boundary_eval_check.py

Go to the documentation of this file.

# Copyright (c) 2025-present Aquin Labs Private Limited. All Rights Reserved.
"""`aquin boundary-eval --check` — save JSON + diagram to the current working directory."""
 
from __future__ import annotations
 
import json
from datetime import datetime, timezone
from pathlib import Path
from typing import Any
 
_JSON_NAME = "boundary-eval-check.json"
_PNG_NAME = "boundary-eval-check.png"
 
 
def _normalize_boundary_eval_result(result: dict[str, Any]) -> dict[str, Any]:
    if not isinstance(result, dict):
        return {"error": "Invalid boundary-eval result"}
 
    if result.get("error"):
        return dict(result)
 
    data = result
    if isinstance(result.get("content"), dict):
        data = result["content"]
    if isinstance(data.get("boundary"), dict):
        data = data["boundary"]
    card = result.get("card")
    if isinstance(card, dict) and isinstance(card.get("data"), dict):
        nested = card["data"].get("boundary")
        if isinstance(nested, dict):
            data = nested
 
    return dict(data)
 
 
def has_boundary_eval_payload(result: dict[str, Any]) -> bool:
    data = _normalize_boundary_eval_result(result)
    if data.get("error"):
        return False
    return data.get("mean_robustness") is not None or bool(data.get("probes"))
 
 
def write_boundary_eval_check(
    result: dict[str, Any],
    *,
    tool_name: str | None,
    cwd: str | Path,
) -> tuple[Path, Path]:
    cwd = Path(cwd)
    json_path = cwd / _JSON_NAME
    png_path = cwd / _PNG_NAME
 
    flat = _normalize_boundary_eval_result(result)
    payload = {
        "saved_at": datetime.now(timezone.utc).isoformat(),
        "tool": tool_name,
        **flat,
    }
    json_path.write_text(json.dumps(payload, indent=2, default=str), encoding="utf-8")
    _plot_boundary_eval_check(flat, png_path)
    return json_path, png_path
 
 
def _plot_boundary_eval_check(result: dict[str, Any], png_path: Path) -> None:
    import matplotlib
 
    matplotlib.use("Agg")
    import matplotlib.pyplot as plt
 
    if result.get("error"):
        fig, ax = plt.subplots(figsize=(6, 2), facecolor="#0f1117")
        ax.axis("off")
        ax.text(0.5, 0.5, f"Error: {result['error']}", ha="center", va="center", wrap=True, color="#e5e7eb")
        fig.savefig(png_path, dpi=140, bbox_inches="tight", facecolor="#0f1117")
        plt.close(fig)
        return
 
    probes = result.get("probes") or []
    mean_rob = float(result.get("mean_robustness") or 0)
    title = f"Boundary eval — mean robustness {mean_rob:.3f}  ·  {len(probes)} probe(s)"
 
    fig, axes = plt.subplots(1, 2, figsize=(11, max(4.5, min(10, len(probes) * 0.35 + 2))), facecolor="#0f1117")
    fig.suptitle(title, color="#e5e7eb", fontsize=11, y=1.02)
 
    _plot_robustness_panel(axes[0], probes, mean_rob)
    _plot_degradation_panel(axes[1], probes)
 
    for ax in axes:
        ax.set_facecolor("#0f1117")
        for spine in ax.spines.values():
            spine.set_color("#374151")
 
    fig.tight_layout()
    fig.savefig(png_path, dpi=140, bbox_inches="tight", facecolor="#0f1117")
    plt.close(fig)
 
 
def _plot_robustness_panel(ax: Any, probes: list[dict[str, Any]], mean_rob: float) -> None:
    ax.set_title("Robustness by prompt", color="#e5e7eb", fontsize=10)
 
    if not probes:
        color = _score_color(mean_rob)
        ax.barh([0], [mean_rob], color=color, height=0.45, alpha=0.92)
        ax.set_xlim(0, 1.05)
        ax.set_yticks([0])
        ax.set_yticklabels(["mean"], fontsize=9, color="#d1d5db")
        ax.set_xlabel("Robustness (0–1)", color="#9ca3af", fontsize=8)
        ax.text(min(mean_rob + 0.02, 0.98), 0, f"{mean_rob:.3f}", va="center", fontsize=8, color="#9ca3af")
        ax.tick_params(axis="x", colors="#6b7280", labelsize=8)
        ax.grid(axis="x", alpha=0.2, color="#4b5563")
        return
 
    labels = [_trunc(str(p.get("prompt") or ""), 22) for p in probes]
    vals = [float(p.get("robustness_score") or 0) for p in probes]
    colors = [_score_color(v) for v in vals]
    y_pos = list(range(len(probes)))
 
    ax.barh(y_pos, vals, color=colors, height=0.65, alpha=0.9)
    ax.axvline(mean_rob, color="#38bdf8", linestyle="--", linewidth=0.8, alpha=0.7)
    ax.set_yticks(y_pos)
    ax.set_yticklabels(labels, fontsize=7, color="#d1d5db")
    ax.invert_yaxis()
    ax.set_xlim(0, 1.05)
    ax.set_xlabel("Robustness", color="#9ca3af", fontsize=8)
    ax.tick_params(axis="x", colors="#6b7280", labelsize=8)
    ax.grid(axis="x", alpha=0.2, color="#4b5563")
 
 
def _plot_degradation_panel(ax: Any, probes: list[dict[str, Any]]) -> None:
    ax.set_title("Confidence drop & KL under corruption", color="#e5e7eb", fontsize=10)
 
    if not probes:
        ax.axis("off")
        ax.text(0.5, 0.5, "No probes", ha="center", va="center", color="#9ca3af")
        return
 
    labels = [_trunc(str(p.get("prompt") or ""), 12) for p in probes]
    x = list(range(len(probes)))
    width = 0.38
    drops = [float(p.get("mean_confidence_drop") or 0) for p in probes]
    kls = [float(p.get("mean_kl") or 0) for p in probes]
 
    ax.bar([i - width / 2 for i in x], drops, width=width, color="#f87171", alpha=0.9, label="conf drop")
    ax.bar([i + width / 2 for i in x], kls, width=width, color="#6366f1", alpha=0.9, label="mean KL")
    ax.set_xticks(x)
    ax.set_xticklabels(labels, fontsize=7, color="#d1d5db", rotation=30, ha="right")
    ax.set_ylabel("Magnitude", color="#9ca3af", fontsize=8)
    ax.legend(fontsize=7, facecolor="#0f1117", edgecolor="#374151", labelcolor="#d1d5db")
    ax.tick_params(colors="#6b7280", labelsize=7)
    ax.grid(axis="y", alpha=0.2, color="#4b5563")
 
 
def _score_color(val: float) -> str:
    if val >= 0.75:
        return "#34d399"
    if val >= 0.5:
        return "#fbbf24"
    return "#f87171"
 
 
def _trunc(text: str, limit: int) -> str:
    t = text.strip()
    if len(t) <= limit:
        return t or "·"
    return t[: limit - 1] + "…"