AINativeBench / data /processed /RQ2 /plot_total_classified_ecdf.py
王子睿
Update processed data
c4e7970 7 months ago
42.3 kB
#!/usr/bin/env python3

import csv
from collections import defaultdict
from pathlib import Path
from typing import Dict, List

import matplotlib.pyplot as plt
import numpy as np

# Use Times New Roman for all text to match other figures
plt.rcParams["font.family"] = "Times New Roman"

# Fixed model order (must match the names used in the CSV files)
MODEL_ORDER: List[str] = [
    "GPT-5",
    "GPT-4o-mini",
    "DeepSeek-V3-1",
    "DeepSeek-R1",
    "Gemini-2.5-flash",
    "Gemini-2.5-flash-nothinking",
    "Qwen3-235b",
]

# Display labels (can be slightly pretty-printed)
MODEL_LABELS: Dict[str, str] = {
    "GPT-5": "GPT-5",
    "GPT-4o-mini": "GPT-4o-mini",
    "DeepSeek-V3-1": "DeepSeek-V3.1",
    "DeepSeek-R1": "DeepSeek-R1",
    "Gemini-2.5-flash": "Gemini-2.5",
    "Gemini-2.5-flash-nothinking": "Gemini-2.5-NT",
    "Qwen3-235b": "Qwen3-235b",
}

# Color palette for different models - professional academic colors (Tableau 10 style)
MODEL_COLORS: Dict[str, str] = {
    "GPT-5": "#1f77b4",  # Blue
    "GPT-4o-mini": "#ff7f0e",  # Orange
    "DeepSeek-V3-1": "#2ca02c",  # Green
    "DeepSeek-R1": "#d62728",  # Red
    "Gemini-2.5-flash": "#9467bd",  # Purple
    "Gemini-2.5-flash-nothinking": "#8c564b",  # Brown
    "Qwen3-235b": "#e377c2",  # Pink
}


def load_total_classified(csv_path: Path) -> Dict[str, List[float]]:
    """Load total_classified values per model from a CSV file.
    Returns:
        data[model] = sorted list of total_classified values (floats).
    """

    by_model: Dict[str, List[float]] = defaultdict(list)

    with csv_path.open("r", encoding="utf-8", newline="") as f:
        reader = csv.DictReader(f)
        for row in reader:
            model = (row.get("model") or "").strip()
            if not model:
                continue
            if model not in MODEL_ORDER:
                # Ignore unknown models so that colors/order stay consistent
                continue
            val_raw = row.get("total_classified")
            if val_raw is None or val_raw == "":
                continue
            try:
                # total_classified is in milliseconds in existing CSVs
                val = float(val_raw)
            except ValueError:
                continue
            by_model[model].append(val)

    # Sort values for ECDF computation
    for m in list(by_model.keys()):
        by_model[m].sort()

    return by_model


def compute_ecdf(values: List[float]):
    """Return x, y for the empirical CDF of a 1D sample.
    x: sorted values
    y: ECDF in [0, 1]
    """

    if not values:
        return np.array([]), np.array([])

    x = np.asarray(values, dtype=float)
    n = x.size
    # Standard ECDF: y_i = i / n for sorted x_i
    y = np.arange(1, n + 1, dtype=float) / float(n)
    return x, y


def compute_weighted_ecdf(values: List[float], weights: List[float]):
    if not values or not weights or len(values) != len(weights):
        return np.array([]), np.array([])

    x = np.asarray(values, dtype=float)
    w = np.asarray(weights, dtype=float)

    if np.all(w <= 0.0):
        return np.array([]), np.array([])

    order = np.argsort(x)
    x_sorted = x[order]
    w_sorted = w[order]

    cum_w = np.cumsum(w_sorted)
    total_w = cum_w[-1]
    if total_w <= 0.0:
        return np.array([]), np.array([])

    y = cum_w / float(total_w)
    return x_sorted, y


def create_legend_pdf_horizontal(output_path: Path) -> None:
    fig, ax = plt.subplots(figsize=(12, 1.0))
    ax.axis("off")

    handles = []
    labels = []
    for model in MODEL_ORDER:
        color = MODEL_COLORS.get(model, "black")
        (handle,) = ax.plot([], [], "-", linewidth=2, color=color)
        handles.append(handle)
        labels.append(MODEL_LABELS.get(model, model))

    ax.legend(
        handles,
        labels,
        loc="center",
        ncol=len(handles),
        frameon=False,
        fancybox=False,
        shadow=False,
        borderaxespad=0.1,
        borderpad=0.3,
        handletextpad=0.4,
        labelspacing=0.2,
        prop={"size": 14},
    )

    fig.tight_layout(pad=0.0)
    fig.savefig(
        output_path,
        format="pdf",
        dpi=300,
        bbox_inches="tight",
        pad_inches=0.0,
    )
    plt.close(fig)
    print(f"saved horizontal legend: {output_path}")


def plot_overall_ecdf(
    overall_values_by_model: Dict[str, List[float]],
    overall_weights_by_model: Dict[str, List[float]],
    out_dir: Path,
) -> None:
    fig, ax = plt.subplots(figsize=(6, 4))

    any_line = False
    for model in MODEL_ORDER:
        values = overall_values_by_model.get(model)
        weights = overall_weights_by_model.get(model)
        if not values or not weights or len(values) != len(weights):
            continue
        x, y = compute_weighted_ecdf(values, weights)
        if x.size == 0:
            continue
        x_plot = x / 1_000_000.0
        color = MODEL_COLORS.get(model, "black")
        label = MODEL_LABELS.get(model, model)
        ax.plot(x_plot, y, label=label, color=color, linewidth=2.0)
        any_line = True

    if not any_line:
        plt.close(fig)
        print("no ECDF lines drawn for overall, skip figure")
        return

    ax.set_xlabel(r"$\mathbf{Trace\ duration\ [10^3\ s]}$", fontsize=22)
    ax.set_ylabel("", fontsize=22)

    ax.set_ylim(0.0, 1.0)

    ax.grid(True, which="both", axis="both", linestyle="-", linewidth=0.5, alpha=0.4)

    ax.tick_params(axis="both", labelsize=22)
    plt.setp(ax.get_xticklabels(), fontweight="bold")
    plt.setp(ax.get_yticklabels(), fontweight="bold")

    ax.margins(x=0.01)
    fig.tight_layout(pad=0.0)

    out_dir.mkdir(parents=True, exist_ok=True)
    out_file = out_dir / "ecdf_overall_time_weighted.pdf"
    fig.savefig(out_file, dpi=200, bbox_inches="tight", pad_inches=0.02)
    plt.close(fig)
    print(f"saved ECDF figure: {out_file}")


def compute_time_statistics(values: List[float]) -> Dict[str, float]:
    """Compute time statistics for a list of values (in milliseconds).
    Returns:
        Dictionary with mean, median, and total in seconds.
    """
    if not values:
        return {"mean": 0.0, "median": 0.0, "total": 0.0, "count": 0}

    # Convert from milliseconds to seconds
    values_sec = [v / 1000.0 for v in values]
    return {
        "mean": np.mean(values_sec),
        "median": np.median(values_sec),
        "total": np.sum(values_sec),
        "count": len(values),
    }


def generate_mcp_vs_hardcoded_comparison(
    base_project: str, scenario_time_data: Dict[str, Dict[str, List[float]]]
) -> str:
    """Generate comparison between MCP and hardcoded versions for a project."""
    lines = []
    lines.append(f"# {base_project}: MCP vs Hardcoded\n\n")

    mcp_scenario = f"{base_project}-MCP"
    hardcoded_scenario = base_project

    if (
        mcp_scenario not in scenario_time_data
        or hardcoded_scenario not in scenario_time_data
    ):
        lines.append("_Data not available for comparison_\n\n")
        return "".join(lines)

    mcp_data = scenario_time_data[mcp_scenario]
    hardcoded_data = scenario_time_data[hardcoded_scenario]

    all_models = sorted(set(mcp_data.keys()) | set(hardcoded_data.keys()))

    # Calculate overall averages first
    overall_stats = {
        "mcp": {"total": 0.0, "count": 0},
        "hard": {"total": 0.0, "count": 0},
    }
    for model in all_models:
        mcp_stats = compute_time_statistics(mcp_data.get(model, []))
        hard_stats = compute_time_statistics(hardcoded_data.get(model, []))
        overall_stats["mcp"]["total"] += mcp_stats["total"]
        overall_stats["mcp"]["count"] += mcp_stats["count"]
        overall_stats["hard"]["total"] += hard_stats["total"]
        overall_stats["hard"]["count"] += hard_stats["count"]

    # Add overall summary section
    lines.append("## Overall Summary (Averaged Across All Models)\n\n")
    lines.append("| MCP Mean (s) | Hardcoded Mean (s) | Diff (MCP-Hard) |\n")
    lines.append("| --- | --- | --- |\n")

    if overall_stats["mcp"]["count"] > 0 and overall_stats["hard"]["count"] > 0:
        avg_mcp = overall_stats["mcp"]["total"] / overall_stats["mcp"]["count"]
        avg_hard = overall_stats["hard"]["total"] / overall_stats["hard"]["count"]
        diff = avg_mcp - avg_hard
        pct = (diff / avg_hard * 100) if avg_hard > 0 else 0
        lines.append(
            f"| {avg_mcp:.2f} | {avg_hard:.2f} | {diff:+.2f}s ({pct:+.1f}%) |\n"
        )

    lines.append("\n---\n\n")

    # Per-model comparison
    lines.append("## Per-Model Comparison\n\n")
    lines.append("| Model | MCP Mean (s) | Hardcoded Mean (s) | Diff (MCP-Hard) |\n")
    lines.append("| --- | --- | --- | --- |\n")

    for model in all_models:
        mcp_stats = compute_time_statistics(mcp_data.get(model, []))
        hard_stats = compute_time_statistics(hardcoded_data.get(model, []))

        mean_diff = mcp_stats["mean"] - hard_stats["mean"]
        mean_pct = (
            (mean_diff / hard_stats["mean"] * 100) if hard_stats["mean"] > 0 else 0
        )

        lines.append(
            f"| {model} | {mcp_stats['mean']:.2f} | {hard_stats['mean']:.2f} | "
            f"{mean_diff:+.2f}s ({mean_pct:+.1f}%) |\n"
        )
    lines.append("\n")

    return "".join(lines)


def generate_mcp_vs_hardcoded_overall_comparison(
    projects: List[str], scenario_time_data: Dict[str, Dict[str, List[float]]]
) -> str:
    """Generate overall comparison across all MCP vs hardcoded projects."""
    lines = []
    lines.append("# Overall MCP vs Hardcoded Comparison\n\n")
    lines.append(f"Averaged across all projects: {', '.join(projects)}\n\n")

    # Collect data from all projects
    overall_data = {}
    framework_stats = {
        "mcp": {"total": 0.0, "count": 0},
        "hard": {"total": 0.0, "count": 0},
    }

    for project in projects:
        mcp_scenario = f"{project}-MCP"
        hardcoded_scenario = project

        if (
            mcp_scenario not in scenario_time_data
            or hardcoded_scenario not in scenario_time_data
        ):
            continue

        mcp_data = scenario_time_data[mcp_scenario]
        hardcoded_data = scenario_time_data[hardcoded_scenario]

        all_models = set(mcp_data.keys()) | set(hardcoded_data.keys())

        for model in all_models:
            if model not in overall_data:
                overall_data[model] = {"mcp": [], "hard": []}

            mcp_vals = mcp_data.get(model, [])
            hard_vals = hardcoded_data.get(model, [])

            overall_data[model]["mcp"].extend(mcp_vals)
            overall_data[model]["hard"].extend(hard_vals)

            # Add to framework-level stats
            mcp_stats = compute_time_statistics(mcp_vals)
            hard_stats = compute_time_statistics(hard_vals)
            framework_stats["mcp"]["total"] += mcp_stats["total"]
            framework_stats["mcp"]["count"] += mcp_stats["count"]
            framework_stats["hard"]["total"] += hard_stats["total"]
            framework_stats["hard"]["count"] += hard_stats["count"]

    # Add framework-level comparison
    lines.append("## Framework-Level Comparison (All Models Averaged)\n\n")
    lines.append("| MCP Mean (s) | Hardcoded Mean (s) | Diff (MCP-Hard) |\n")
    lines.append("| --- | --- | --- |\n")

    if framework_stats["mcp"]["count"] > 0 and framework_stats["hard"]["count"] > 0:
        avg_mcp = framework_stats["mcp"]["total"] / framework_stats["mcp"]["count"]
        avg_hard = framework_stats["hard"]["total"] / framework_stats["hard"]["count"]
        diff = avg_mcp - avg_hard
        pct = (diff / avg_hard * 100) if avg_hard > 0 else 0
        lines.append(
            f"| {avg_mcp:.2f} | {avg_hard:.2f} | {diff:+.2f}s ({pct:+.1f}%) |\n"
        )

    lines.append("\n---\n\n")

    # Per-model summary
    lines.append("## Per-Model Summary\n\n")
    lines.append("| Model | MCP Mean (s) | Hard Mean (s) | Diff (MCP-Hard) |\n")
    lines.append("| --- | --- | --- | --- |\n")

    for model in sorted(overall_data.keys()):
        mcp_stats = compute_time_statistics(overall_data[model]["mcp"])
        hard_stats = compute_time_statistics(overall_data[model]["hard"])

        mean_diff = mcp_stats["mean"] - hard_stats["mean"]
        mean_pct = (
            (mean_diff / hard_stats["mean"] * 100) if hard_stats["mean"] > 0 else 0
        )

        lines.append(
            f"| {model} | {mcp_stats['mean']:.2f} | {hard_stats['mean']:.2f} | "
            f"{mean_diff:+.2f}s ({mean_pct:+.1f}%) |\n"
        )
    lines.append("\n---\n\n")

    return "".join(lines)


def generate_version_comparison(
    base_project: str,
    version_a_suffix: str,
    version_b_suffix: str,
    scenario_time_data: Dict[str, Dict[str, List[float]]],
    version_a_name: str,
    version_b_name: str,
) -> str:
    """Generate comparison between two versions of a project."""
    lines = []
    lines.append(f"# {base_project}: {version_a_name} vs {version_b_name}\n\n")

    scenario_a = f"{base_project}{version_a_suffix}"
    scenario_b = f"{base_project}{version_b_suffix}"

    if scenario_a not in scenario_time_data or scenario_b not in scenario_time_data:
        lines.append("_Data not available for comparison_\n\n")
        return "".join(lines)

    data_a = scenario_time_data[scenario_a]
    data_b = scenario_time_data[scenario_b]

    all_models = sorted(set(data_a.keys()) | set(data_b.keys()))

    # Calculate overall averages first
    overall_stats = {"a": {"total": 0.0, "count": 0}, "b": {"total": 0.0, "count": 0}}
    for model in all_models:
        stats_a = compute_time_statistics(data_a.get(model, []))
        stats_b = compute_time_statistics(data_b.get(model, []))
        overall_stats["a"]["total"] += stats_a["total"]
        overall_stats["a"]["count"] += stats_a["count"]
        overall_stats["b"]["total"] += stats_b["total"]
        overall_stats["b"]["count"] += stats_b["count"]

    # Add overall summary section
    lines.append("## Overall Summary (Averaged Across All Models)\n\n")
    lines.append(
        f"| {version_a_name} Mean (s) | {version_b_name} Mean (s) | Diff ({version_a_name}-{version_b_name}) |\n"
    )
    lines.append("| --- | --- | --- |\n")

    if overall_stats["a"]["count"] > 0 and overall_stats["b"]["count"] > 0:
        avg_a = overall_stats["a"]["total"] / overall_stats["a"]["count"]
        avg_b = overall_stats["b"]["total"] / overall_stats["b"]["count"]
        diff = avg_a - avg_b
        pct = (diff / avg_b * 100) if avg_b > 0 else 0
        lines.append(f"| {avg_a:.2f} | {avg_b:.2f} | {diff:+.2f}s ({pct:+.1f}%) |\n")

    lines.append("\n---\n\n")

    # Per-model comparison
    lines.append("## Per-Model Comparison\n\n")
    lines.append(
        f"| Model | {version_a_name} Mean (s) | {version_b_name} Mean (s) | Diff ({version_a_name}-{version_b_name}) |\n"
    )
    lines.append("| --- | --- | --- | --- |\n")

    for model in all_models:
        stats_a = compute_time_statistics(data_a.get(model, []))
        stats_b = compute_time_statistics(data_b.get(model, []))

        mean_diff = stats_a["mean"] - stats_b["mean"]
        mean_pct = (mean_diff / stats_b["mean"] * 100) if stats_b["mean"] > 0 else 0

        lines.append(
            f"| {model} | {stats_a['mean']:.2f} | {stats_b['mean']:.2f} | "
            f"{mean_diff:+.2f}s ({mean_pct:+.1f}%) |\n"
        )
    lines.append("\n")

    return "".join(lines)


def generate_version_overall_comparison(
    projects: List[str],
    version_a_suffix: str,
    version_b_suffix: str,
    scenario_time_data: Dict[str, Dict[str, List[float]]],
    version_a_name: str,
    version_b_name: str,
    comparison_title: str,
) -> str:
    """Generate overall comparison across all projects for two versions."""
    lines = []
    lines.append(f"# Overall {comparison_title}\n\n")
    lines.append(f"Averaged across all projects: {', '.join(projects)}\n\n")

    # Collect data from all projects
    overall_data = {}
    framework_stats = {"a": {"total": 0.0, "count": 0}, "b": {"total": 0.0, "count": 0}}

    for project in projects:
        scenario_a = f"{project}{version_a_suffix}"
        scenario_b = f"{project}{version_b_suffix}"

        if scenario_a not in scenario_time_data or scenario_b not in scenario_time_data:
            continue

        data_a = scenario_time_data[scenario_a]
        data_b = scenario_time_data[scenario_b]

        all_models = set(data_a.keys()) | set(data_b.keys())

        for model in all_models:
            if model not in overall_data:
                overall_data[model] = {"a": [], "b": []}

            vals_a = data_a.get(model, [])
            vals_b = data_b.get(model, [])

            overall_data[model]["a"].extend(vals_a)
            overall_data[model]["b"].extend(vals_b)

            # Add to framework-level stats
            stats_a = compute_time_statistics(vals_a)
            stats_b = compute_time_statistics(vals_b)
            framework_stats["a"]["total"] += stats_a["total"]
            framework_stats["a"]["count"] += stats_a["count"]
            framework_stats["b"]["total"] += stats_b["total"]
            framework_stats["b"]["count"] += stats_b["count"]

    # Add framework-level comparison
    lines.append("## Framework-Level Comparison (All Models Averaged)\n\n")
    lines.append(
        f"| {version_a_name} Mean (s) | {version_b_name} Mean (s) | Diff ({version_a_name}-{version_b_name}) |\n"
    )
    lines.append("| --- | --- | --- |\n")

    if framework_stats["a"]["count"] > 0 and framework_stats["b"]["count"] > 0:
        avg_a = framework_stats["a"]["total"] / framework_stats["a"]["count"]
        avg_b = framework_stats["b"]["total"] / framework_stats["b"]["count"]
        diff = avg_a - avg_b
        pct = (diff / avg_b * 100) if avg_b > 0 else 0
        lines.append(f"| {avg_a:.2f} | {avg_b:.2f} | {diff:+.2f}s ({pct:+.1f}%) |\n")

    lines.append("\n---\n\n")

    # Per-model summary
    lines.append("## Per-Model Summary\n\n")
    lines.append(
        f"| Model | {version_a_name} Mean (s) | {version_b_name} Mean (s) | Diff ({version_a_name}-{version_b_name}) |\n"
    )
    lines.append("| --- | --- | --- | --- |\n")

    for model in sorted(overall_data.keys()):
        stats_a = compute_time_statistics(overall_data[model]["a"])
        stats_b = compute_time_statistics(overall_data[model]["b"])

        mean_diff = stats_a["mean"] - stats_b["mean"]
        mean_pct = (mean_diff / stats_b["mean"] * 100) if stats_b["mean"] > 0 else 0

        lines.append(
            f"| {model} | {stats_a['mean']:.2f} | {stats_b['mean']:.2f} | "
            f"{mean_diff:+.2f}s ({mean_pct:+.1f}%) |\n"
        )
    lines.append("\n---\n\n")

    return "".join(lines)


def get_base_project_name(scenario: str) -> str:
    if scenario.endswith("-A2A_mix"):
        return scenario[: -len("-A2A_mix")]
    if scenario.endswith("-H_A2A"):
        return scenario[: -len("-H_A2A")]
    if scenario.endswith("-A2A"):
        return scenario[: -len("-A2A")]
    if scenario.endswith("-MCP"):
        return scenario[: -len("-MCP")]
    return scenario


def generate_overall_model_comparison(
    scenario_time_data: Dict[str, Dict[str, List[float]]],
) -> str:
    """Generate an all-projects summary comparing models across every scenario."""

    lines: List[str] = []
    lines.append("## All Projects Combined (Summary Across All Projects, by Model)\n\n")

    # Aggregate all total_classified samples per model across every project scenario
    aggregated: Dict[str, List[float]] = defaultdict(list)
    for project_data in scenario_time_data.values():
        for model, vals in project_data.items():
            aggregated[model].extend(vals)

    if not aggregated:
        lines.append("_No data available across projects_\n\n")
        return "".join(lines)

    # Respect fixed display order, then any remaining models alphabetically
    ordered_models: List[str] = [
        m for m in MODEL_ORDER if m in aggregated and aggregated[m]
    ]
    remaining_models = sorted(
        m for m in aggregated.keys() if m not in ordered_models and aggregated[m]
    )
    all_models = ordered_models + remaining_models

    # Compute statistics for each model
    model_stats: Dict[str, Dict[str, float]] = {}
    for model in all_models:
        model_stats[model] = compute_time_statistics(aggregated[model])

    # Summary table
    lines.append("### Model Statistics Summary\n\n")
    lines.append("| Model | Mean (s) | Median (s) | Count | Total (s) |\n")
    lines.append("| --- | --- | --- | --- | --- |\n")

    for model in all_models:
        stats = model_stats[model]
        lines.append(
            f"| {model} | {stats['mean']:.2f} | {stats['median']:.2f} | "
            f"{stats['count']} | {stats['total']:.2f} |\n"
        )

    # Relative performance vs fastest (only if at least two models with data)
    positive_models = [m for m in all_models if model_stats[m]["mean"] > 0]
    if len(positive_models) > 1:
        fastest_model = min(positive_models, key=lambda m: model_stats[m]["mean"])
        fastest_mean = model_stats[fastest_model]["mean"]

        lines.append("\n### Relative Performance (vs. Fastest Model)\n\n")
        lines.append(
            f"Baseline (fastest): **{fastest_model}** ({fastest_mean:.2f}s mean)\n\n"
        )
        lines.append("| Model | Mean (s) | Slowdown vs Baseline |\n")
        lines.append("| --- | --- | --- |\n")

        for model in all_models:
            stats = model_stats[model]
            if stats["mean"] > 0 and fastest_mean > 0:
                slowdown = (stats["mean"] - fastest_mean) / fastest_mean * 100
                lines.append(f"| {model} | {stats['mean']:.2f} | {slowdown:+.1f}% |\n")
            else:
                lines.append(f"| {model} | {stats['mean']:.2f} | N/A |\n")

        # Add slowest model summary
        slowest_model = max(
            positive_models,
            key=lambda m: model_stats[m]["mean"] if model_stats[m]["mean"] > 0 else 0,
        )
        slowest_mean = model_stats[slowest_model]["mean"]
        if fastest_mean > 0 and slowest_mean > 0:
            slowest_slowdown = (slowest_mean - fastest_mean) / fastest_mean * 100
            lines.append(
                f"\n**Slowest model:** {slowest_model} ({slowest_mean:.2f}s mean, "
                f"{slowest_slowdown:+.1f}% slower than baseline)\n"
            )

    lines.append("\n")
    return "".join(lines)


def generate_project_model_comparison(
    project_name: str, scenario_time_data: Dict[str, Dict[str, List[float]]]
) -> str:
    """Generate model-to-model comparison for a single project.
    Compares all models within the same project scenario.
    """
    lines = []
    lines.append(f"# {project_name}: Model Comparison\n\n")

    if project_name not in scenario_time_data:
        lines.append("_Data not available for this project_\n\n")
        return "".join(lines)

    project_data = scenario_time_data[project_name]
    all_models = sorted(project_data.keys())

    if not all_models:
        lines.append("_No model data available_\n\n")
        return "".join(lines)

    # Compute statistics for each model
    model_stats = {}
    for model in all_models:
        vals = project_data.get(model, [])
        model_stats[model] = compute_time_statistics(vals)

    # Summary table
    lines.append("## Model Statistics Summary\n\n")
    lines.append("| Model | Mean (s) | Median (s) | Count | Total (s) |\n")
    lines.append("| --- | --- | --- | --- | --- |\n")

    for model in all_models:
        stats = model_stats[model]
        lines.append(
            f"| {model} | {stats['mean']:.2f} | {stats['median']:.2f} | "
            f"{stats['count']} | {stats['total']:.2f} |\n"
        )

    # Pairwise comparison: compare each model against the fastest one
    if len(all_models) > 1:
        lines.append("\n## Relative Performance (vs. Fastest Model)\n\n")

        # Find fastest model by mean
        fastest_model = min(
            all_models,
            key=lambda m: (
                model_stats[m]["mean"] if model_stats[m]["mean"] > 0 else float("inf")
            ),
        )
        fastest_mean = model_stats[fastest_model]["mean"]

        # Find slowest model by mean
        slowest_model = max(
            all_models,
            key=lambda m: model_stats[m]["mean"] if model_stats[m]["mean"] > 0 else 0,
        )
        slowest_mean = model_stats[slowest_model]["mean"]

        lines.append(
            f"Baseline (fastest): **{fastest_model}** ({fastest_mean:.2f}s mean)\n\n"
        )
        lines.append("| Model | Mean (s) | Slowdown vs Baseline |\n")
        lines.append("| --- | --- | --- |\n")

        for model in all_models:
            stats = model_stats[model]
            if stats["mean"] > 0 and fastest_mean > 0:
                slowdown = (stats["mean"] - fastest_mean) / fastest_mean * 100
                lines.append(f"| {model} | {stats['mean']:.2f} | {slowdown:+.1f}% |\n")
            else:
                lines.append(f"| {model} | {stats['mean']:.2f} | N/A |\n")

        # Add slowest model summary
        if fastest_mean > 0 and slowest_mean > 0:
            slowest_slowdown = (slowest_mean - fastest_mean) / fastest_mean * 100
            lines.append(
                f"\n**Slowest model:** {slowest_model} ({slowest_mean:.2f}s mean, "
                f"{slowest_slowdown:+.1f}% slower than baseline)\n"
            )

    lines.append("\n")
    return "".join(lines)


def plot_ecdf_for_project(
    project_dir: Path, csv_path: Path, out_dir: Path, x_max_ms: float
) -> None:
    """Plot ECDF of total_classified for all models in one project.
    One figure per project, up to 7 lines (one per model present in the CSV).
    """

    data_by_model = load_total_classified(csv_path)
    if not data_by_model:
        print(f"no total_classified data in {csv_path}, skip")
        return

    # If no global series maximum was provided, fall back to this project's
    # own maximum so that the function remains usable in isolation.
    if x_max_ms <= 0.0:
        local_max = 0.0
        for vals in data_by_model.values():
            if vals:
                v_max = max(vals)
                if v_max > local_max:
                    local_max = v_max
        x_max_ms = local_max

    fig, ax = plt.subplots(figsize=(6, 4))

    # For a consistent legend order, iterate in fixed MODEL_ORDER
    any_line = False
    for model in MODEL_ORDER:
        values = data_by_model.get(model)
        if not values:
            continue
        x, y = compute_ecdf(values)
        if x.size == 0:
            continue
        # Plot in units of 10^3 s so that the 1e3 scaling factor is
        # explicitly captured in the axis label rather than as a separate
        # offset text.
        x_plot = x / 1_000_000.0
        color = MODEL_COLORS.get(model, "black")
        label = MODEL_LABELS.get(model, model)
        ax.plot(x_plot, y, label=label, color=color, linewidth=2.0)
        any_line = True

    if not any_line:
        plt.close(fig)
        print(f"no ECDF lines drawn for {csv_path}, skip figure")
        return

    # Use a shared x-axis upper bound (in milliseconds) for all scenarios in
    # the same project series so that their ECDFs are directly comparable.
    x_max_plot = x_max_ms / 1_000_000.0
    if x_max_plot > 0.0:
        ax.set_xlim(0.0, x_max_plot)

    # Show axis units directly in terms of 10^3 s using LaTeX bold.
    ax.set_xlabel(r"$\mathbf{Trace\ duration\ [10^3\ s]}$", fontsize=22)
    # No explicit y-axis label (ECDF) to keep the figure clean.
    ax.set_ylabel("", fontsize=22)

    ax.set_ylim(0.0, 1.0)

    # Add light grid similar to typical ECDF examples
    ax.grid(True, which="both", axis="both", linestyle="-", linewidth=0.5, alpha=0.4)

    ax.tick_params(axis="both", labelsize=22)
    plt.setp(ax.get_xticklabels(), fontweight="bold")
    plt.setp(ax.get_yticklabels(), fontweight="bold")

    # Slightly reduce margins so curves fill the axes area; no in-figure legend or title
    ax.margins(x=0.01)
    fig.tight_layout(pad=0.0)

    out_dir.mkdir(parents=True, exist_ok=True)
    out_file = out_dir / f"ecdf_{project_dir.name}.pdf"
    # Use a tiny padding so the right axis spine is fully preserved while
    # keeping extra whitespace visually negligible.
    fig.savefig(out_file, dpi=200, bbox_inches="tight", pad_inches=0.02)
    plt.close(fig)
    print(f"saved ECDF figure: {out_file}")


def main() -> None:
    # Assume this script is placed in Part2 directory
    part2_dir = Path(__file__).resolve().parent

    # Output directory for ECDF figures
    out_dir = part2_dir / "ECDFs"

    overall_values_by_model: Dict[str, List[float]] = defaultdict(list)
    overall_weights_by_model: Dict[str, List[float]] = defaultdict(list)

    # Dictionary to store time data per scenario for comparisons
    scenario_time_data: Dict[str, Dict[str, List[float]]] = {}

    # Per-series maximum of total_classified (in milliseconds) so that
    # scenarios sharing the same base project name use a common x-axis range.
    series_x_max_ms: Dict[str, float] = {}

    # First pass: load data, build overall ECDF inputs, and compute per-series maxima.
    for sub in sorted(p for p in part2_dir.iterdir() if p.is_dir()):
        if sub.name.startswith("z_"):
            # Skip output directories
            continue
        csv_path = sub / "performance_breakdown_summary.csv"
        if not csv_path.exists():
            continue
        data_by_model = load_total_classified(csv_path)

        # Store scenario data for comparisons
        scenario_time_data[sub.name] = data_by_model

        # Update weights used for the overall ECDF
        for model, vals in data_by_model.items():
            if not vals:
                continue
            total_time = float(sum(vals))
            count = len(vals)
            if total_time <= 0.0 or count <= 0:
                continue
            weight_per_sample = total_time / float(count)
            for v in vals:
                overall_values_by_model[model].append(v)
                overall_weights_by_model[model].append(weight_per_sample)

        # Track the maximum total_classified for this scenario and propagate it
        # to the corresponding project series.
        scenario_max = 0.0
        for vals in data_by_model.values():
            if vals:
                v_max = max(vals)
                if v_max > scenario_max:
                    scenario_max = v_max
        if scenario_max > 0.0:
            base_name = get_base_project_name(sub.name)
            prev_max = series_x_max_ms.get(base_name, 0.0)
            if scenario_max > prev_max:
                series_x_max_ms[base_name] = scenario_max

    # Second pass: draw ECDF for each scenario using the shared x-axis maximum
    # per project series.
    for sub in sorted(p for p in part2_dir.iterdir() if p.is_dir()):
        if sub.name.startswith("z_"):
            continue
        csv_path = sub / "performance_breakdown_summary.csv"
        if not csv_path.exists():
            continue
        base_name = get_base_project_name(sub.name)
        x_max_ms = series_x_max_ms.get(base_name, 0.0)
        print(f"processing {csv_path}")
        try:
            plot_ecdf_for_project(sub, csv_path, out_dir, x_max_ms)
        except Exception as exc:  # pragma: no cover - defensive
            print(f"  error while plotting {csv_path}: {exc}")

    if overall_values_by_model:
        plot_overall_ecdf(overall_values_by_model, overall_weights_by_model, out_dir)

    # Also create a standalone horizontal legend PDF (one per script run)
    legend_path = out_dir / "ECDF_Model_Legend_horizontal.pdf"
    create_legend_pdf_horizontal(legend_path)

    # Generate time comparison markdown files
    print("\n" + "=" * 60)
    print("Generating time comparison summaries...")
    print("=" * 60)

    # 1. MCP vs Hardcoded comparisons
    mcp_hardcoded_projects = [
        "MarkdownValidator",
        "GameBuilder",
        "EmailResponder",
    ]

    comparison_lines = []
    # Add overall summary first
    comparison_lines.append(
        generate_mcp_vs_hardcoded_overall_comparison(
            mcp_hardcoded_projects, scenario_time_data
        )
    )

    # Then add per-project comparisons
    for project in mcp_hardcoded_projects:
        comparison_lines.append(
            generate_mcp_vs_hardcoded_comparison(project, scenario_time_data)
        )

    comparison_md_path = out_dir / "Time_Comparison_MCP_vs_Hardcoded.md"
    comparison_md_path.write_text("".join(comparison_lines), encoding="utf-8")
    print(f"Created: {comparison_md_path}")

    # 2. MCP vs A2A comparisons
    version_projects = [
        "SQL_assistant",
        "intelligent_recruitment_platform",
        "landing_page_generator",
        "self_evaluation_loop_flow",
        "write_a_book_with_flows",
    ]

    comparison_lines = []
    # Add overall summary first
    comparison_lines.append(
        generate_version_overall_comparison(
            version_projects,
            "-MCP",
            "-A2A",
            scenario_time_data,
            "MCP",
            "A2A",
            "MCP vs A2A Time Comparison",
        )
    )

    # Then add per-project comparisons
    for project in version_projects:
        comparison_lines.append(
            generate_version_comparison(
                project, "-MCP", "-A2A", scenario_time_data, "MCP", "A2A"
            )
        )

    comparison_md_path = out_dir / "Time_Comparison_MCP_vs_A2A.md"
    comparison_md_path.write_text("".join(comparison_lines), encoding="utf-8")
    print(f"Created: {comparison_md_path}")

    # 3. A2A vs A2A_mix comparisons
    comparison_lines = []
    # Add overall summary first
    comparison_lines.append(
        generate_version_overall_comparison(
            version_projects,
            "-A2A",
            "-A2A_mix",
            scenario_time_data,
            "A2A",
            "A2A_mix",
            "A2A vs A2A_mix Time Comparison",
        )
    )

    # Then add per-project comparisons
    for project in version_projects:
        comparison_lines.append(
            generate_version_comparison(
                project, "-A2A", "-A2A_mix", scenario_time_data, "A2A", "A2A_mix"
            )
        )

    comparison_md_path = out_dir / "Time_Comparison_A2A_vs_A2A_mix.md"
    comparison_md_path.write_text("".join(comparison_lines), encoding="utf-8")
    print(f"Created: {comparison_md_path}")

    # 4. Generate detailed A2A vs A2A_mix comparison for each project
    print("\nGenerating detailed A2A vs A2A_mix per-project comparisons...")
    a2a_mix_comparison_lines = []
    a2a_mix_comparison_lines.append(
        "# A2A vs A2A_mix: Detailed Per-Project Comparison\n\n"
    )
    a2a_mix_comparison_lines.append(
        "This document compares A2A and A2A_mix architectures for each project, "
    )
    a2a_mix_comparison_lines.append(
        "showing both per-model and overall statistics.\n\n"
    )
    a2a_mix_comparison_lines.append("---\n\n")

    # Aggregators for cross-project summary
    global_all_models = sorted(
        set().union(*[set(d.keys()) for d in scenario_time_data.values()])
    )
    overall_deltas: List[Dict[str, float]] = []
    per_model_global: Dict[str, Dict[str, float]] = defaultdict(
        lambda: {"a2a_sum": 0.0, "a2a_cnt": 0, "mix_sum": 0.0, "mix_cnt": 0}
    )

    for project in version_projects:
        scenario_a2a = f"{project}-A2A"
        scenario_a2a_mix = f"{project}-A2A_mix"

        if (
            scenario_a2a not in scenario_time_data
            or scenario_a2a_mix not in scenario_time_data
        ):
            continue

        a2a_mix_comparison_lines.append(f"## {project}\n\n")
        a2a_mix_comparison_lines.append("### Project-Level Summary\n\n")

        data_a2a = scenario_time_data[scenario_a2a]
        data_a2a_mix = scenario_time_data[scenario_a2a_mix]
        all_models = sorted(set(data_a2a.keys()) | set(data_a2a_mix.keys()))

        # Overall comparison
        overall_a2a_total = 0.0
        overall_a2a_count = 0
        overall_a2a_mix_total = 0.0
        overall_a2a_mix_count = 0

        for model in all_models:
            stats_a2a = compute_time_statistics(data_a2a.get(model, []))
            stats_a2a_mix = compute_time_statistics(data_a2a_mix.get(model, []))
            overall_a2a_total += stats_a2a["total"]
            overall_a2a_count += stats_a2a["count"]
            overall_a2a_mix_total += stats_a2a_mix["total"]
            overall_a2a_mix_count += stats_a2a_mix["count"]
            # accumulate for global per-model summary
            per_model_global[model]["a2a_sum"] += stats_a2a["total"]
            per_model_global[model]["a2a_cnt"] += stats_a2a["count"]
            per_model_global[model]["mix_sum"] += stats_a2a_mix["total"]
            per_model_global[model]["mix_cnt"] += stats_a2a_mix["count"]

        if overall_a2a_count > 0 and overall_a2a_mix_count > 0:
            avg_a2a = overall_a2a_total / overall_a2a_count
            avg_a2a_mix = overall_a2a_mix_total / overall_a2a_mix_count
            diff = avg_a2a_mix - avg_a2a
            pct = (diff / avg_a2a * 100) if avg_a2a > 0 else 0
            a2a_mix_comparison_lines.append("### Overall Summary\n\n")
            a2a_mix_comparison_lines.append(
                "| A2A Mean (s) | A2A_mix Mean (s) | Diff (A2A_mix - A2A) |\n"
            )
            a2a_mix_comparison_lines.append("| --- | --- | --- |\n")
            a2a_mix_comparison_lines.append(
                f"| {avg_a2a:.2f} | {avg_a2a_mix:.2f} | {diff:+.2f}s ({pct:+.1f}%) |\n\n"
            )
            overall_deltas.append(
                {
                    "project": project,
                    "a2a": avg_a2a,
                    "mix": avg_a2a_mix,
                    "diff": diff,
                    "pct": pct,
                }
            )

        # Per-model comparison
        a2a_mix_comparison_lines.append("### Per-Model Comparison\n\n")
        a2a_mix_comparison_lines.append(
            "| Model | A2A Mean (s) | A2A_mix Mean (s) | Diff (A2A_mix - A2A) |\n"
        )
        a2a_mix_comparison_lines.append("| --- | --- | --- | --- |\n")

        for model in all_models:
            stats_a2a = compute_time_statistics(data_a2a.get(model, []))
            stats_a2a_mix = compute_time_statistics(data_a2a_mix.get(model, []))
            mean_diff = stats_a2a_mix["mean"] - stats_a2a["mean"]
            mean_pct = (
                (mean_diff / stats_a2a["mean"] * 100) if stats_a2a["mean"] > 0 else 0
            )
            a2a_mix_comparison_lines.append(
                f"| {model} | {stats_a2a['mean']:.2f} | {stats_a2a_mix['mean']:.2f} | "
                f"{mean_diff:+.2f}s ({mean_pct:+.1f}%) |\n"
            )

        a2a_mix_comparison_lines.append("\n---\n\n")

    # Global cross-project summaries (coarse)
    if overall_deltas:
        a2a_mix_comparison_lines.insert(
            4,
            "## Overall (All Projects)\n\n"
            "| Project | A2A Mean (s) | A2A_mix Mean (s) | Diff (A2A_mix - A2A) |\n"
            "| --- | --- | --- | --- |\n"
            + "".join(
                f"| {d['project']} | {d['a2a']:.2f} | {d['mix']:.2f} | {d['diff']:+.2f}s ({d['pct']:+.1f}%) |\n"
                for d in overall_deltas
            )
            + "\n",
        )

    if per_model_global:
        per_model_lines = []
        per_model_lines.append("## All Projects Combined (Per-Model)\n\n")
        per_model_lines.append(
            "| Model | A2A Mean (s) | A2A_mix Mean (s) | Diff (A2A_mix - A2A) |\n"
        )
        per_model_lines.append("| --- | --- | --- | --- |\n")
        for model in MODEL_ORDER:
            stats = per_model_global.get(model)
            if not stats:
                continue
            a2a_cnt = stats["a2a_cnt"]
            mix_cnt = stats["mix_cnt"]
            if a2a_cnt <= 0 or mix_cnt <= 0:
                continue
            avg_a2a = stats["a2a_sum"] / a2a_cnt
            avg_mix = stats["mix_sum"] / mix_cnt
            diff = avg_mix - avg_a2a
            pct = (diff / avg_a2a * 100) if avg_a2a > 0 else 0
            per_model_lines.append(
                f"| {model} | {avg_a2a:.2f} | {avg_mix:.2f} | {diff:+.2f}s ({pct:+.1f}%) |\n"
            )
        per_model_lines.append("\n---\n\n")
        # insert after top intro (after first 4 elements added earlier)
        a2a_mix_comparison_lines[5:5] = per_model_lines

    a2a_mix_md_path = out_dir / "A2A_vs_A2A_mix_Detailed_Comparison.md"
    a2a_mix_md_path.write_text("".join(a2a_mix_comparison_lines), encoding="utf-8")
    print(f"Created: {a2a_mix_md_path}")

    # 5. Generate per-project model comparison for all 21 projects
    print("\nGenerating per-project model comparisons (21 projects)...")
    all_project_names = sorted(scenario_time_data.keys())

    model_comparison_lines = []
    model_comparison_lines.append("# Per-Project Model Performance Comparison\n\n")
    model_comparison_lines.append(
        f"This document compares model performance within each of the {len(all_project_names)} projects.\n\n"
    )
    model_comparison_lines.append("Each project shows:\n")
    model_comparison_lines.append(
        "- Model statistics (mean, median, count, total time)\n"
    )
    model_comparison_lines.append(
        "- Relative performance compared to the fastest model\n\n"
    )

    # Add global summary across all projects first
    model_comparison_lines.append("---\n\n")
    model_comparison_lines.append(generate_overall_model_comparison(scenario_time_data))
    model_comparison_lines.append("---\n\n")

    for project_name in all_project_names:
        model_comparison_lines.append(
            generate_project_model_comparison(project_name, scenario_time_data)
        )
        model_comparison_lines.append("---\n\n")

    model_comparison_md_path = out_dir / "Per_Project_Model_Comparison.md"
    model_comparison_md_path.write_text(
        "".join(model_comparison_lines), encoding="utf-8"
    )
    print(f"Created: {model_comparison_md_path}")

    print("\n" + "=" * 60)
    print("All time comparison summaries generated!")
    print("=" * 60)


if __name__ == "__main__":
    main()