Upload benchmark query-category table assets

evaluation/tables/benchmark_query_category_table/README.md

@@ -0,0 +1,38 @@
+# Benchmark Query Category Table
+
+This folder contains a compact benchmark summary table that aggregates the retained benchmark channels by dataset category.
+
+## Goal
+
+Create a single standalone LaTeX table with:
+
+- rows: `Categorical`, `Numerical`, `Mix`
+- columns: `Distance`, `Subgroup`, `Conditional`, `Tail`, `Missingness`, `Cardinality`
+- cell values: arithmetic means across the available synthetic generators
+- direct PDF export for review
+- companion CSV exports for traceability
+
+## Files
+
+- `build_benchmark_query_category_table.py`
+  - Rebuilds the compact table from the current evaluation assets.
+- `final/benchmark_query_category_table_real.tex`
+  - Standalone LaTeX source for the review PDF.
+- `final/benchmark_query_category_table_real_source.tex`
+  - Paper-facing embedded table snippet copied into the active paper folder.
+- `final/benchmark_query_category_table_real_summary.csv`
+  - Category-level summary with per-metric generator-averaged means, stds, and model counts.
+- `final/benchmark_query_category_table_real_dataset_level.csv`
+  - Dataset-level metric values used to compute the compact summary.
+- `final/benchmark_query_category_table_real_sources.csv`
+  - Metric-to-source manifest for traceability.
+
+## Rebuild
+
+From the repo root:
+
+```powershell
+python Evaluation\benchmark_query_category_table\build_benchmark_query_category_table.py
+```
+
+The script also attempts to compile the standalone PDF and to mirror the paper-facing source snippet into the active paper figure directory.

evaluation/tables/benchmark_query_category_table/build_benchmark_query_category_table.py

@@ -0,0 +1,793 @@
+from __future__ import annotations
+
+import re
+import shutil
+import subprocess
+from pathlib import Path
+
+import pandas as pd
+
+
+REPO_ROOT = Path(__file__).resolve().parents[2]
+EVAL_ROOT = REPO_ROOT / "Evaluation"
+TABLE_ROOT = EVAL_ROOT / "benchmark_query_category_table"
+OUT_DIR = TABLE_ROOT / "final"
+PAPER_FIGURE_SUBDIR = "figures/benchmark_query_category_table"
+FINAL_BASENAME = "benchmark_query_category_table_real"
+
+PAPER_MODEL_ORDER = [
+    "real",
+    "arf",
+    "bayesnet",
+    "ctgan",
+    "forestdiffusion",
+    "realtabformer",
+    "tabbyflow",
+    "tabddpm",
+    "tabdiff",
+    "tabpfgen",
+    "tabsyn",
+    "tvae",
+]
+SYNTHETIC_MODEL_ORDER = [model_id for model_id in PAPER_MODEL_ORDER if model_id != "real"]
+MODEL_ORDER_MAP = {model_id: idx for idx, model_id in enumerate(PAPER_MODEL_ORDER)}
+
+PREFIX_ORDER = ["c", "n", "m"]
+PREFIX_LABELS = {
+    "c": "Categorical",
+    "n": "Numerical",
+    "m": "Mix",
+}
+PREFIX_ORDER_MAP = {prefix: idx for idx, prefix in enumerate(PREFIX_ORDER)}
+
+METRICS = [
+    {
+        "key": "distance_overall",
+        "family": "Distance",
+        "group_label": "Distance",
+        "header_label": "Dist. overall",
+        "full_label": "Distance overall",
+        "highlight_column": True,
+        "source_kind": "distance_dataset_export",
+        "source_file": "Evaluation/distance/runs/*/datasets/*/distance_summary__*.csv",
+        "source_note": "Latest deduplicated dataset-model overall fidelity score.",
+    },
+    {
+        "key": "internal_profile_stability",
+        "family": "Subgroup",
+        "group_label": "Subgroup",
+        "header_label": "Internal profile",
+        "full_label": "Internal profile stability",
+        "highlight_column": False,
+        "source_kind": "query_subitem_dataset_export",
+        "source_file": "Evaluation/query_fivepart_breakdown/subgroup_breakdown/data/dataset_model_subitems.csv",
+        "source_note": "Canonical subgroup sub-item score from dataset-level sub-item export.",
+    },
+    {
+        "key": "subgroup_size_stability",
+        "family": "Subgroup",
+        "group_label": "Subgroup",
+        "header_label": "Size stability",
+        "full_label": "Subgroup size stability",
+        "highlight_column": False,
+        "source_kind": "query_subitem_dataset_export",
+        "source_file": "Evaluation/query_fivepart_breakdown/subgroup_breakdown/data/dataset_model_subitems.csv",
+        "source_note": "Canonical subgroup sub-item score from dataset-level sub-item export.",
+    },
+    {
+        "key": "dependency_strength_similarity",
+        "family": "Conditional",
+        "group_label": "Conditional",
+        "header_label": "Dependency strength",
+        "full_label": "Dependency strength similarity",
+        "highlight_column": False,
+        "source_kind": "query_subitem_dataset_export",
+        "source_file": "Evaluation/query_fivepart_breakdown/conditional_breakdown/data/dataset_model_subitems.csv",
+        "source_note": "Canonical conditional sub-item score from dataset-level sub-item export.",
+    },
+    {
+        "key": "direction_consistency",
+        "family": "Conditional",
+        "group_label": "Conditional",
+        "header_label": "Direction",
+        "full_label": "Direction consistency",
+        "highlight_column": False,
+        "source_kind": "query_subitem_dataset_export",
+        "source_file": "Evaluation/query_fivepart_breakdown/conditional_breakdown/data/dataset_model_subitems.csv",
+        "source_note": "Canonical conditional sub-item score from dataset-level sub-item export.",
+    },
+    {
+        "key": "slice_level_consistency",
+        "family": "Conditional",
+        "group_label": "Conditional",
+        "header_label": "Slice-level",
+        "full_label": "Slice-level consistency",
+        "highlight_column": False,
+        "source_kind": "query_subitem_dataset_export",
+        "source_file": "Evaluation/query_fivepart_breakdown/conditional_breakdown/data/dataset_model_subitems.csv",
+        "source_note": "Canonical conditional sub-item score from dataset-level sub-item export.",
+    },
+    {
+        "key": "tail_set_consistency",
+        "family": "Tail",
+        "group_label": "Tail",
+        "header_label": "Tail set",
+        "full_label": "Tail set consistency",
+        "highlight_column": False,
+        "source_kind": "query_subitem_dataset_export",
+        "source_file": "Evaluation/query_fivepart_breakdown/tail_breakdown/data/dataset_model_scores.csv",
+        "source_note": "Canonical tail sub-item score from dataset-level breakdown export.",
+    },
+    {
+        "key": "tail_mass_similarity",
+        "family": "Tail",
+        "group_label": "Tail",
+        "header_label": "Tail mass",
+        "full_label": "Tail mass similarity",
+        "highlight_column": False,
+        "source_kind": "query_subitem_dataset_export",
+        "source_file": "Evaluation/query_fivepart_breakdown/tail_breakdown/data/dataset_model_scores.csv",
+        "source_note": "Canonical tail sub-item score from dataset-level breakdown export.",
+    },
+    {
+        "key": "tail_concentration_consistency",
+        "family": "Tail",
+        "group_label": "Tail",
+        "header_label": "Tail concentration",
+        "full_label": "Tail concentration consistency",
+        "highlight_column": False,
+        "source_kind": "query_subitem_dataset_export",
+        "source_file": "Evaluation/query_fivepart_breakdown/tail_breakdown/data/dataset_model_scores.csv",
+        "source_note": "Canonical tail sub-item score from dataset-level breakdown export.",
+    },
+    {
+        "key": "marginal_missing_rate_consistency",
+        "family": "Missingness",
+        "group_label": "Missingness",
+        "header_label": "Marginal rate",
+        "full_label": "Marginal missing-rate consistency",
+        "highlight_column": False,
+        "source_kind": "query_subitem_dataset_export",
+        "source_file": "Evaluation/query_fivepart_breakdown/missingness_breakdown/data/dataset_model_subitems.csv",
+        "source_note": "Canonical missingness sub-item score from dataset-level sub-item export.",
+    },
+    {
+        "key": "co_missingness_pattern_consistency",
+        "family": "Missingness",
+        "group_label": "Missingness",
+        "header_label": "Co-missingness",
+        "full_label": "Co-missingness pattern consistency",
+        "highlight_column": False,
+        "source_kind": "query_subitem_dataset_export",
+        "source_file": "Evaluation/query_fivepart_breakdown/missingness_breakdown/data/dataset_model_subitems.csv",
+        "source_note": "Canonical missingness sub-item score from dataset-level sub-item export.",
+    },
+    {
+        "key": "support_rank_profile_consistency",
+        "family": "Cardinality",
+        "group_label": "Cardinality",
+        "header_label": "Support-rank",
+        "full_label": "Support-rank profile consistency",
+        "highlight_column": False,
+        "source_kind": "cardinality_subitem_dataset_export",
+        "source_file": "Evaluation/query_fivepart_breakdown/cardinality/data/cleaned_results.csv",
+        "source_note": "Dataset-level mean of the discrete support-retention branch.",
+    },
+    {
+        "key": "high_cardinality_response_stability",
+        "family": "Cardinality",
+        "group_label": "Cardinality",
+        "header_label": "High-cardinality",
+        "full_label": "High-cardinality response stability",
+        "highlight_column": False,
+        "source_kind": "cardinality_subitem_dataset_export",
+        "source_file": "Evaluation/query_fivepart_breakdown/cardinality/data/cleaned_results.csv",
+        "source_note": "Dataset-level mean over high-dynamic cardinality/range units.",
+    },
+]
+
+METRIC_SPECS = {metric["key"]: metric for metric in METRICS}
+METRIC_ORDER = [metric["key"] for metric in METRICS]
+FAMILY_ORDER = ["Distance", "Subgroup", "Conditional", "Tail", "Missingness", "Cardinality"]
+
+
+def ensure_out_dir() -> None:
+    OUT_DIR.mkdir(parents=True, exist_ok=True)
+
+
+def pick_active_paper_dir() -> Path:
+    candidates = sorted((REPO_ROOT / "Paper").glob("*/main.tex"))
+    if not candidates:
+        raise FileNotFoundError("Could not find Paper/*/main.tex.")
+    return candidates[0].parent
+
+
+def dataset_prefix(dataset_id: str) -> str:
+    return str(dataset_id)[0].lower() if dataset_id else ""
+
+
+def dataset_sort_tuple(dataset_id: str) -> tuple[int, int, str]:
+    text = str(dataset_id)
+    match = re.match(r"([A-Za-z]+)(\d+)", text)
+    if not match:
+        return (99, 10**9, text)
+    prefix, number = match.groups()
+    prefix_order = {"c": 0, "m": 1, "n": 2}.get(prefix.lower(), 50)
+    return (prefix_order, int(number), text)
+
+
+def latex_escape(text: str) -> str:
+    replacements = {
+        "\\": r"\textbackslash{}",
+        "&": r"\&",
+        "%": r"\%",
+        "$": r"\$",
+        "#": r"\#",
+        "_": r"\_",
+        "{": r"\{",
+        "}": r"\}",
+        "~": r"\textasciitilde{}",
+        "^": r"\textasciicircum{}",
+    }
+    escaped = str(text)
+    for src, dst in replacements.items():
+        escaped = escaped.replace(src, dst)
+    return escaped
+
+
+def fmt_num(value: float) -> str:
+    if abs(value) < 5e-4:
+        value = 0.0
+    return f"{value:.2f}"
+
+
+def normalize_long_metric(
+    df: pd.DataFrame,
+    *,
+    metric_key: str,
+    value_col: str,
+    source_path: str,
+    source_kind: str,
+) -> pd.DataFrame:
+    working = df.copy()
+    working = working[working["model_id"].isin(SYNTHETIC_MODEL_ORDER)].copy()
+    if "dataset_prefix" not in working.columns:
+        working["dataset_prefix"] = working["dataset_id"].map(dataset_prefix)
+    working["dataset_prefix_label"] = working["dataset_prefix"].map(PREFIX_LABELS)
+    working["metric_key"] = metric_key
+    working["metric_family"] = METRIC_SPECS[metric_key]["family"]
+    working["metric_group_label"] = METRIC_SPECS[metric_key]["group_label"]
+    working["metric_header_label"] = METRIC_SPECS[metric_key]["header_label"]
+    working["metric_full_label"] = METRIC_SPECS[metric_key]["full_label"]
+    working["metric_value"] = pd.to_numeric(working[value_col], errors="coerce")
+    working["source_kind"] = source_kind
+    working["source_path"] = source_path
+    working["row_kind"] = "synthetic"
+    working["model_order"] = working["model_id"].map(MODEL_ORDER_MAP)
+    return working[
+        [
+            "dataset_id",
+            "dataset_prefix",
+            "dataset_prefix_label",
+            "model_id",
+            "metric_key",
+            "metric_family",
+            "metric_group_label",
+            "metric_header_label",
+            "metric_full_label",
+            "metric_value",
+            "source_kind",
+            "source_path",
+            "row_kind",
+            "model_order",
+        ]
+    ]
+
+
+def load_distance_long() -> pd.DataFrame:
+    frames = []
+    cols = ["dataset_id", "model_id", "timestamp_utc", "overall_fidelity_score"]
+    for path in (EVAL_ROOT / "distance" / "runs").glob("*/*/distance_summary__*.csv"):
+        try:
+            df = pd.read_csv(path, usecols=cols, low_memory=False)
+        except Exception:
+            continue
+        frames.append(df)
+    if not frames:
+        raise FileNotFoundError("No distance dataset exports were found under Evaluation/distance/runs.")
+
+    df = pd.concat(frames, ignore_index=True)
+    df = df[df["model_id"].isin(SYNTHETIC_MODEL_ORDER)].copy()
+    df["timestamp_utc"] = pd.to_datetime(df["timestamp_utc"], errors="coerce")
+    df["dataset_prefix"] = df["dataset_id"].map(dataset_prefix)
+    df = df.sort_values(["dataset_id", "model_id", "timestamp_utc"]).drop_duplicates(
+        ["dataset_id", "model_id"],
+        keep="last",
+    )
+    return normalize_long_metric(
+        df[["dataset_id", "dataset_prefix", "model_id", "overall_fidelity_score"]],
+        metric_key="distance_overall",
+        value_col="overall_fidelity_score",
+        source_path="Evaluation/distance/runs/*/datasets/*/distance_summary__*.csv (latest dataset-model export)",
+        source_kind="distance_dataset_export",
+    )
+
+
+def load_subgroup_long() -> pd.DataFrame:
+    subitems_path = EVAL_ROOT / "query_fivepart_breakdown" / "subgroup_breakdown" / "data" / "dataset_model_subitems.csv"
+    subitems_df = pd.read_csv(subitems_path)
+
+    frames = []
+
+    for subitem_key in ["internal_profile_stability", "subgroup_size_stability"]:
+        subset = subitems_df[subitems_df["subitem_id"] == subitem_key].copy()
+        frames.append(
+            normalize_long_metric(
+                subset[["dataset_id", "dataset_prefix", "model_id", "subitem_score"]],
+                metric_key=subitem_key,
+                value_col="subitem_score",
+                source_path=str(subitems_path.relative_to(REPO_ROOT)),
+                source_kind="query_subitem_dataset_export",
+            )
+        )
+    return pd.concat(frames, ignore_index=True)
+
+
+def load_conditional_long() -> pd.DataFrame:
+    subitems_path = EVAL_ROOT / "query_fivepart_breakdown" / "conditional_breakdown" / "data" / "dataset_model_subitems.csv"
+    subitems_df = pd.read_csv(subitems_path)
+
+    frames = []
+
+    for subitem_key in ["dependency_strength_similarity", "direction_consistency", "slice_level_consistency"]:
+        subset = subitems_df[subitems_df["subitem_id"] == subitem_key].copy()
+        frames.append(
+            normalize_long_metric(
+                subset[["dataset_id", "dataset_prefix", "model_id", "subitem_score"]],
+                metric_key=subitem_key,
+                value_col="subitem_score",
+                source_path=str(subitems_path.relative_to(REPO_ROOT)),
+                source_kind="query_subitem_dataset_export",
+            )
+        )
+    return pd.concat(frames, ignore_index=True)
+
+
+def load_tail_long() -> pd.DataFrame:
+    path = EVAL_ROOT / "query_fivepart_breakdown" / "tail_breakdown" / "data" / "dataset_model_scores.csv"
+    df = pd.read_csv(path)
+    frames = []
+    for metric_key in [
+        "tail_set_consistency",
+        "tail_mass_similarity",
+        "tail_concentration_consistency",
+    ]:
+        frames.append(
+            normalize_long_metric(
+                df[["dataset_id", "dataset_prefix", "model_id", metric_key]],
+                metric_key=metric_key,
+                value_col=metric_key,
+                source_path=str(path.relative_to(REPO_ROOT)),
+                source_kind="query_subitem_dataset_export",
+            )
+        )
+    return pd.concat(frames, ignore_index=True)
+
+
+def load_missingness_long() -> pd.DataFrame:
+    subitems_path = EVAL_ROOT / "query_fivepart_breakdown" / "missingness_breakdown" / "data" / "dataset_model_subitems.csv"
+    subitems_df = pd.read_csv(subitems_path)
+
+    frames = []
+
+    for subitem_key in ["marginal_missing_rate_consistency", "co_missingness_pattern_consistency"]:
+        subset = subitems_df[subitems_df["subitem_id"] == subitem_key].copy()
+        frames.append(
+            normalize_long_metric(
+                subset[["dataset_id", "dataset_prefix", "model_id", "subitem_score"]],
+                metric_key=subitem_key,
+                value_col="subitem_score",
+                source_path=str(subitems_path.relative_to(REPO_ROOT)),
+                source_kind="query_subitem_dataset_export",
+            )
+        )
+    return pd.concat(frames, ignore_index=True)
+
+
+def load_cardinality_long() -> pd.DataFrame:
+    path = EVAL_ROOT / "query_fivepart_breakdown" / "cardinality" / "data" / "cleaned_results.csv"
+    df = pd.read_csv(
+        path,
+        low_memory=False,
+        usecols=[
+            "dataset",
+            "model",
+            "channel",
+            "score",
+            "dynamic_regime",
+        ],
+    )
+    df = df.rename(columns={"dataset": "dataset_id", "model": "model_id"}).copy()
+    df["model_id"] = df["model_id"].astype(str).str.strip().str.lower()
+    df["dataset_prefix"] = df["dataset_id"].map(dataset_prefix)
+
+    support_rank_df = (
+        df[df["channel"] == "discrete"]
+        .groupby(["dataset_id", "dataset_prefix", "model_id"], as_index=False)["score"]
+        .mean()
+        .rename(columns={"score": "support_rank_profile_consistency"})
+    )
+    high_dynamic_df = (
+        df[df["dynamic_regime"] == "high"]
+        .groupby(["dataset_id", "dataset_prefix", "model_id"], as_index=False)["score"]
+        .mean()
+        .rename(columns={"score": "high_cardinality_response_stability"})
+    )
+
+    frames = [
+        normalize_long_metric(
+            support_rank_df,
+            metric_key="support_rank_profile_consistency",
+            value_col="support_rank_profile_consistency",
+            source_path=str(path.relative_to(REPO_ROOT)),
+            source_kind="cardinality_subitem_dataset_export",
+        ),
+        normalize_long_metric(
+            high_dynamic_df,
+            metric_key="high_cardinality_response_stability",
+            value_col="high_cardinality_response_stability",
+            source_path=str(path.relative_to(REPO_ROOT)),
+            source_kind="cardinality_subitem_dataset_export",
+        ),
+    ]
+    return pd.concat(frames, ignore_index=True)
+
+
+def assemble_dataset_level_table() -> pd.DataFrame:
+    frames = [
+        load_distance_long(),
+        load_subgroup_long(),
+        load_conditional_long(),
+        load_tail_long(),
+        load_missingness_long(),
+        load_cardinality_long(),
+    ]
+    long_df = pd.concat(frames, ignore_index=True)
+    long_df["metric_value"] = pd.to_numeric(long_df["metric_value"], errors="coerce")
+    long_df["prefix_order"] = long_df["dataset_prefix"].map(PREFIX_ORDER_MAP)
+    long_df["metric_order"] = long_df["metric_key"].map({metric_key: idx for idx, metric_key in enumerate(METRIC_ORDER)})
+    long_df["dataset_sort_key"] = long_df["dataset_id"].map(dataset_sort_tuple)
+    return (
+        long_df.sort_values(["prefix_order", "metric_order", "dataset_sort_key", "model_order"])
+        .drop(columns=["dataset_sort_key"])
+        .reset_index(drop=True)
+    )
+
+
+def build_category_summary(dataset_level: pd.DataFrame) -> pd.DataFrame:
+    per_model = (
+        dataset_level.groupby(
+            [
+                "dataset_prefix",
+                "dataset_prefix_label",
+                "model_id",
+                "metric_key",
+                "metric_family",
+                "metric_group_label",
+                "metric_header_label",
+                "metric_full_label",
+            ],
+            as_index=False,
+        )
+        .agg(
+            model_metric_mean=("metric_value", "mean"),
+            covered_datasets=("dataset_id", "nunique"),
+        )
+    )
+
+    summary = (
+        per_model.groupby(
+            [
+                "dataset_prefix",
+                "dataset_prefix_label",
+                "metric_key",
+                "metric_family",
+                "metric_group_label",
+                "metric_header_label",
+                "metric_full_label",
+            ],
+            as_index=False,
+        )
+        .agg(
+            category_mean=("model_metric_mean", "mean"),
+            category_std=("model_metric_mean", "std"),
+            model_count=("model_metric_mean", "count"),
+            min_covered_datasets=("covered_datasets", "min"),
+            max_covered_datasets=("covered_datasets", "max"),
+        )
+    )
+    summary.loc[summary["model_count"] <= 1, "category_std"] = 0.0
+    summary["prefix_order"] = summary["dataset_prefix"].map(PREFIX_ORDER_MAP)
+    summary["metric_order"] = summary["metric_key"].map({metric_key: idx for idx, metric_key in enumerate(METRIC_ORDER)})
+    summary["rank"] = pd.NA
+
+    for metric_key in METRIC_ORDER:
+        subset = summary[(summary["metric_key"] == metric_key) & summary["category_mean"].notna()].sort_values(
+            "category_mean",
+            ascending=False,
+        )
+        for idx, (_, row) in enumerate(subset.iterrows(), start=1):
+            summary.loc[
+                (summary["dataset_prefix"] == row["dataset_prefix"]) & (summary["metric_key"] == metric_key),
+                "rank",
+            ] = idx
+
+    return summary.sort_values(["prefix_order", "metric_order"]).reset_index(drop=True)
+
+
+def render_category_cell(summary: pd.DataFrame, prefix: str, metric_key: str) -> str:
+    row = summary[
+        (summary["dataset_prefix"] == prefix)
+        & (summary["metric_key"] == metric_key)
+    ]
+    spec = METRIC_SPECS[metric_key]
+    prefix_text = r"\cellcolor{OverallTint} " if spec["highlight_column"] else ""
+    if row.empty or pd.isna(row.iloc[0]["category_mean"]):
+        return prefix_text + r"\textit{N/A}"
+
+    mean_value = float(row.iloc[0]["category_mean"])
+    std_value = 0.0 if pd.isna(row.iloc[0]["category_std"]) else float(row.iloc[0]["category_std"])
+    rank_value = None if pd.isna(row.iloc[0]["rank"]) else int(row.iloc[0]["rank"])
+    body = f"{fmt_num(mean_value)}$_{{\\pm {fmt_num(std_value)}}}$"
+    if rank_value == 1:
+        body = rf"{{\color{{FirstPlace}}\textbf{{{body}}}}}"
+    elif rank_value == 2:
+        body = rf"{{\color{{SecondPlace}}\textbf{{{body}}}}}"
+    elif rank_value == 3:
+        body = rf"{{\color{{ThirdPlace}}\textbf{{{body}}}}}"
+    return prefix_text + body
+
+
+def render_embedded_source(summary: pd.DataFrame) -> str:
+    tabular_spec = "@{}l " + " ".join("c" for _ in METRIC_ORDER) + "@{}"
+
+    top_header_parts = []
+    cmidrules = []
+    start_col = 2
+    for family in FAMILY_ORDER:
+        family_metrics = [metric for metric in METRICS if metric["family"] == family]
+        span = len(family_metrics)
+        top_header_parts.append(rf"\multicolumn{{{span}}}{{c}}{{\textbf{{{latex_escape(family)}}}}}")
+        cmidrules.append(rf"\cmidrule(lr){{{start_col}-{start_col + span - 1}}}")
+        start_col += span
+    top_header = " & ".join(top_header_parts)
+    second_header = " & ".join(
+        (
+            (rf"\cellcolor{{OverallTint}} {latex_escape(metric['header_label'])} $\uparrow$")
+            if metric["highlight_column"]
+            else rf"{latex_escape(metric['header_label'])} $\uparrow$"
+        )
+        for metric in METRICS
+    )
+
+    body_lines = []
+    for prefix in PREFIX_ORDER:
+        cells = [latex_escape(PREFIX_LABELS[prefix])]
+        for metric_key in METRIC_ORDER:
+            cells.append(render_category_cell(summary, prefix, metric_key))
+        body_lines.append(" & ".join(cells) + r" \\")
+
+    return rf"""{{
+\definecolor{{FirstPlace}}{{HTML}}{{1397B8}}%
+\definecolor{{SecondPlace}}{{HTML}}{{7B45E5}}%
+\definecolor{{ThirdPlace}}{{HTML}}{{000000}}%
+\definecolor{{OverallTint}}{{HTML}}{{F8F1DA}}%
+\definecolor{{RuleGray}}{{HTML}}{{C8CDD3}}%
+\arrayrulecolor{{RuleGray}}%
+\setlength{{\tabcolsep}}{{3.0pt}}%
+\renewcommand{{\arraystretch}}{{1.08}}%
+\scriptsize
+\begin{{adjustbox}}{{max width=\textwidth}}
+\begin{{tabular}}{{{tabular_spec}}}
+\toprule
+\multirow{{2}}{{*}}{{\textbf{{Category}}}} & {top_header} \\
+{chr(10).join(cmidrules)}
+ & {second_header} \\
+\midrule
+{chr(10).join(body_lines)}
+\bottomrule
+\end{{tabular}}
+\end{{adjustbox}}
+
+\vspace{{0.35em}}
+\begin{{minipage}}{{0.96\textwidth}}
+\footnotesize\textit{{Note.}} Each cell reports mean $\pm$ std across synthetic generators after first averaging that metric within the covered datasets of the corresponding category for each generator. \texttt{{Distance}} keeps only \texttt{{Dist. overall}}, while the five query-centric families retain only their sub-items. The category rows are ranked within each column and highlighted as {{\color{{FirstPlace}}\textbf{{First}}}}, {{\color{{SecondPlace}}\textbf{{Second}}}}, and {{\color{{ThirdPlace}}\textbf{{Third}}}}.
+\end{{minipage}}
+}}
+"""
+
+
+def render_standalone_latex(summary: pd.DataFrame) -> str:
+    embedded = render_embedded_source(summary)
+    return rf"""\documentclass[10pt]{{article}}
+\usepackage[a4paper,landscape,margin=0.60in]{{geometry}}
+\usepackage[T1]{{fontenc}}
+\usepackage[utf8]{{inputenc}}
+\usepackage{{newtxtext,newtxmath}}
+\usepackage{{booktabs}}
+\usepackage[table]{{xcolor}}
+\usepackage{{array}}
+\usepackage{{multirow}}
+\usepackage{{adjustbox}}
+\usepackage{{caption}}
+\usepackage{{microtype}}
+\captionsetup{{font=small,labelfont=bf}}
+
+\begin{{document}}
+\thispagestyle{{empty}}
+
+\noindent{{\small\textit{{Category-wise benchmark summary with expanded query families}}}}\\[-0.15em]
+\noindent\color{{gray}}\rule{{\textwidth}}{{0.5pt}}
+
+\begin{{table}}[ht]
+\centering
+\caption{{Category-wise benchmark summary aligned to the benchmark reference-table style. \texttt{{Distance}} retains only \texttt{{Dist. overall}}. The five query-centric families (\texttt{{Subgroup}}, \texttt{{Conditional}}, \texttt{{Tail}}, \texttt{{Missingness}}, and \texttt{{Cardinality}}) retain only their sub-items. Rows are the three dataset categories, and each cell is a generator-averaged category score shown as mean $\pm$ std. Within each column, category ranks are marked as \textcolor[HTML]{{1397B8}}{{\textbf{{First}}}}, \textcolor[HTML]{{7B45E5}}{{\textbf{{Second}}}}, and \textcolor[HTML]{{000000}}{{\textbf{{Third}}}}.}}
+\label{{tab:benchmark_query_category_real}}
+{embedded}
+\end{{table}}
+
+\end{{document}}
+"""
+
+
+def build_source_manifest() -> pd.DataFrame:
+    return pd.DataFrame(
+        [
+            {
+                "metric_key": metric["key"],
+                "metric_family": metric["family"],
+                "metric_header_label": metric["header_label"],
+                "metric_full_label": metric["full_label"],
+                "source_kind": metric["source_kind"],
+                "source_file": metric["source_file"],
+                "source_note": metric["source_note"],
+            }
+            for metric in METRICS
+        ]
+    )
+
+
+def try_compile_pdf(tex_path: Path) -> tuple[bool, str]:
+    commands = []
+    bundled_tectonic = REPO_ROOT / "Evaluation" / "model_radar" / "_build_tools" / "tectonic" / "tectonic.exe"
+    if shutil.which("latexmk"):
+        commands.append(["latexmk", "-pdf", "-interaction=nonstopmode", tex_path.name])
+    if shutil.which("tectonic"):
+        commands.append(["tectonic", tex_path.name])
+    elif bundled_tectonic.exists():
+        commands.append([str(bundled_tectonic), tex_path.name])
+
+    last_message = "No local TeX engine was found."
+    for command in commands:
+        try:
+            proc = subprocess.run(
+                command,
+                cwd=tex_path.parent,
+                stdout=subprocess.PIPE,
+                stderr=subprocess.STDOUT,
+                text=True,
+                check=False,
+            )
+        except Exception as exc:
+            last_message = str(exc)
+            continue
+        if proc.returncode == 0:
+            return True, proc.stdout[-2000:]
+        last_message = proc.stdout[-2000:]
+    return False, last_message
+
+
+def try_render_png(pdf_path: Path, png_path: Path) -> tuple[bool, str]:
+    try:
+        import fitz  # type: ignore
+
+        doc = fitz.open(pdf_path)
+        page = doc.load_page(0)
+        pix = page.get_pixmap(matrix=fitz.Matrix(2.0, 2.0), alpha=False)
+        pix.save(png_path)
+        return True, "Rendered with PyMuPDF."
+    except Exception as exc:
+        last_message = str(exc)
+
+    if shutil.which("pdftoppm"):
+        prefix = png_path.with_suffix("")
+        try:
+            proc = subprocess.run(
+                ["pdftoppm", "-png", "-singlefile", str(pdf_path), str(prefix)],
+                cwd=pdf_path.parent,
+                stdout=subprocess.PIPE,
+                stderr=subprocess.STDOUT,
+                text=True,
+                check=False,
+            )
+            if proc.returncode == 0 and png_path.exists():
+                return True, proc.stdout[-1000:]
+            last_message = proc.stdout[-1000:]
+        except Exception as exc:
+            last_message = str(exc)
+
+    if shutil.which("magick"):
+        try:
+            proc = subprocess.run(
+                ["magick", "-density", "220", str(pdf_path), str(png_path)],
+                cwd=pdf_path.parent,
+                stdout=subprocess.PIPE,
+                stderr=subprocess.STDOUT,
+                text=True,
+                check=False,
+            )
+            if proc.returncode == 0 and png_path.exists():
+                return True, proc.stdout[-1000:]
+            last_message = proc.stdout[-1000:]
+        except Exception as exc:
+            last_message = str(exc)
+
+    return False, last_message
+
+
+def write_outputs(dataset_level: pd.DataFrame, summary: pd.DataFrame) -> dict[str, Path]:
+    ensure_out_dir()
+    output_paths = {
+        "dataset_csv": OUT_DIR / f"{FINAL_BASENAME}_dataset_level.csv",
+        "summary_csv": OUT_DIR / f"{FINAL_BASENAME}_summary.csv",
+        "legacy_summary_csv": OUT_DIR / f"{FINAL_BASENAME}_model_summary.csv",
+        "sources_csv": OUT_DIR / f"{FINAL_BASENAME}_sources.csv",
+        "tex": OUT_DIR / f"{FINAL_BASENAME}.tex",
+        "pdf": OUT_DIR / f"{FINAL_BASENAME}.pdf",
+        "png": OUT_DIR / f"{FINAL_BASENAME}.png",
+        "source_tex": OUT_DIR / f"{FINAL_BASENAME}_source.tex",
+    }
+
+    dataset_level.to_csv(output_paths["dataset_csv"], index=False)
+    summary.to_csv(output_paths["summary_csv"], index=False)
+    summary.to_csv(output_paths["legacy_summary_csv"], index=False)
+    build_source_manifest().to_csv(output_paths["sources_csv"], index=False)
+    output_paths["tex"].write_text(render_standalone_latex(summary), encoding="utf-8")
+    output_paths["source_tex"].write_text(render_embedded_source(summary), encoding="utf-8")
+    return output_paths
+
+
+def sync_paper_source(source_tex_path: Path) -> Path:
+    paper_dir = pick_active_paper_dir()
+    target_dir = paper_dir / PAPER_FIGURE_SUBDIR
+    target_dir.mkdir(parents=True, exist_ok=True)
+    target_path = target_dir / source_tex_path.name
+    target_path.write_text(source_tex_path.read_text(encoding="utf-8"), encoding="utf-8")
+    return target_path
+
+
+def main() -> None:
+    dataset_level = assemble_dataset_level_table()
+    summary = build_category_summary(dataset_level)
+    output_paths = write_outputs(dataset_level, summary)
+    paper_target = sync_paper_source(output_paths["source_tex"])
+
+    compiled, compile_note = try_compile_pdf(output_paths["tex"])
+    if compiled:
+        print(f"Compiled PDF: {output_paths['pdf']}")
+        rendered, render_note = try_render_png(output_paths["pdf"], output_paths["png"])
+        if rendered:
+            print(f"Rendered PNG: {output_paths['png']}")
+        else:
+            print(f"PNG rendering skipped/failed: {render_note}")
+    else:
+        print(f"PDF compilation skipped/failed: {compile_note}")
+
+    print(f"Wrote dataset-level CSV: {output_paths['dataset_csv']}")
+    print(f"Wrote summary CSV: {output_paths['summary_csv']}")
+    print(f"Wrote source manifest: {output_paths['sources_csv']}")
+    print(f"Wrote embedded source snippet: {paper_target}")
+
+
+if __name__ == "__main__":
+    main()

evaluation/tables/benchmark_query_category_table/final/README.md

@@ -0,0 +1,11 @@
+# Final Assets
+
+- `benchmark_query_category_table_real.tex`: standalone LaTeX source for the review table
+- `benchmark_query_category_table_real.pdf`: compiled review PDF when a local TeX engine is available
+- `benchmark_query_category_table_real.png`: rendered one-page preview when PDF-to-image conversion is available
+- `benchmark_query_category_table_real_source.tex`: paper-facing embedded table snippet
+- `benchmark_query_category_table_real_summary.csv`: category-level generator-averaged means, stds, and model counts
+- `benchmark_query_category_table_real_dataset_level.csv`: dataset-level metric values used for aggregation
+- `benchmark_query_category_table_real_sources.csv`: source manifest for each displayed metric
+
+This table keeps only six columns: `Distance`, `Subgroup`, `Conditional`, `Tail`, `Missingness`, and `Cardinality`. Rows are `Categorical`, `Numerical`, and `Mix`, and each cell is the arithmetic mean across the available synthetic generators for that category/metric combination.

evaluation/tables/benchmark_query_category_table/final/benchmark_query_category_table_real.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c05cce08671806ea82d464f556f4717fb6c2e9e08c949ec4c01c0eac63c8c017
+size 28662

evaluation/tables/benchmark_query_category_table/final/benchmark_query_category_table_real.tex

@@ -0,0 +1,62 @@
+\documentclass[10pt]{article}
+\usepackage[a4paper,landscape,margin=0.60in]{geometry}
+\usepackage[T1]{fontenc}
+\usepackage[utf8]{inputenc}
+\usepackage{newtxtext,newtxmath}
+\usepackage{booktabs}
+\usepackage[table]{xcolor}
+\usepackage{array}
+\usepackage{multirow}
+\usepackage{adjustbox}
+\usepackage{caption}
+\usepackage{microtype}
+\captionsetup{font=small,labelfont=bf}
+
+\begin{document}
+\thispagestyle{empty}
+
+\noindent{\small\textit{Category-wise benchmark summary with expanded query families}}\\[-0.15em]
+\noindent\color{gray}\rule{\textwidth}{0.5pt}
+
+\begin{table}[ht]
+\centering
+\caption{Category-wise benchmark summary aligned to the benchmark reference-table style. \texttt{Distance} retains only \texttt{Dist. overall}. The five query-centric families (\texttt{Subgroup}, \texttt{Conditional}, \texttt{Tail}, \texttt{Missingness}, and \texttt{Cardinality}) retain only their sub-items. Rows are the three dataset categories, and each cell is a generator-averaged category score shown as mean $\pm$ std. Within each column, category ranks are marked as \textcolor[HTML]{1397B8}{\textbf{First}}, \textcolor[HTML]{7B45E5}{\textbf{Second}}, and \textcolor[HTML]{000000}{\textbf{Third}}.}
+\label{tab:benchmark_query_category_real}
+{
+\definecolor{FirstPlace}{HTML}{1397B8}%
+\definecolor{SecondPlace}{HTML}{7B45E5}%
+\definecolor{ThirdPlace}{HTML}{000000}%
+\definecolor{OverallTint}{HTML}{F8F1DA}%
+\definecolor{RuleGray}{HTML}{C8CDD3}%
+\arrayrulecolor{RuleGray}%
+\setlength{\tabcolsep}{3.0pt}%
+\renewcommand{\arraystretch}{1.08}%
+\scriptsize
+\begin{adjustbox}{max width=\textwidth}
+\begin{tabular}{@{}l c c c c c c c c c c c c c@{}}
+\toprule
+\multirow{2}{*}{\textbf{Category}} & \multicolumn{1}{c}{\textbf{Distance}} & \multicolumn{2}{c}{\textbf{Subgroup}} & \multicolumn{3}{c}{\textbf{Conditional}} & \multicolumn{3}{c}{\textbf{Tail}} & \multicolumn{2}{c}{\textbf{Missingness}} & \multicolumn{2}{c}{\textbf{Cardinality}} \\
+\cmidrule(lr){2-2}
+\cmidrule(lr){3-4}
+\cmidrule(lr){5-7}
+\cmidrule(lr){8-10}
+\cmidrule(lr){11-12}
+\cmidrule(lr){13-14}
+ & \cellcolor{OverallTint} Dist. overall $\uparrow$ & Internal profile $\uparrow$ & Size stability $\uparrow$ & Dependency strength $\uparrow$ & Direction $\uparrow$ & Slice-level $\uparrow$ & Tail set $\uparrow$ & Tail mass $\uparrow$ & Tail concentration $\uparrow$ & Marginal rate $\uparrow$ & Co-missingness $\uparrow$ & Support-rank $\uparrow$ & High-cardinality $\uparrow$ \\
+\midrule
+Categorical & \cellcolor{OverallTint} {\color{ThirdPlace}\textbf{0.74$_{\pm 0.16}$}} & {\color{FirstPlace}\textbf{0.47$_{\pm 0.07}$}} & {\color{SecondPlace}\textbf{0.40$_{\pm 0.08}$}} & {\color{FirstPlace}\textbf{0.53$_{\pm 0.10}$}} & {\color{FirstPlace}\textbf{0.54$_{\pm 0.11}$}} & {\color{SecondPlace}\textbf{0.39$_{\pm 0.09}$}} & {\color{FirstPlace}\textbf{0.14$_{\pm 0.12}$}} & {\color{SecondPlace}\textbf{0.31$_{\pm 0.24}$}} & {\color{SecondPlace}\textbf{0.37$_{\pm 0.26}$}} & {\color{FirstPlace}\textbf{0.88$_{\pm 0.07}$}} & {\color{FirstPlace}\textbf{0.88$_{\pm 0.07}$}} & {\color{SecondPlace}\textbf{0.46$_{\pm 0.37}$}} & {\color{ThirdPlace}\textbf{0.66$_{\pm 0.15}$}} \\
+Numerical & \cellcolor{OverallTint} {\color{SecondPlace}\textbf{0.78$_{\pm 0.12}$}} & {\color{ThirdPlace}\textbf{0.32$_{\pm 0.08}$}} & {\color{ThirdPlace}\textbf{0.37$_{\pm 0.11}$}} & {\color{ThirdPlace}\textbf{0.35$_{\pm 0.14}$}} & {\color{ThirdPlace}\textbf{0.38$_{\pm 0.13}$}} & {\color{ThirdPlace}\textbf{0.37$_{\pm 0.05}$}} & {\color{ThirdPlace}\textbf{0.07$_{\pm 0.07}$}} & {\color{FirstPlace}\textbf{0.60$_{\pm 0.11}$}} & {\color{FirstPlace}\textbf{0.67$_{\pm 0.07}$}} & {\color{ThirdPlace}\textbf{0.55$_{\pm 0.25}$}} & {\color{ThirdPlace}\textbf{0.55$_{\pm 0.25}$}} & {\color{ThirdPlace}\textbf{0.39$_{\pm 0.28}$}} & {\color{FirstPlace}\textbf{0.88$_{\pm 0.09}$}} \\
+Mix & \cellcolor{OverallTint} {\color{FirstPlace}\textbf{0.86$_{\pm 0.11}$}} & {\color{SecondPlace}\textbf{0.40$_{\pm 0.06}$}} & {\color{FirstPlace}\textbf{0.46$_{\pm 0.06}$}} & {\color{SecondPlace}\textbf{0.51$_{\pm 0.07}$}} & {\color{SecondPlace}\textbf{0.51$_{\pm 0.08}$}} & {\color{FirstPlace}\textbf{0.47$_{\pm 0.05}$}} & {\color{SecondPlace}\textbf{0.10$_{\pm 0.12}$}} & {\color{ThirdPlace}\textbf{0.23$_{\pm 0.18}$}} & {\color{ThirdPlace}\textbf{0.31$_{\pm 0.20}$}} & {\color{SecondPlace}\textbf{0.84$_{\pm 0.11}$}} & {\color{SecondPlace}\textbf{0.84$_{\pm 0.11}$}} & {\color{FirstPlace}\textbf{0.48$_{\pm 0.36}$}} & {\color{SecondPlace}\textbf{0.80$_{\pm 0.09}$}} \\
+\bottomrule
+\end{tabular}
+\end{adjustbox}
+
+\vspace{0.35em}
+\begin{minipage}{0.96\textwidth}
+\footnotesize\textit{Note.} Each cell reports mean $\pm$ std across synthetic generators after first averaging that metric within the covered datasets of the corresponding category for each generator. \texttt{Distance} keeps only \texttt{Dist. overall}, while the five query-centric families retain only their sub-items. The category rows are ranked within each column and highlighted as {\color{FirstPlace}\textbf{First}}, {\color{SecondPlace}\textbf{Second}}, and {\color{ThirdPlace}\textbf{Third}}.
+\end{minipage}
+}
+
+\end{table}
+
+\end{document}

evaluation/tables/benchmark_query_category_table/final/benchmark_query_category_table_real_model_summary.csv

@@ -0,0 +1,40 @@
+dataset_prefix,dataset_prefix_label,metric_key,metric_family,metric_group_label,metric_header_label,metric_full_label,category_mean,category_std,model_count,min_covered_datasets,max_covered_datasets,prefix_order,metric_order,rank
+c,Categorical,distance_overall,Distance,Distance,Dist. overall,Distance overall,0.7378319956864181,0.15882813583964822,11,8,20,0,0,3
+c,Categorical,internal_profile_stability,Subgroup,Subgroup,Internal profile,Internal profile stability,0.4695235274722863,0.06998974703211315,11,8,19,0,1,1
+c,Categorical,subgroup_size_stability,Subgroup,Subgroup,Size stability,Subgroup size stability,0.398626941876227,0.07623829611451803,11,8,19,0,2,2
+c,Categorical,dependency_strength_similarity,Conditional,Conditional,Dependency strength,Dependency strength similarity,0.5297446115302926,0.10311066073871228,11,8,19,0,3,1
+c,Categorical,direction_consistency,Conditional,Conditional,Direction,Direction consistency,0.5362975178302296,0.10704072139884171,11,8,19,0,4,1
+c,Categorical,slice_level_consistency,Conditional,Conditional,Slice-level,Slice-level consistency,0.38875129467250535,0.08983283734580416,11,7,18,0,5,2
+c,Categorical,tail_set_consistency,Tail,Tail,Tail set,Tail set consistency,0.14253322839467797,0.12341820140990105,11,8,18,0,6,1
+c,Categorical,tail_mass_similarity,Tail,Tail,Tail mass,Tail mass similarity,0.31028592376298464,0.23675650633035966,11,8,18,0,7,2
+c,Categorical,tail_concentration_consistency,Tail,Tail,Tail concentration,Tail concentration consistency,0.3740018230680052,0.26482213384488884,11,8,18,0,8,2
+c,Categorical,marginal_missing_rate_consistency,Missingness,Missingness,Marginal rate,Marginal missing-rate consistency,0.8813004799999999,0.06970819812648324,10,8,19,0,9,1
+c,Categorical,co_missingness_pattern_consistency,Missingness,Missingness,Co-missingness,Co-missingness pattern consistency,0.8783717266666666,0.0666524265510059,10,8,19,0,10,1
+c,Categorical,support_rank_profile_consistency,Cardinality,Cardinality,Support-rank,Support-rank profile consistency,0.4602752186068887,0.366861408505073,11,8,20,0,11,2
+c,Categorical,high_cardinality_response_stability,Cardinality,Cardinality,High-cardinality,High-cardinality response stability,0.6620046773285082,0.14963038372953083,11,1,12,0,12,3
+n,Numerical,distance_overall,Distance,Distance,Dist. overall,Distance overall,0.7781988849850697,0.11701571515037142,11,2,20,1,0,2
+n,Numerical,internal_profile_stability,Subgroup,Subgroup,Internal profile,Internal profile stability,0.32300778920404777,0.08227942926217155,11,2,19,1,1,3
+n,Numerical,subgroup_size_stability,Subgroup,Subgroup,Size stability,Subgroup size stability,0.36957429421192695,0.11325409439634704,11,1,16,1,2,3
+n,Numerical,dependency_strength_similarity,Conditional,Conditional,Dependency strength,Dependency strength similarity,0.34568107721322067,0.13836647733257204,11,2,19,1,3,3
+n,Numerical,direction_consistency,Conditional,Conditional,Direction,Direction consistency,0.38364716916321734,0.13041214216436803,11,2,19,1,4,3
+n,Numerical,slice_level_consistency,Conditional,Conditional,Slice-level,Slice-level consistency,0.36704052493014905,0.053452389974622545,11,2,19,1,5,3
+n,Numerical,tail_set_consistency,Tail,Tail,Tail set,Tail set consistency,0.07464324919104746,0.0706525986620942,10,2,19,1,6,3
+n,Numerical,tail_mass_similarity,Tail,Tail,Tail mass,Tail mass similarity,0.5977028937383901,0.11297410538607572,10,2,19,1,7,1
+n,Numerical,tail_concentration_consistency,Tail,Tail,Tail concentration,Tail concentration consistency,0.6691425957667698,0.07388882174370075,10,2,19,1,8,1
+n,Numerical,marginal_missing_rate_consistency,Missingness,Missingness,Marginal rate,Marginal missing-rate consistency,0.5514571851851852,0.2512177564827895,9,2,19,1,9,3
+n,Numerical,co_missingness_pattern_consistency,Missingness,Missingness,Co-missingness,Co-missingness pattern consistency,0.5521162222222222,0.249249116869312,9,2,19,1,10,3
+n,Numerical,support_rank_profile_consistency,Cardinality,Cardinality,Support-rank,Support-rank profile consistency,0.3945733820275187,0.28447649711307266,10,2,19,1,11,3
+n,Numerical,high_cardinality_response_stability,Cardinality,Cardinality,High-cardinality,High-cardinality response stability,0.8777668822001493,0.09134254610672132,10,2,14,1,12,1
+m,Mix,distance_overall,Distance,Distance,Dist. overall,Distance overall,0.8563024788267684,0.10559535058410295,11,6,11,2,0,1
+m,Mix,internal_profile_stability,Subgroup,Subgroup,Internal profile,Internal profile stability,0.3959704766595662,0.05525315390211962,11,6,11,2,1,2
+m,Mix,subgroup_size_stability,Subgroup,Subgroup,Size stability,Subgroup size stability,0.4604416994197091,0.06456896618839744,11,6,11,2,2,1
+m,Mix,dependency_strength_similarity,Conditional,Conditional,Dependency strength,Dependency strength similarity,0.5092219044921619,0.06698324290685187,11,6,11,2,3,2
+m,Mix,direction_consistency,Conditional,Conditional,Direction,Direction consistency,0.5051151576853998,0.07874618397987326,11,6,11,2,4,2
+m,Mix,slice_level_consistency,Conditional,Conditional,Slice-level,Slice-level consistency,0.4712745279835051,0.0537725139680299,11,6,11,2,5,1
+m,Mix,tail_set_consistency,Tail,Tail,Tail set,Tail set consistency,0.09766786832185491,0.12306172722651405,11,6,11,2,6,2
+m,Mix,tail_mass_similarity,Tail,Tail,Tail mass,Tail mass similarity,0.22915647893841007,0.18489994930514467,11,6,11,2,7,3
+m,Mix,tail_concentration_consistency,Tail,Tail,Tail concentration,Tail concentration consistency,0.3105914506349206,0.202725881012202,11,6,11,2,8,3
+m,Mix,marginal_missing_rate_consistency,Missingness,Missingness,Marginal rate,Marginal missing-rate consistency,0.8425754999999999,0.10734148666154718,11,6,11,2,9,2
+m,Mix,co_missingness_pattern_consistency,Missingness,Missingness,Co-missingness,Co-missingness pattern consistency,0.8413917727272726,0.10844381859544887,11,6,11,2,10,2
+m,Mix,support_rank_profile_consistency,Cardinality,Cardinality,Support-rank,Support-rank profile consistency,0.4813761323423284,0.36480806865690174,11,6,11,2,11,1
+m,Mix,high_cardinality_response_stability,Cardinality,Cardinality,High-cardinality,High-cardinality response stability,0.7957809127934102,0.09392061292496333,11,5,10,2,12,2

evaluation/tables/benchmark_query_category_table/final/benchmark_query_category_table_real_source.tex

@@ -0,0 +1,34 @@
+{
+\definecolor{FirstPlace}{HTML}{1397B8}%
+\definecolor{SecondPlace}{HTML}{7B45E5}%
+\definecolor{ThirdPlace}{HTML}{000000}%
+\definecolor{OverallTint}{HTML}{F8F1DA}%
+\definecolor{RuleGray}{HTML}{C8CDD3}%
+\arrayrulecolor{RuleGray}%
+\setlength{\tabcolsep}{3.0pt}%
+\renewcommand{\arraystretch}{1.08}%
+\scriptsize
+\begin{adjustbox}{max width=\textwidth}
+\begin{tabular}{@{}l c c c c c c c c c c c c c@{}}
+\toprule
+\multirow{2}{*}{\textbf{Category}} & \multicolumn{1}{c}{\textbf{Distance}} & \multicolumn{2}{c}{\textbf{Subgroup}} & \multicolumn{3}{c}{\textbf{Conditional}} & \multicolumn{3}{c}{\textbf{Tail}} & \multicolumn{2}{c}{\textbf{Missingness}} & \multicolumn{2}{c}{\textbf{Cardinality}} \\
+\cmidrule(lr){2-2}
+\cmidrule(lr){3-4}
+\cmidrule(lr){5-7}
+\cmidrule(lr){8-10}
+\cmidrule(lr){11-12}
+\cmidrule(lr){13-14}
+ & \cellcolor{OverallTint} Dist. overall $\uparrow$ & Internal profile $\uparrow$ & Size stability $\uparrow$ & Dependency strength $\uparrow$ & Direction $\uparrow$ & Slice-level $\uparrow$ & Tail set $\uparrow$ & Tail mass $\uparrow$ & Tail concentration $\uparrow$ & Marginal rate $\uparrow$ & Co-missingness $\uparrow$ & Support-rank $\uparrow$ & High-cardinality $\uparrow$ \\
+\midrule
+Categorical & \cellcolor{OverallTint} {\color{ThirdPlace}\textbf{0.74$_{\pm 0.16}$}} & {\color{FirstPlace}\textbf{0.47$_{\pm 0.07}$}} & {\color{SecondPlace}\textbf{0.40$_{\pm 0.08}$}} & {\color{FirstPlace}\textbf{0.53$_{\pm 0.10}$}} & {\color{FirstPlace}\textbf{0.54$_{\pm 0.11}$}} & {\color{SecondPlace}\textbf{0.39$_{\pm 0.09}$}} & {\color{FirstPlace}\textbf{0.14$_{\pm 0.12}$}} & {\color{SecondPlace}\textbf{0.31$_{\pm 0.24}$}} & {\color{SecondPlace}\textbf{0.37$_{\pm 0.26}$}} & {\color{FirstPlace}\textbf{0.88$_{\pm 0.07}$}} & {\color{FirstPlace}\textbf{0.88$_{\pm 0.07}$}} & {\color{SecondPlace}\textbf{0.46$_{\pm 0.37}$}} & {\color{ThirdPlace}\textbf{0.66$_{\pm 0.15}$}} \\
+Numerical & \cellcolor{OverallTint} {\color{SecondPlace}\textbf{0.78$_{\pm 0.12}$}} & {\color{ThirdPlace}\textbf{0.32$_{\pm 0.08}$}} & {\color{ThirdPlace}\textbf{0.37$_{\pm 0.11}$}} & {\color{ThirdPlace}\textbf{0.35$_{\pm 0.14}$}} & {\color{ThirdPlace}\textbf{0.38$_{\pm 0.13}$}} & {\color{ThirdPlace}\textbf{0.37$_{\pm 0.05}$}} & {\color{ThirdPlace}\textbf{0.07$_{\pm 0.07}$}} & {\color{FirstPlace}\textbf{0.60$_{\pm 0.11}$}} & {\color{FirstPlace}\textbf{0.67$_{\pm 0.07}$}} & {\color{ThirdPlace}\textbf{0.55$_{\pm 0.25}$}} & {\color{ThirdPlace}\textbf{0.55$_{\pm 0.25}$}} & {\color{ThirdPlace}\textbf{0.39$_{\pm 0.28}$}} & {\color{FirstPlace}\textbf{0.88$_{\pm 0.09}$}} \\
+Mix & \cellcolor{OverallTint} {\color{FirstPlace}\textbf{0.86$_{\pm 0.11}$}} & {\color{SecondPlace}\textbf{0.40$_{\pm 0.06}$}} & {\color{FirstPlace}\textbf{0.46$_{\pm 0.06}$}} & {\color{SecondPlace}\textbf{0.51$_{\pm 0.07}$}} & {\color{SecondPlace}\textbf{0.51$_{\pm 0.08}$}} & {\color{FirstPlace}\textbf{0.47$_{\pm 0.05}$}} & {\color{SecondPlace}\textbf{0.10$_{\pm 0.12}$}} & {\color{ThirdPlace}\textbf{0.23$_{\pm 0.18}$}} & {\color{ThirdPlace}\textbf{0.31$_{\pm 0.20}$}} & {\color{SecondPlace}\textbf{0.84$_{\pm 0.11}$}} & {\color{SecondPlace}\textbf{0.84$_{\pm 0.11}$}} & {\color{FirstPlace}\textbf{0.48$_{\pm 0.36}$}} & {\color{SecondPlace}\textbf{0.80$_{\pm 0.09}$}} \\
+\bottomrule
+\end{tabular}
+\end{adjustbox}
+
+\vspace{0.35em}
+\begin{minipage}{0.96\textwidth}
+\footnotesize\textit{Note.} Each cell reports mean $\pm$ std across synthetic generators after first averaging that metric within the covered datasets of the corresponding category for each generator. \texttt{Distance} keeps only \texttt{Dist. overall}, while the five query-centric families retain only their sub-items. The category rows are ranked within each column and highlighted as {\color{FirstPlace}\textbf{First}}, {\color{SecondPlace}\textbf{Second}}, and {\color{ThirdPlace}\textbf{Third}}.
+\end{minipage}
+}

evaluation/tables/benchmark_query_category_table/final/benchmark_query_category_table_real_sources.csv

@@ -0,0 +1,14 @@
+metric_key,metric_family,metric_header_label,metric_full_label,source_kind,source_file,source_note
+distance_overall,Distance,Dist. overall,Distance overall,distance_dataset_export,Evaluation/distance/runs/*/datasets/*/distance_summary__*.csv,Latest deduplicated dataset-model overall fidelity score.
+internal_profile_stability,Subgroup,Internal profile,Internal profile stability,query_subitem_dataset_export,Evaluation/query_fivepart_breakdown/subgroup_breakdown/data/dataset_model_subitems.csv,Canonical subgroup sub-item score from dataset-level sub-item export.
+subgroup_size_stability,Subgroup,Size stability,Subgroup size stability,query_subitem_dataset_export,Evaluation/query_fivepart_breakdown/subgroup_breakdown/data/dataset_model_subitems.csv,Canonical subgroup sub-item score from dataset-level sub-item export.
+dependency_strength_similarity,Conditional,Dependency strength,Dependency strength similarity,query_subitem_dataset_export,Evaluation/query_fivepart_breakdown/conditional_breakdown/data/dataset_model_subitems.csv,Canonical conditional sub-item score from dataset-level sub-item export.
+direction_consistency,Conditional,Direction,Direction consistency,query_subitem_dataset_export,Evaluation/query_fivepart_breakdown/conditional_breakdown/data/dataset_model_subitems.csv,Canonical conditional sub-item score from dataset-level sub-item export.
+slice_level_consistency,Conditional,Slice-level,Slice-level consistency,query_subitem_dataset_export,Evaluation/query_fivepart_breakdown/conditional_breakdown/data/dataset_model_subitems.csv,Canonical conditional sub-item score from dataset-level sub-item export.
+tail_set_consistency,Tail,Tail set,Tail set consistency,query_subitem_dataset_export,Evaluation/query_fivepart_breakdown/tail_breakdown/data/dataset_model_scores.csv,Canonical tail sub-item score from dataset-level breakdown export.
+tail_mass_similarity,Tail,Tail mass,Tail mass similarity,query_subitem_dataset_export,Evaluation/query_fivepart_breakdown/tail_breakdown/data/dataset_model_scores.csv,Canonical tail sub-item score from dataset-level breakdown export.
+tail_concentration_consistency,Tail,Tail concentration,Tail concentration consistency,query_subitem_dataset_export,Evaluation/query_fivepart_breakdown/tail_breakdown/data/dataset_model_scores.csv,Canonical tail sub-item score from dataset-level breakdown export.
+marginal_missing_rate_consistency,Missingness,Marginal rate,Marginal missing-rate consistency,query_subitem_dataset_export,Evaluation/query_fivepart_breakdown/missingness_breakdown/data/dataset_model_subitems.csv,Canonical missingness sub-item score from dataset-level sub-item export.
+co_missingness_pattern_consistency,Missingness,Co-missingness,Co-missingness pattern consistency,query_subitem_dataset_export,Evaluation/query_fivepart_breakdown/missingness_breakdown/data/dataset_model_subitems.csv,Canonical missingness sub-item score from dataset-level sub-item export.
+support_rank_profile_consistency,Cardinality,Support-rank,Support-rank profile consistency,cardinality_subitem_dataset_export,Evaluation/query_fivepart_breakdown/cardinality/data/cleaned_results.csv,Dataset-level mean of the discrete support-retention branch.
+high_cardinality_response_stability,Cardinality,High-cardinality,High-cardinality response stability,cardinality_subitem_dataset_export,Evaluation/query_fivepart_breakdown/cardinality/data/cleaned_results.csv,Dataset-level mean over high-dynamic cardinality/range units.

evaluation/tables/benchmark_query_category_table/final/benchmark_query_category_table_real_summary.csv

@@ -0,0 +1,40 @@
+dataset_prefix,dataset_prefix_label,metric_key,metric_family,metric_group_label,metric_header_label,metric_full_label,category_mean,category_std,model_count,min_covered_datasets,max_covered_datasets,prefix_order,metric_order,rank
+c,Categorical,distance_overall,Distance,Distance,Dist. overall,Distance overall,0.7378319956864181,0.15882813583964822,11,8,20,0,0,3
+c,Categorical,internal_profile_stability,Subgroup,Subgroup,Internal profile,Internal profile stability,0.4695235274722863,0.06998974703211315,11,8,19,0,1,1
+c,Categorical,subgroup_size_stability,Subgroup,Subgroup,Size stability,Subgroup size stability,0.398626941876227,0.07623829611451803,11,8,19,0,2,2
+c,Categorical,dependency_strength_similarity,Conditional,Conditional,Dependency strength,Dependency strength similarity,0.5297446115302926,0.10311066073871228,11,8,19,0,3,1
+c,Categorical,direction_consistency,Conditional,Conditional,Direction,Direction consistency,0.5362975178302296,0.10704072139884171,11,8,19,0,4,1
+c,Categorical,slice_level_consistency,Conditional,Conditional,Slice-level,Slice-level consistency,0.38875129467250535,0.08983283734580416,11,7,18,0,5,2
+c,Categorical,tail_set_consistency,Tail,Tail,Tail set,Tail set consistency,0.14253322839467797,0.12341820140990105,11,8,18,0,6,1
+c,Categorical,tail_mass_similarity,Tail,Tail,Tail mass,Tail mass similarity,0.31028592376298464,0.23675650633035966,11,8,18,0,7,2
+c,Categorical,tail_concentration_consistency,Tail,Tail,Tail concentration,Tail concentration consistency,0.3740018230680052,0.26482213384488884,11,8,18,0,8,2
+c,Categorical,marginal_missing_rate_consistency,Missingness,Missingness,Marginal rate,Marginal missing-rate consistency,0.8813004799999999,0.06970819812648324,10,8,19,0,9,1
+c,Categorical,co_missingness_pattern_consistency,Missingness,Missingness,Co-missingness,Co-missingness pattern consistency,0.8783717266666666,0.0666524265510059,10,8,19,0,10,1
+c,Categorical,support_rank_profile_consistency,Cardinality,Cardinality,Support-rank,Support-rank profile consistency,0.4602752186068887,0.366861408505073,11,8,20,0,11,2
+c,Categorical,high_cardinality_response_stability,Cardinality,Cardinality,High-cardinality,High-cardinality response stability,0.6620046773285082,0.14963038372953083,11,1,12,0,12,3
+n,Numerical,distance_overall,Distance,Distance,Dist. overall,Distance overall,0.7781988849850697,0.11701571515037142,11,2,20,1,0,2
+n,Numerical,internal_profile_stability,Subgroup,Subgroup,Internal profile,Internal profile stability,0.32300778920404777,0.08227942926217155,11,2,19,1,1,3
+n,Numerical,subgroup_size_stability,Subgroup,Subgroup,Size stability,Subgroup size stability,0.36957429421192695,0.11325409439634704,11,1,16,1,2,3
+n,Numerical,dependency_strength_similarity,Conditional,Conditional,Dependency strength,Dependency strength similarity,0.34568107721322067,0.13836647733257204,11,2,19,1,3,3
+n,Numerical,direction_consistency,Conditional,Conditional,Direction,Direction consistency,0.38364716916321734,0.13041214216436803,11,2,19,1,4,3
+n,Numerical,slice_level_consistency,Conditional,Conditional,Slice-level,Slice-level consistency,0.36704052493014905,0.053452389974622545,11,2,19,1,5,3
+n,Numerical,tail_set_consistency,Tail,Tail,Tail set,Tail set consistency,0.07464324919104746,0.0706525986620942,10,2,19,1,6,3
+n,Numerical,tail_mass_similarity,Tail,Tail,Tail mass,Tail mass similarity,0.5977028937383901,0.11297410538607572,10,2,19,1,7,1
+n,Numerical,tail_concentration_consistency,Tail,Tail,Tail concentration,Tail concentration consistency,0.6691425957667698,0.07388882174370075,10,2,19,1,8,1
+n,Numerical,marginal_missing_rate_consistency,Missingness,Missingness,Marginal rate,Marginal missing-rate consistency,0.5514571851851852,0.2512177564827895,9,2,19,1,9,3
+n,Numerical,co_missingness_pattern_consistency,Missingness,Missingness,Co-missingness,Co-missingness pattern consistency,0.5521162222222222,0.249249116869312,9,2,19,1,10,3
+n,Numerical,support_rank_profile_consistency,Cardinality,Cardinality,Support-rank,Support-rank profile consistency,0.3945733820275187,0.28447649711307266,10,2,19,1,11,3
+n,Numerical,high_cardinality_response_stability,Cardinality,Cardinality,High-cardinality,High-cardinality response stability,0.8777668822001493,0.09134254610672132,10,2,14,1,12,1
+m,Mix,distance_overall,Distance,Distance,Dist. overall,Distance overall,0.8563024788267684,0.10559535058410295,11,6,11,2,0,1
+m,Mix,internal_profile_stability,Subgroup,Subgroup,Internal profile,Internal profile stability,0.3959704766595662,0.05525315390211962,11,6,11,2,1,2
+m,Mix,subgroup_size_stability,Subgroup,Subgroup,Size stability,Subgroup size stability,0.4604416994197091,0.06456896618839744,11,6,11,2,2,1
+m,Mix,dependency_strength_similarity,Conditional,Conditional,Dependency strength,Dependency strength similarity,0.5092219044921619,0.06698324290685187,11,6,11,2,3,2
+m,Mix,direction_consistency,Conditional,Conditional,Direction,Direction consistency,0.5051151576853998,0.07874618397987326,11,6,11,2,4,2
+m,Mix,slice_level_consistency,Conditional,Conditional,Slice-level,Slice-level consistency,0.4712745279835051,0.0537725139680299,11,6,11,2,5,1
+m,Mix,tail_set_consistency,Tail,Tail,Tail set,Tail set consistency,0.09766786832185491,0.12306172722651405,11,6,11,2,6,2
+m,Mix,tail_mass_similarity,Tail,Tail,Tail mass,Tail mass similarity,0.22915647893841007,0.18489994930514467,11,6,11,2,7,3
+m,Mix,tail_concentration_consistency,Tail,Tail,Tail concentration,Tail concentration consistency,0.3105914506349206,0.202725881012202,11,6,11,2,8,3
+m,Mix,marginal_missing_rate_consistency,Missingness,Missingness,Marginal rate,Marginal missing-rate consistency,0.8425754999999999,0.10734148666154718,11,6,11,2,9,2
+m,Mix,co_missingness_pattern_consistency,Missingness,Missingness,Co-missingness,Co-missingness pattern consistency,0.8413917727272726,0.10844381859544887,11,6,11,2,10,2
+m,Mix,support_rank_profile_consistency,Cardinality,Cardinality,Support-rank,Support-rank profile consistency,0.4813761323423284,0.36480806865690174,11,6,11,2,11,1
+m,Mix,high_cardinality_response_stability,Cardinality,Cardinality,High-cardinality,High-cardinality response stability,0.7957809127934102,0.09392061292496333,11,5,10,2,12,2