.948784) (3,0.524149)};
+\addlegendentry{Panel mean}
+\addplot+[only marks, mark=none, draw=summaryblack, error bars/.cd, y dir=both, y explicit] coordinates { (1,0.496910) +- (0,0.032741) };
+\addplot+[only marks, mark=none, draw=summaryblack, error bars/.cd, y dir=both, y explicit] coordinates { (2,0.948784) +- (0,0.013789) };
+\addplot+[only marks, mark=none, draw=summaryblack, error bars/.cd, y dir=both, y explicit] coordinates { (3,0.524149) +- (0,0.045843) };
+\end{axis}
+\end{tikzpicture}
+\end{document}

evaluation/query_family/conditional/locality_support_diagnostics/runs/20260519_192327_conditional_locality_support/figures/fig_conditional_locality_main.tex

@@ -0,0 +1,53 @@
+\documentclass[tikz,border=4pt]{standalone}
+\usepackage{pgfplots}
+\usepgfplotslibrary{groupplots}
+\usepackage{xcolor}
+\pgfplotsset{compat=1.18}
+
+\definecolor{modelarf}{HTML}{777777}
+\definecolor{modelbayesnet}{HTML}{CCBB44}
+\definecolor{modelctgan}{HTML}{EE6677}
+\definecolor{modelforestdiffusion}{HTML}{228833}
+\definecolor{modelrealtabformer}{HTML}{332288}
+\definecolor{modeltabbyflow}{HTML}{882255}
+\definecolor{modeltabddpm}{HTML}{EE7733}
+\definecolor{modeltabdiff}{HTML}{AA3377}
+\definecolor{modeltabpfgen}{HTML}{009988}
+\definecolor{modeltabsyn}{HTML}{66CCEE}
+\definecolor{modeltvae}{HTML}{4477AA}
+\definecolor{summaryblack}{HTML}{000000}
+\begin{document}
+\begin{tikzpicture}
+\begin{axis}[
+width=13.8cm,
+height=8.4cm,
+ymin=0.0, ymax=1.0,
+title={Conditional locality decomposition},
+ylabel={Conditional fidelity score},
+xtick={1,2,3},
+xticklabels={Grouped / Global,2D Surface,Filtered / Local},
+ymajorgrids,
+grid style={draw=gray!20},
+major grid style={draw=gray!28},
+axis line style={draw=black!70},
+tick style={draw=black!70},
+legend style={draw=none, fill=none, font=\scriptsize, at={(0.98,0.98)}, anchor=north east},
+]
+\addplot+[mark=*, mark size=1.8pt, line width=0.9pt, draw=modelarf, fill=modelarf, opacity=0.82] coordinates {(1,0.514410) (2,0.943241) (3,0.578316)};
+\addplot+[mark=*, mark size=1.8pt, line width=0.9pt, draw=modelbayesnet, fill=modelbayesnet, opacity=0.82] coordinates {(1,0.528563) (2,0.943418) (3,0.590767)};
+\addplot+[mark=*, mark size=1.8pt, line width=0.9pt, draw=modelctgan, fill=modelctgan, opacity=0.82] coordinates {(1,0.518313) (2,0.938868) (3,0.502965)};
+\addplot+[mark=*, mark size=1.8pt, line width=0.9pt, draw=modelforestdiffusion, fill=modelforestdiffusion, opacity=0.82] coordinates {(1,0.428868) (2,0.928066) (3,0.405873)};
+\addplot+[mark=*, mark size=1.8pt, line width=0.9pt, draw=modelrealtabformer, fill=modelrealtabformer, opacity=0.82] coordinates {(1,0.642171) (2,0.991771) (3,0.725290)};
+\addplot+[mark=*, mark size=1.8pt, line width=0.9pt, draw=modeltabbyflow, fill=modeltabbyflow, opacity=0.82] coordinates {(1,0.443399) (2,0.938558) (3,0.479184)};
+\addplot+[mark=*, mark size=1.8pt, line width=0.9pt, draw=modeltabddpm, fill=modeltabddpm, opacity=0.82] coordinates {(1,0.435330) (2,0.960794) (3,0.480608)};
+\addplot+[mark=*, mark size=1.8pt, line width=0.9pt, draw=modeltabdiff, fill=modeltabdiff, opacity=0.82] coordinates {(1,0.467464) (2,0.966405) (3,0.490968)};
+\addplot+[mark=*, mark size=1.8pt, line width=0.9pt, draw=modeltabpfgen, fill=modeltabpfgen, opacity=0.82] coordinates {(1,0.500275) (2,0.920881) (3,0.553496)};
+\addplot+[mark=*, mark size=1.8pt, line width=0.9pt, draw=modeltabsyn, fill=modeltabsyn, opacity=0.82] coordinates {(1,0.478248) (2,0.946274) (3,0.546588)};
+\addplot+[mark=*, mark size=1.8pt, line width=0.9pt, draw=modeltvae, fill=modeltvae, opacity=0.82] coordinates {(1,0.486116) (2,0.958343) (3,0.404814)};
+\addplot+[mark=*, mark size=2.6pt, line width=1.8pt, draw=summaryblack, fill=summaryblack] coordinates {(1,0.496910) (2,0.948784) (3,0.524149)};
+\addplot+[only marks, mark=none, draw=summaryblack, error bars/.cd, y dir=both, y explicit] coordinates { (1,0.496910) +- (0,0.032741) };
+\addplot+[only marks, mark=none, draw=summaryblack, error bars/.cd, y dir=both, y explicit] coordinates { (2,0.948784) +- (0,0.013789) };
+\addplot+[only marks, mark=none, draw=summaryblack, error bars/.cd, y dir=both, y explicit] coordinates { (3,0.524149) +- (0,0.045843) };
+\end{axis}
+\end{tikzpicture}
+\end{document}

evaluation/query_family/conditional/locality_support_diagnostics/runs/20260519_192327_conditional_locality_support/figures/fig_conditional_locality_support_combined.tex

@@ -0,0 +1,66 @@
+\documentclass[tikz,border=4pt]{standalone}
+\usepackage{pgfplots}
+\usepgfplotslibrary{groupplots}
+\usepackage{xcolor}
+\pgfplotsset{compat=1.18}
+
+\definecolor{modelarf}{HTML}{777777}
+\definecolor{modelbayesnet}{HTML}{CCBB44}
+\definecolor{modelctgan}{HTML}{EE6677}
+\definecolor{modelforestdiffusion}{HTML}{228833}
+\definecolor{modelrealtabformer}{HTML}{332288}
+\definecolor{modeltabbyflow}{HTML}{882255}
+\definecolor{modeltabddpm}{HTML}{EE7733}
+\definecolor{modeltabdiff}{HTML}{AA3377}
+\definecolor{modeltabpfgen}{HTML}{009988}
+\definecolor{modeltabsyn}{HTML}{66CCEE}
+\definecolor{modeltvae}{HTML}{4477AA}
+\definecolor{summaryblack}{HTML}{000000}
+\begin{document}
+\begin{tikzpicture}
+\begin{groupplot}[
+group style={group size=2 by 1, horizontal sep=1.3cm},
+width=6.6cm,
+height=7.6cm,
+ymin=0.0, ymax=1.0,
+ymajorgrids,
+grid style={draw=gray!20},
+major grid style={draw=gray!28},
+axis line style={draw=black!70},
+tick style={draw=black!70},
+]
+\nextgroupplot[title={Panel A. Locality decomposition}, ylabel={Conditional fidelity score}, xtick={1,2,3}, xticklabels={Grouped / Global,2D Surface,Filtered / Local}]
+\addplot+[mark=*, mark size=1.6pt, line width=0.85pt, draw=modelarf, fill=modelarf, opacity=0.78] coordinates {(1,0.514410) (2,0.943241) (3,0.578316)};
+\addplot+[mark=*, mark size=1.6pt, line width=0.85pt, draw=modelbayesnet, fill=modelbayesnet, opacity=0.78] coordinates {(1,0.528563) (2,0.943418) (3,0.590767)};
+\addplot+[mark=*, mark size=1.6pt, line width=0.85pt, draw=modelctgan, fill=modelctgan, opacity=0.78] coordinates {(1,0.518313) (2,0.938868) (3,0.502965)};
+\addplot+[mark=*, mark size=1.6pt, line width=0.85pt, draw=modelforestdiffusion, fill=modelforestdiffusion, opacity=0.78] coordinates {(1,0.428868) (2,0.928066) (3,0.405873)};
+\addplot+[mark=*, mark size=1.6pt, line width=0.85pt, draw=modelrealtabformer, fill=modelrealtabformer, opacity=0.78] coordinates {(1,0.642171) (2,0.991771) (3,0.725290)};
+\addplot+[mark=*, mark size=1.6pt, line width=0.85pt, draw=modeltabbyflow, fill=modeltabbyflow, opacity=0.78] coordinates {(1,0.443399) (2,0.938558) (3,0.479184)};
+\addplot+[mark=*, mark size=1.6pt, line width=0.85pt, draw=modeltabddpm, fill=modeltabddpm, opacity=0.78] coordinates {(1,0.435330) (2,0.960794) (3,0.480608)};
+\addplot+[mark=*, mark size=1.6pt, line width=0.85pt, draw=modeltabdiff, fill=modeltabdiff, opacity=0.78] coordinates {(1,0.467464) (2,0.966405) (3,0.490968)};
+\addplot+[mark=*, mark size=1.6pt, line width=0.85pt, draw=modeltabpfgen, fill=modeltabpfgen, opacity=0.78] coordinates {(1,0.500275) (2,0.920881) (3,0.553496)};
+\addplot+[mark=*, mark size=1.6pt, line width=0.85pt, draw=modeltabsyn, fill=modeltabsyn, opacity=0.78] coordinates {(1,0.478248) (2,0.946274) (3,0.546588)};
+\addplot+[mark=*, mark size=1.6pt, line width=0.85pt, draw=modeltvae, fill=modeltvae, opacity=0.78] coordinates {(1,0.486116) (2,0.958343) (3,0.404814)};
+\addplot+[mark=*, mark size=2.5pt, line width=1.8pt, draw=summaryblack, fill=summaryblack] coordinates {(1,0.496910) (2,0.948784) (3,0.524149)};
+\addplot+[only marks, mark=none, draw=summaryblack, error bars/.cd, y dir=both, y explicit] coordinates { (1,0.496910) +- (0,0.032741) };
+\addplot+[only marks, mark=none, draw=summaryblack, error bars/.cd, y dir=both, y explicit] coordinates { (2,0.948784) +- (0,0.013789) };
+\addplot+[only marks, mark=none, draw=summaryblack, error bars/.cd, y dir=both, y explicit] coordinates { (3,0.524149) +- (0,0.045843) };
+\nextgroupplot[title={Panel B. Support decomposition}, xtick={1,2,3}, xticklabels={Dense,Medium,Sparse}]
+\addplot+[mark=*, mark size=1.6pt, line width=0.85pt, draw=modelarf, fill=modelarf, opacity=0.78] coordinates {(1,0.662726) (2,0.583871) (3,0.583843)};
+\addplot+[mark=*, mark size=1.6pt, line width=0.85pt, draw=modelbayesnet, fill=modelbayesnet, opacity=0.78] coordinates {(1,0.696954) (2,0.613269) (3,0.553258)};
+\addplot+[mark=*, mark size=1.6pt, line width=0.85pt, draw=modelctgan, fill=modelctgan, opacity=0.78] coordinates {(1,0.614868) (2,0.480298) (3,0.455327)};
+\addplot+[mark=*, mark size=1.6pt, line width=0.85pt, draw=modelforestdiffusion, fill=modelforestdiffusion, opacity=0.78] coordinates {(1,0.445168) (2,0.482988) (3,0.362554)};
+\addplot+[mark=*, mark size=1.6pt, line width=0.85pt, draw=modelrealtabformer, fill=modelrealtabformer, opacity=0.78] coordinates {(1,0.846226) (2,0.759283) (3,0.708563)};
+\addplot+[mark=*, mark size=1.6pt, line width=0.85pt, draw=modeltabbyflow, fill=modeltabbyflow, opacity=0.78] coordinates {(1,0.550897) (2,0.504824) (3,0.433465)};
+\addplot+[mark=*, mark size=1.6pt, line width=0.85pt, draw=modeltabddpm, fill=modeltabddpm, opacity=0.78] coordinates {(1,0.577591) (2,0.523680) (3,0.435008)};
+\addplot+[mark=*, mark size=1.6pt, line width=0.85pt, draw=modeltabdiff, fill=modeltabdiff, opacity=0.78] coordinates {(1,0.571185) (2,0.527559) (3,0.437426)};
+\addplot+[mark=*, mark size=1.6pt, line width=0.85pt, draw=modeltabpfgen, fill=modeltabpfgen, opacity=0.78] coordinates {(1,0.645873) (2,0.605467) (3,0.505942)};
+\addplot+[mark=*, mark size=1.6pt, line width=0.85pt, draw=modeltabsyn, fill=modeltabsyn, opacity=0.78] coordinates {(1,0.617580) (2,0.572811) (3,0.472952)};
+\addplot+[mark=*, mark size=1.6pt, line width=0.85pt, draw=modeltvae, fill=modeltvae, opacity=0.78] coordinates {(1,0.536632) (2,0.382542) (3,0.311541)};
+\addplot+[mark=*, mark size=2.5pt, line width=1.8pt, draw=summaryblack, fill=summaryblack] coordinates {(1,0.616144) (2,0.548679) (3,0.479482)};
+\addplot+[only marks, mark=none, draw=summaryblack, error bars/.cd, y dir=both, y explicit] coordinates { (1,0.616144) +- (0,0.048622) };
+\addplot+[only marks, mark=none, draw=summaryblack, error bars/.cd, y dir=both, y explicit] coordinates { (2,0.548679) +- (0,0.051566) };
+\addplot+[only marks, mark=none, draw=summaryblack, error bars/.cd, y dir=both, y explicit] coordinates { (3,0.479482) +- (0,0.048690) };
+\end{groupplot}
+\end{tikzpicture}
+\end{document}

evaluation/query_family/conditional/locality_support_diagnostics/runs/20260519_192327_conditional_locality_support/figures/fig_conditional_support_by_model.tex

@@ -0,0 +1,65 @@
+\documentclass[tikz,border=4pt]{standalone}
+\usepackage{pgfplots}
+\usepgfplotslibrary{groupplots}
+\usepackage{xcolor}
+\pgfplotsset{compat=1.18}
+
+\definecolor{modelarf}{HTML}{777777}
+\definecolor{modelbayesnet}{HTML}{CCBB44}
+\definecolor{modelctgan}{HTML}{EE6677}
+\definecolor{modelforestdiffusion}{HTML}{228833}
+\definecolor{modelrealtabformer}{HTML}{332288}
+\definecolor{modeltabbyflow}{HTML}{882255}
+\definecolor{modeltabddpm}{HTML}{EE7733}
+\definecolor{modeltabdiff}{HTML}{AA3377}
+\definecolor{modeltabpfgen}{HTML}{009988}
+\definecolor{modeltabsyn}{HTML}{66CCEE}
+\definecolor{modeltvae}{HTML}{4477AA}
+\definecolor{summaryblack}{HTML}{000000}
+\begin{document}
+\begin{tikzpicture}
+\begin{axis}[
+width=13.8cm,
+height=8.4cm,
+ymin=0.0, ymax=1.0,
+title={Conditional support decomposition by model},
+ylabel={Filtered-local conditional fidelity},
+xtick={1,2,3},
+xticklabels={Dense,Medium,Sparse},
+ymajorgrids,
+grid style={draw=gray!20},
+major grid style={draw=gray!28},
+axis line style={draw=black!70},
+tick style={draw=black!70},
+legend style={draw=none, fill=none, font=\scriptsize, at={(0.98,0.98)}, anchor=north east},
+]
+\addplot+[mark=*, mark size=1.8pt, line width=0.9pt, draw=modelarf, fill=modelarf, opacity=0.82] coordinates {(1,0.662726) (2,0.583871) (3,0.583843)};
+\addlegendentry{ARF}
+\addplot+[mark=*, mark size=1.8pt, line width=0.9pt, draw=modelbayesnet, fill=modelbayesnet, opacity=0.82] coordinates {(1,0.696954) (2,0.613269) (3,0.553258)};
+\addlegendentry{BayesNet}
+\addplot+[mark=*, mark size=1.8pt, line width=0.9pt, draw=modelctgan, fill=modelctgan, opacity=0.82] coordinates {(1,0.614868) (2,0.480298) (3,0.455327)};
+\addlegendentry{CTGAN}
+\addplot+[mark=*, mark size=1.8pt, line width=0.9pt, draw=modelforestdiffusion, fill=modelforestdiffusion, opacity=0.82] coordinates {(1,0.445168) (2,0.482988) (3,0.362554)};
+\addlegendentry{ForestDiffusion}
+\addplot+[mark=*, mark size=1.8pt, line width=0.9pt, draw=modelrealtabformer, fill=modelrealtabformer, opacity=0.82] coordinates {(1,0.846226) (2,0.759283) (3,0.708563)};
+\addlegendentry{RealTabFormer}
+\addplot+[mark=*, mark size=1.8pt, line width=0.9pt, draw=modeltabbyflow, fill=modeltabbyflow, opacity=0.82] coordinates {(1,0.550897) (2,0.504824) (3,0.433465)};
+\addlegendentry{TabbyFlow}
+\addplot+[mark=*, mark size=1.8pt, line width=0.9pt, draw=modeltabddpm, fill=modeltabddpm, opacity=0.82] coordinates {(1,0.577591) (2,0.523680) (3,0.435008)};
+\addlegendentry{TabDDPM}
+\addplot+[mark=*, mark size=1.8pt, line width=0.9pt, draw=modeltabdiff, fill=modeltabdiff, opacity=0.82] coordinates {(1,0.571185) (2,0.527559) (3,0.437426)};
+\addlegendentry{TabDiff}
+\addplot+[mark=*, mark size=1.8pt, line width=0.9pt, draw=modeltabpfgen, fill=modeltabpfgen, opacity=0.82] coordinates {(1,0.645873) (2,0.605467) (3,0.505942)};
+\addlegendentry{TabPFGen}
+\addplot+[mark=*, mark size=1.8pt, line width=0.9pt, draw=modeltabsyn, fill=modeltabsyn, opacity=0.82] coordinates {(1,0.617580) (2,0.572811) (3,0.472952)};
+\addlegendentry{TabSyn}
+\addplot+[mark=*, mark size=1.8pt, line width=0.9pt, draw=modeltvae, fill=modeltvae, opacity=0.82] coordinates {(1,0.536632) (2,0.382542) (3,0.311541)};
+\addlegendentry{TVAE}
+\addplot+[mark=*, mark size=2.6pt, line width=1.8pt, draw=summaryblack, fill=summaryblack] coordinates {(1,0.616144) (2,0.548679) (3,0.479482)};
+\addlegendentry{Panel mean}
+\addplot+[only marks, mark=none, draw=summaryblack, error bars/.cd, y dir=both, y explicit] coordinates { (1,0.616144) +- (0,0.048622) };
+\addplot+[only marks, mark=none, draw=summaryblack, error bars/.cd, y dir=both, y explicit] coordinates { (2,0.548679) +- (0,0.051566) };
+\addplot+[only marks, mark=none, draw=summaryblack, error bars/.cd, y dir=both, y explicit] coordinates { (3,0.479482) +- (0,0.048690) };
+\end{axis}
+\end{tikzpicture}
+\end{document}

evaluation/query_family/conditional/locality_support_diagnostics/runs/20260519_192327_conditional_locality_support/figures/fig_conditional_support_main.tex

@@ -0,0 +1,53 @@
+\documentclass[tikz,border=4pt]{standalone}
+\usepackage{pgfplots}
+\usepgfplotslibrary{groupplots}
+\usepackage{xcolor}
+\pgfplotsset{compat=1.18}
+
+\definecolor{modelarf}{HTML}{777777}
+\definecolor{modelbayesnet}{HTML}{CCBB44}
+\definecolor{modelctgan}{HTML}{EE6677}
+\definecolor{modelforestdiffusion}{HTML}{228833}
+\definecolor{modelrealtabformer}{HTML}{332288}
+\definecolor{modeltabbyflow}{HTML}{882255}
+\definecolor{modeltabddpm}{HTML}{EE7733}
+\definecolor{modeltabdiff}{HTML}{AA3377}
+\definecolor{modeltabpfgen}{HTML}{009988}
+\definecolor{modeltabsyn}{HTML}{66CCEE}
+\definecolor{modeltvae}{HTML}{4477AA}
+\definecolor{summaryblack}{HTML}{000000}
+\begin{document}
+\begin{tikzpicture}
+\begin{axis}[
+width=13.8cm,
+height=8.4cm,
+ymin=0.0, ymax=1.0,
+title={Conditional support decomposition},
+ylabel={Filtered-local conditional fidelity},
+xtick={1,2,3},
+xticklabels={Dense,Medium,Sparse},
+ymajorgrids,
+grid style={draw=gray!20},
+major grid style={draw=gray!28},
+axis line style={draw=black!70},
+tick style={draw=black!70},
+legend style={draw=none, fill=none, font=\scriptsize, at={(0.98,0.98)}, anchor=north east},
+]
+\addplot+[mark=*, mark size=1.8pt, line width=0.9pt, draw=modelarf, fill=modelarf, opacity=0.82] coordinates {(1,0.662726) (2,0.583871) (3,0.583843)};
+\addplot+[mark=*, mark size=1.8pt, line width=0.9pt, draw=modelbayesnet, fill=modelbayesnet, opacity=0.82] coordinates {(1,0.696954) (2,0.613269) (3,0.553258)};
+\addplot+[mark=*, mark size=1.8pt, line width=0.9pt, draw=modelctgan, fill=modelctgan, opacity=0.82] coordinates {(1,0.614868) (2,0.480298) (3,0.455327)};
+\addplot+[mark=*, mark size=1.8pt, line width=0.9pt, draw=modelforestdiffusion, fill=modelforestdiffusion, opacity=0.82] coordinates {(1,0.445168) (2,0.482988) (3,0.362554)};
+\addplot+[mark=*, mark size=1.8pt, line width=0.9pt, draw=modelrealtabformer, fill=modelrealtabformer, opacity=0.82] coordinates {(1,0.846226) (2,0.759283) (3,0.708563)};
+\addplot+[mark=*, mark size=1.8pt, line width=0.9pt, draw=modeltabbyflow, fill=modeltabbyflow, opacity=0.82] coordinates {(1,0.550897) (2,0.504824) (3,0.433465)};
+\addplot+[mark=*, mark size=1.8pt, line width=0.9pt, draw=modeltabddpm, fill=modeltabddpm, opacity=0.82] coordinates {(1,0.577591) (2,0.523680) (3,0.435008)};
+\addplot+[mark=*, mark size=1.8pt, line width=0.9pt, draw=modeltabdiff, fill=modeltabdiff, opacity=0.82] coordinates {(1,0.571185) (2,0.527559) (3,0.437426)};
+\addplot+[mark=*, mark size=1.8pt, line width=0.9pt, draw=modeltabpfgen, fill=modeltabpfgen, opacity=0.82] coordinates {(1,0.645873) (2,0.605467) (3,0.505942)};
+\addplot+[mark=*, mark size=1.8pt, line width=0.9pt, draw=modeltabsyn, fill=modeltabsyn, opacity=0.82] coordinates {(1,0.617580) (2,0.572811) (3,0.472952)};
+\addplot+[mark=*, mark size=1.8pt, line width=0.9pt, draw=modeltvae, fill=modeltvae, opacity=0.82] coordinates {(1,0.536632) (2,0.382542) (3,0.311541)};
+\addplot+[mark=*, mark size=2.6pt, line width=1.8pt, draw=summaryblack, fill=summaryblack] coordinates {(1,0.616144) (2,0.548679) (3,0.479482)};
+\addplot+[only marks, mark=none, draw=summaryblack, error bars/.cd, y dir=both, y explicit] coordinates { (1,0.616144) +- (0,0.048622) };
+\addplot+[only marks, mark=none, draw=summaryblack, error bars/.cd, y dir=both, y explicit] coordinates { (2,0.548679) +- (0,0.051566) };
+\addplot+[only marks, mark=none, draw=summaryblack, error bars/.cd, y dir=both, y explicit] coordinates { (3,0.479482) +- (0,0.048690) };
+\end{axis}
+\end{tikzpicture}
+\end{document}

evaluation/query_family/conditional/locality_support_diagnostics/runs/20260519_192327_conditional_locality_support/manifest.json

@@ -0,0 +1,261 @@
+{
+  "task": "conditional_locality_support_diagnostic",
+  "generated_at_utc": "2026-05-19T19:23:40.029454+00:00",
+  "source_analysis_run": "trainonly_v2_current_success_official_20way_official20_20260519_232817",
+  "source_conditional_root": "/data/jialinzhang/TabQueryBench/code_snapshot/Evaluation/query_fivepart_breakdown/conditional_breakdown",
+  "run_tag": "20260519_192327_conditional_locality_support",
+  "run_dir": "/data/jialinzhang/TabQueryBench/code_snapshot/Evaluation/query_fivepart_breakdown/conditional_breakdown/locality_support_diagnostics/runs/20260519_192327_conditional_locality_support",
+  "primary_support_variant": "all_filtered_local",
+  "primary_support_reason": "primary_scalar_variant_missing",
+  "coverage": {
+    "conditional_query_rows": 17606,
+    "locality_dataset_model_panels": 404,
+    "filtered_local_query_rows": 2317,
+    "support_unique_cases": 223,
+    "support_primary_panel_rows": 744
+  },
+  "support_method_summary": {
+    "case_count": 223,
+    "template_count": 1,
+    "mode_counts": {
+      "exact": 214,
+      "unavailable": 9
+    },
+    "sql_artifact_found_count": 223,
+    "sql_artifact_missing_count": 0,
+    "main_eligible_case_count": 0
+  },
+  "support_variant_summary": {
+    "scalar_filtered_local": {
+      "eligible_case_count": 0,
+      "supported_dataset_count": 0,
+      "unsupported_dataset_count": 0,
+      "dataset_notes": []
+    },
+    "all_filtered_local": {
+      "eligible_case_count": 214,
+      "supported_dataset_count": 24,
+      "unsupported_dataset_count": 2,
+      "dataset_notes": [
+        {
+          "analysis_variant": "all_filtered_local",
+          "dataset_id": "c4",
+          "case_count": 9,
+          "unique_support_values": 9,
+          "bucketing_status": "ok"
+        },
+        {
+          "analysis_variant": "all_filtered_local",
+          "dataset_id": "m1",
+          "case_count": 9,
+          "unique_support_values": 9,
+          "bucketing_status": "ok"
+        },
+        {
+          "analysis_variant": "all_filtered_local",
+          "dataset_id": "c5",
+          "case_count": 2,
+          "unique_support_values": 2,
+          "bucketing_status": "unsupported_degenerate_within_dataset"
+        },
+        {
+          "analysis_variant": "all_filtered_local",
+          "dataset_id": "m4",
+          "case_count": 8,
+          "unique_support_values": 8,
+          "bucketing_status": "ok"
+        },
+        {
+          "analysis_variant": "all_filtered_local",
+          "dataset_id": "m10",
+          "case_count": 9,
+          "unique_support_values": 9,
+          "bucketing_status": "ok"
+        },
+        {
+          "analysis_variant": "all_filtered_local",
+          "dataset_id": "n8",
+          "case_count": 4,
+          "unique_support_values": 4,
+          "bucketing_status": "ok"
+        },
+        {
+          "analysis_variant": "all_filtered_local",
+          "dataset_id": "c7",
+          "case_count": 9,
+          "unique_support_values": 9,
+          "bucketing_status": "ok"
+        },
+        {
+          "analysis_variant": "all_filtered_local",
+          "dataset_id": "c8",
+          "case_count": 9,
+          "unique_support_values": 9,
+          "bucketing_status": "ok"
+        },
+        {
+          "analysis_variant": "all_filtered_local",
+          "dataset_id": "n15",
+          "case_count": 9,
+          "unique_support_values": 7,
+          "bucketing_status": "ok"
+        },
+        {
+          "analysis_variant": "all_filtered_local",
+          "dataset_id": "m7",
+          "case_count": 9,
+          "unique_support_values": 9,
+          "bucketing_status": "ok"
+        },
+        {
+          "analysis_variant": "all_filtered_local",
+          "dataset_id": "c16",
+          "case_count": 9,
+          "unique_support_values": 7,
+          "bucketing_status": "ok"
+        },
+        {
+          "analysis_variant": "all_filtered_local",
+          "dataset_id": "c17",
+          "case_count": 8,
+          "unique_support_values": 5,
+          "bucketing_status": "ok"
+        },
+        {
+          "analysis_variant": "all_filtered_local",
+          "dataset_id": "m6",
+          "case_count": 9,
+          "unique_support_values": 9,
+          "bucketing_status": "ok"
+        },
+        {
+          "analysis_variant": "all_filtered_local",
+          "dataset_id": "n17",
+          "case_count": 9,
+          "unique_support_values": 7,
+          "bucketing_status": "ok"
+        },
+        {
+          "analysis_variant": "all_filtered_local",
+          "dataset_id": "m9",
+          "case_count": 9,
+          "unique_support_values": 9,
+          "bucketing_status": "ok"
+        },
+        {
+          "analysis_variant": "all_filtered_local",
+          "dataset_id": "c11",
+          "case_count": 9,
+          "unique_support_values": 9,
+          "bucketing_status": "ok"
+        },
+        {
+          "analysis_variant": "all_filtered_local",
+          "dataset_id": "n5",
+          "case_count": 9,
+          "unique_support_values": 7,
+          "bucketing_status": "ok"
+        },
+        {
+          "analysis_variant": "all_filtered_local",
+          "dataset_id": "c9",
+          "case_count": 9,
+          "unique_support_values": 6,
+          "bucketing_status": "ok"
+        },
+        {
+          "analysis_variant": "all_filtered_local",
+          "dataset_id": "c19",
+          "case_count": 8,
+          "unique_support_values": 7,
+          "bucketing_status": "ok"
+        },
+        {
+          "analysis_variant": "all_filtered_local",
+          "dataset_id": "c18",
+          "case_count": 8,
+          "unique_support_values": 8,
+          "bucketing_status": "ok"
+        },
+        {
+          "analysis_variant": "all_filtered_local",
+          "dataset_id": "m12",
+          "case_count": 9,
+          "unique_support_values": 7,
+          "bucketing_status": "ok"
+        },
+        {
+          "analysis_variant": "all_filtered_local",
+          "dataset_id": "n19",
+          "case_count": 6,
+          "unique_support_values": 2,
+          "bucketing_status": "unsupported_degenerate_within_dataset"
+        },
+        {
+          "analysis_variant": "all_filtered_local",
+          "dataset_id": "c14",
+          "case_count": 9,
+          "unique_support_values": 9,
+          "bucketing_status": "ok"
+        },
+        {
+          "analysis_variant": "all_filtered_local",
+          "dataset_id": "n12",
+          "case_count": 8,
+          "unique_support_values": 4,
+          "bucketing_status": "ok"
+        },
+        {
+          "analysis_variant": "all_filtered_local",
+          "dataset_id": "c15",
+          "case_count": 9,
+          "unique_support_values": 9,
+          "bucketing_status": "ok"
+        },
+        {
+          "analysis_variant": "all_filtered_local",
+          "dataset_id": "c10",
+          "case_count": 9,
+          "unique_support_values": 9,
+          "bucketing_status": "ok"
+        }
+      ]
+    }
+  },
+  "compile_notes": {
+    "fig_conditional_locality_main": {
+      "success": false,
+      "note": "latexmk not available"
+    },
+    "fig_conditional_locality_by_model": {
+      "success": false,
+      "note": "latexmk not available"
+    },
+    "fig_conditional_support_main": {
+      "success": false,
+      "note": "latexmk not available"
+    },
+    "fig_conditional_support_by_model": {
+      "success": false,
+      "note": "latexmk not available"
+    },
+    "fig_conditional_locality_support_combined": {
+      "success": false,
+      "note": "latexmk not available"
+    },
+    "table_conditional_locality_summary": {
+      "success": false,
+      "note": "latexmk not available"
+    },
+    "table_conditional_support_summary": {
+      "success": false,
+      "note": "latexmk not available"
+    }
+  },
+  "key_findings": {
+    "locality_global": "Across panel means, conditional fidelity declines from grouped/global summaries (0.497) to 2D surfaces (0.949) and then to filtered/local slices (0.524).",
+    "locality_model": "The strongest grouped/global to filtered/local drop appears for TVAE, falling from 0.486 to 0.405.",
+    "support_global": "Within the exact-support filtered-local subset, dense slices score 0.616, medium slices 0.549, and sparse slices 0.479, consistent with a sparse-support penalty.",
+    "support_model": "Model behavior is mixed: 11 models have positive dense-minus-sparse gaps and 0 show the reverse; the largest positive gap is TVAE at 0.225."
+  }
+}

evaluation/query_family/conditional/locality_support_diagnostics/runs/20260519_192327_conditional_locality_support/report/conditional_locality_diagnostic.md

@@ -0,0 +1,30 @@
+# Conditional locality diagnostic
+
+## Classification audit
+
+Template-level semantics, not raw SQL column counts, define the primary locality buckets. The explicit mapping below keeps the two-axis filtered template in `filtered_local` while preserving `axis_arity = 2D` as a secondary annotation.
+
+| template_id                       | template_name                        | structure_type   | axis_arity   |   n_query_rows |   n_datasets |   n_models |
+|:----------------------------------|:-------------------------------------|:-----------------|:-------------|---------------:|-------------:|-----------:|
+| tpl_m4_group_condition_rate       | Grouped Condition Rate               | grouped_global   | 1D           |           3777 |           32 |         11 |
+| tpl_m4_group_ratio_two_conditions | Grouped Ratio of Two Conditions      | grouped_global   | 1D           |           2745 |           35 |         11 |
+| tpl_m4_window_partition_avg       | Window Partition Average             | grouped_global   | 1D           |           1808 |           21 |         11 |
+| tpl_tpcds_within_group_share      | Within-Group Share of Total          | grouped_global   | 1D           |           6343 |           33 |         11 |
+| tpl_c2_two_dim_target_rate        | Two-Axis Target Rate Surface         | surface_2d       | 2D           |            616 |            3 |         11 |
+| tpl_c2_filtered_group_count_2d    | Filtered Two-Dimensional Group Count | filtered_local   | 2D           |           2317 |           27 |         11 |
+
+## Coverage and scores
+
+| structure_type   | bucket_label     |   query_row_count |   dataset_count |   model_count |   panel_count |   template_count |   mean_score |   ci95_radius | coverage_note                        |
+|:-----------------|:-----------------|------------------:|----------------:|--------------:|--------------:|-----------------:|-------------:|--------------:|:-------------------------------------|
+| grouped_global   | Grouped / Global |             14673 |              39 |            11 |           404 |                4 |     0.49691  |      0.032741 | adequate                             |
+| surface_2d       | 2D Surface       |               616 |               3 |            11 |            33 |                1 |     0.948784 |      0.013789 | low_dataset_coverage,single_template |
+| filtered_local   | Filtered / Local |              2317 |              27 |            11 |           280 |                1 |     0.524149 |      0.045843 | single_template                      |
+
+## Diagnostic takeaways
+
+- Panel means decline from `0.497` for grouped/global queries to `0.949` for 2D surfaces and `0.524` for filtered/local slices.
+- The steepest grouped/global to filtered/local decline appears for `TVAE`: `0.486` to `0.405`.
+- `surface_2d` still rests on one template family, so the locality trend should be treated as structured diagnostic evidence rather than a universal law over all possible 2D conditional tasks.
+- The current conditional row export carries heuristic subitem labels. This locality decomposition therefore anchors on template semantics and panel-level aggregation instead of over-interpreting any single heuristic subitem tag.
+

evaluation/query_family/conditional/locality_support_diagnostics/runs/20260519_192327_conditional_locality_support/report/conditional_locality_support_report.md

@@ -0,0 +1,86 @@
+# Conditional locality and support report
+
+## Scope
+
+- Source conditional breakdown: `/data/jialinzhang/TabQueryBench/code_snapshot/Evaluation/query_fivepart_breakdown/conditional_breakdown`
+- Source analysis run: `trainonly_v2_current_success_official_20way_official20_20260519_232817`
+- Primary support variant: `all_filtered_local` (`primary_scalar_variant_missing`)
+- Stable aggregation is panel-first throughout: query rows -> dataset-model-bucket panels -> model/global summaries.
+- This diagnostic does not rerun the benchmark and does not overwrite upstream conditional outputs.
+
+## Main supported findings
+
+- Across panel means, conditional fidelity declines from grouped/global summaries (0.497) to 2D surfaces (0.949) and then to filtered/local slices (0.524).
+- The strongest grouped/global to filtered/local drop appears for TVAE, falling from 0.486 to 0.405.
+- Within the exact-support filtered-local subset, dense slices score 0.616, medium slices 0.549, and sparse slices 0.479, consistent with a sparse-support penalty.
+- Model behavior is mixed: 11 models have positive dense-minus-sparse gaps and 0 show the reverse; the largest positive gap is TVAE at 0.225.
+
+## Locality diagnostic
+
+# Conditional locality diagnostic
+
+## Classification audit
+
+Template-level semantics, not raw SQL column counts, define the primary locality buckets. The explicit mapping below keeps the two-axis filtered template in `filtered_local` while preserving `axis_arity = 2D` as a secondary annotation.
+
+| template_id                       | template_name                        | structure_type   | axis_arity   |   n_query_rows |   n_datasets |   n_models |
+|:----------------------------------|:-------------------------------------|:-----------------|:-------------|---------------:|-------------:|-----------:|
+| tpl_m4_group_condition_rate       | Grouped Condition Rate               | grouped_global   | 1D           |           3777 |           32 |         11 |
+| tpl_m4_group_ratio_two_conditions | Grouped Ratio of Two Conditions      | grouped_global   | 1D           |           2745 |           35 |         11 |
+| tpl_m4_window_partition_avg       | Window Partition Average             | grouped_global   | 1D           |           1808 |           21 |         11 |
+| tpl_tpcds_within_group_share      | Within-Group Share of Total          | grouped_global   | 1D           |           6343 |           33 |         11 |
+| tpl_c2_two_dim_target_rate        | Two-Axis Target Rate Surface         | surface_2d       | 2D           |            616 |            3 |         11 |
+| tpl_c2_filtered_group_count_2d    | Filtered Two-Dimensional Group Count | filtered_local   | 2D           |           2317 |           27 |         11 |
+
+## Coverage and scores
+
+| structure_type   | bucket_label     |   query_row_count |   dataset_count |   model_count |   panel_count |   template_count |   mean_score |   ci95_radius | coverage_note                        |
+|:-----------------|:-----------------|------------------:|----------------:|--------------:|--------------:|-----------------:|-------------:|--------------:|:-------------------------------------|
+| grouped_global   | Grouped / Global |             14673 |              39 |            11 |           404 |                4 |     0.49691  |      0.032741 | adequate                             |
+| surface_2d       | 2D Surface       |               616 |               3 |            11 |            33 |                1 |     0.948784 |      0.013789 | low_dataset_coverage,single_template |
+| filtered_local   | Filtered / Local |              2317 |              27 |            11 |           280 |                1 |     0.524149 |      0.045843 | single_template                      |
+
+## Diagnostic takeaways
+
+- Panel means decline from `0.497` for grouped/global queries to `0.949` for 2D surfaces and `0.524` for filtered/local slices.
+- The steepest grouped/global to filtered/local decline appears for `TVAE`: `0.486` to `0.405`.
+- `surface_2d` still rests on one template family, so the locality trend should be treated as structured diagnostic evidence rather than a universal law over all possible 2D conditional tasks.
+- The current conditional row export carries heuristic subitem labels. This locality decomposition therefore anchors on template semantics and panel-level aggregation instead of over-interpreting any single heuristic subitem tag.
+
+
+## Support diagnostic
+
+# Conditional support diagnostic
+
+## Feasibility and recovery modes
+
+- Source SQL artifact coverage found: `223` recovered cases; missing: `0`.
+- Support recovery modes on unique filtered-local cases: `{"exact": 214, "unavailable": 9}`.
+- Primary dense/medium/sparse variant: `all_filtered_local` (`primary_scalar_variant_missing`).
+- Exact vs proxy row-level coverage in the audit export: exact=`2227`, derived_exact=`0`, proxy=`0`, unavailable=`90`.
+
+## Coverage and scores
+
+| support_bucket   | bucket_label   |   query_row_count |   dataset_count |   model_count |   panel_count |   template_count |   mean_score |   ci95_radius | coverage_note   |
+|:-----------------|:---------------|------------------:|----------------:|--------------:|--------------:|-----------------:|-------------:|--------------:|:----------------|
+| dense            | Dense          |               726 |              24 |            11 |           248 |                1 |     0.616144 |      0.048622 | single_template |
+| medium           | Medium         |               678 |              24 |            11 |           248 |                1 |     0.548679 |      0.051566 | single_template |
+| sparse           | Sparse         |               735 |              24 |            11 |           248 |                1 |     0.479482 |      0.04869  | single_template |
+
+## Diagnostic takeaways
+
+- The global panel mean declines from `0.616` on dense filtered-local slices to `0.479` on sparse slices.
+- Model behavior is mixed: `11` models have positive dense-minus-sparse gaps, `0` show the reverse, and `0` are flat. The largest positive gap appears for `TVAE` at `0.225`.
+- In the broader `all_filtered_local` sensitivity view (`24` datasets), dense=`0.616` and sparse=`0.479`; the sparse-support penalty is clearer once the filtered 2D local template is included.
+- The primary support analysis intentionally keeps only scalar filtered-local templates in the main dense/medium/sparse comparison so that the support unit remains the count of real rows satisfying the local predicate.
+- Exact per-cell support is still recovered and audited for the filtered 2D group-count template, but that template is left as a sensitivity-only support basis because its natural support statistic is a cell-count distribution rather than a scalar slice size.
+- On this main scalar subset, sparse support does not by itself explain the filtered-local weakness. Any support-mediated interpretation should therefore be limited to model-specific behavior or to the broader sensitivity analysis, not promoted as a universal driver.
+- Any unsupported or unavailable support cases remain explicit in the audit CSV and are not silently folded into the main claim.
+
+
+## Caveats
+
+- `surface_2d` is still represented by one template family, so the locality trend should be described as a template-grounded diagnostic pattern rather than a universal statement about dimensionality alone.
+- The support main figure intentionally excludes the filtered 2D count template from the primary dense/medium/sparse claim because its most faithful support signal is a distribution of cell counts, not a single filtered-row count.
+- Existing heuristic subitem labels in the conditional row export do not perfectly align with template-level semantics, so this diagnostic relies on template semantics for bucket assignment and uses query-score panel means as the primary outcome.
+

evaluation/query_family/conditional/locality_support_diagnostics/runs/20260519_192327_conditional_locality_support/report/conditional_support_bucket_diagnostic.md

@@ -0,0 +1,27 @@
+# Conditional support diagnostic
+
+## Feasibility and recovery modes
+
+- Source SQL artifact coverage found: `223` recovered cases; missing: `0`.
+- Support recovery modes on unique filtered-local cases: `{"exact": 214, "unavailable": 9}`.
+- Primary dense/medium/sparse variant: `all_filtered_local` (`primary_scalar_variant_missing`).
+- Exact vs proxy row-level coverage in the audit export: exact=`2227`, derived_exact=`0`, proxy=`0`, unavailable=`90`.
+
+## Coverage and scores
+
+| support_bucket   | bucket_label   |   query_row_count |   dataset_count |   model_count |   panel_count |   template_count |   mean_score |   ci95_radius | coverage_note   |
+|:-----------------|:---------------|------------------:|----------------:|--------------:|--------------:|-----------------:|-------------:|--------------:|:----------------|
+| dense            | Dense          |               726 |              24 |            11 |           248 |                1 |     0.616144 |      0.048622 | single_template |
+| medium           | Medium         |               678 |              24 |            11 |           248 |                1 |     0.548679 |      0.051566 | single_template |
+| sparse           | Sparse         |               735 |              24 |            11 |           248 |                1 |     0.479482 |      0.04869  | single_template |
+
+## Diagnostic takeaways
+
+- The global panel mean declines from `0.616` on dense filtered-local slices to `0.479` on sparse slices.
+- Model behavior is mixed: `11` models have positive dense-minus-sparse gaps, `0` show the reverse, and `0` are flat. The largest positive gap appears for `TVAE` at `0.225`.
+- In the broader `all_filtered_local` sensitivity view (`24` datasets), dense=`0.616` and sparse=`0.479`; the sparse-support penalty is clearer once the filtered 2D local template is included.
+- The primary support analysis intentionally keeps only scalar filtered-local templates in the main dense/medium/sparse comparison so that the support unit remains the count of real rows satisfying the local predicate.
+- Exact per-cell support is still recovered and audited for the filtered 2D group-count template, but that template is left as a sensitivity-only support basis because its natural support statistic is a cell-count distribution rather than a scalar slice size.
+- On this main scalar subset, sparse support does not by itself explain the filtered-local weakness. Any support-mediated interpretation should therefore be limited to model-specific behavior or to the broader sensitivity analysis, not promoted as a universal driver.
+- Any unsupported or unavailable support cases remain explicit in the audit CSV and are not silently folded into the main claim.
+

evaluation/query_family/conditional/locality_support_diagnostics/runs/20260519_192327_conditional_locality_support/report/paper_caption.txt

@@ -0,0 +1,8 @@
+Figure 1. Conditional locality decomposition.
+Across panel means, conditional fidelity declines from grouped/global summaries (0.497) to 2D surfaces (0.949) and then to filtered/local slices (0.524). Points and error bars show panel means with 95% confidence intervals; colored traces show per-model means under the frozen model roster and color convention.
+
+Figure 2. Conditional support decomposition.
+Within the exact-support filtered-local subset, dense slices score 0.616, medium slices 0.549, and sparse slices 0.479, consistent with a sparse-support penalty. The main support figure uses the all filtered-local templates subset so that support is measured on a comparable exact real-row-count scale within each dataset.
+
+Figure 3. Combined conditional locality/support diagnostic.
+Panel A shows the locality decomposition from grouped/global summaries to filtered/local slices. Panel B shows the dense/medium/sparse comparison inside the filtered-local subset. Both panels use panel-level aggregation and expose coverage caveats in the companion audit tables rather than hiding thin buckets.

evaluation/query_family/conditional/locality_support_diagnostics/runs/20260519_192327_conditional_locality_support/report/paper_paragraphs.md

@@ -0,0 +1,5 @@
+Across panel means, conditional fidelity declines from grouped/global summaries (0.497) to 2D surfaces (0.949) and then to filtered/local slices (0.524). This suggests that axis count alone is not the most interpretable explanation for the conditional-family weakness: grouped/global summaries remain comparatively more stable, while narrow filtered slices are harder to preserve.
+
+Within the exact-support filtered-local subset, dense slices score 0.616, medium slices 0.549, and sparse slices 0.479, consistent with a sparse-support penalty. That pattern indicates that sparse real support explains part of the local-slice collapse, consistent with synthetic generators smoothing away rare conditional interactions.
+
+Model behavior is mixed: 11 models have positive dense-minus-sparse gaps and 0 show the reverse; the largest positive gap is TVAE at 0.225. At the same time, the support diagnostic does not fully explain the conditional gap on its own: even dense local slices can remain weak for some models, and the 2D-surface bucket still rests on limited template coverage.

evaluation/query_family/conditional/locality_support_diagnostics/runs/20260519_192327_conditional_locality_support/tables/table_conditional_locality_summary.tex

@@ -0,0 +1,16 @@
+\documentclass{standalone}
+\usepackage[table]{xcolor}
+\usepackage{booktabs}
+\begin{document}
+\scriptsize
+\emph{Panel-level locality summary.}\\[0.4em]
+\begin{tabular}{llllll}
+\toprule
+Bucket & Panels & Datasets & Templates & Mean & 95\% CI \\
+\midrule
+Grouped / Global & 404 & 39 & 4 & 0.497 & 0.033 \\
+2D Surface & 33 & 3 & 1 & 0.949 & 0.014 \\
+Filtered / Local & 280 & 27 & 1 & 0.524 & 0.046 \\
+\bottomrule
+\end{tabular}
+\end{document}

evaluation/query_family/conditional/locality_support_diagnostics/runs/20260519_192327_conditional_locality_support/tables/table_conditional_support_summary.tex

@@ -0,0 +1,16 @@
+\documentclass{standalone}
+\usepackage[table]{xcolor}
+\usepackage{booktabs}
+\begin{document}
+\scriptsize
+\emph{Panel-level support summary.}\\[0.4em]
+\begin{tabular}{llllll}
+\toprule
+Bucket & Panels & Datasets & Templates & Mean & 95\% CI \\
+\midrule
+Dense & 248 & 24 & 1 & 0.616 & 0.049 \\
+Medium & 248 & 24 & 1 & 0.549 & 0.052 \\
+Sparse & 248 & 24 & 1 & 0.479 & 0.049 \\
+\bottomrule
+\end{tabular}
+\end{document}

evaluation/query_family/conditional/locality_support_diagnostics/runs/20260524_090854_conditional_locality_support/README.md

@@ -0,0 +1,8 @@
+# 20260524_090854_conditional_locality_support
+
+This run contains the full reproducible bundle for the conditional locality/support diagnostic.
+
+- `data/` exports the summary and audit CSVs.
+- `figures/` holds the paper-facing figures plus standalone TeX sources.
+- `tables/` holds LaTeX table snippets.
+- `report/` holds the Markdown narrative, captions, and paper paragraphs.

evaluation/query_family/conditional/locality_support_diagnostics/runs/20260524_090854_conditional_locality_support/data/conditional_locality_summary.csv

@@ -0,0 +1,4 @@
+structure_type,bucket_label,panel_count,dataset_count,model_count,template_count,query_row_count,prefix_coverage,prefix_count,mean_score,std_score,se_score,ci95_low,ci95_high,ci95_radius,coverage_note
+grouped_global,Grouped / Global,404,39,11,4,14673,"c,m,n",3,0.49691,0.335757,0.016705,0.46417,0.529651,0.032741,adequate
+surface_2d,2D Surface,33,3,11,1,616,c,1,0.948784,0.040414,0.007035,0.934994,0.962573,0.013789,"low_dataset_coverage,single_template"
+filtered_local,Filtered / Local,280,27,11,1,2317,"c,m,n",3,0.524149,0.39138,0.023389,0.478305,0.569992,0.045843,single_template