"""Regenerate the blog's four panels from figure-data.json, using no models. Run with Python 3, NumPy, and Matplotlib. Paths resolve beside this script, so the command works from any directory. Original experiment files are not needed and are never opened by this script. """ from __future__ import annotations import argparse import hashlib import json from pathlib import Path import matplotlib matplotlib.use("Agg") import matplotlib.pyplot as plt from matplotlib.ticker import FixedLocator, FuncFormatter, NullLocator import numpy as np HERE = Path(__file__).resolve().parent INK = "#253348" BLUE = "#3268BD" ORANGE = "#C97836" MUTED = "#8490A2" PALETTE = {"smwm": BLUE, "pca": ORANGE, "direct": INK} DIMENSIONS = {"toy-phase": (5.5, 4.8), "panda-cycle": (6.4, 4.4), "activation-control": (5.5, 4.4), "waveform-control": (5.5, 4.4)} DPI = 200 def load_data(path): data = json.loads(Path(path).read_text(encoding="utf-8")) if data.get("schema_version") != 1: raise ValueError("Unsupported figure-data schema") trajectories = data["toy"]["trajectories"] if [row["mu"] for row in trajectories] != [-0.05, 0.05]: raise ValueError("The toy panel requires the two recorded parameter settings") for row in trajectories: values = np.asarray(row["xy"], dtype=np.float64) if values.shape != (600, 2) or not np.isfinite(values).all(): raise ValueError("Expected all 600 recorded two-dimensional toy states") for name in ("input_tail", "unedited_forecast", "edited_forecast"): values = np.asarray(data["panda_cycle"][name], dtype=np.float32) if values.shape != (128,) or not np.isfinite(values).all(): raise ValueError(f"Expected exactly 128 finite values for {name}") rows = data["control"]["series"] expected = {(objective, method, seed) for objective in ("activation", "waveform") for method in ("smwm", "pca", "direct") for seed in (0, 1, 2)} observed = {(row["objective"], row["method"], row["seed"]) for row in rows} if observed != expected or len(rows) != len(expected): raise ValueError("Expected both objectives and all three methods/seeds") for row in rows: values = np.asarray(row["values"], dtype=np.float32) if values.shape != (128,) or not np.isfinite(values).all() or np.any(values <= 0): raise ValueError("Control panels require all 128 positive, finite error ratios") # Preserve the source arrays' float32 reduction, including the even-N median. if float(np.median(values)) != row["median"]: raise ValueError("Saved median does not match the copied case-level ratios") return data def style_axis(ax): ax.spines[["top", "right"]].set_visible(False) ax.spines[["left", "bottom"]].set_color("#CAD0D9") ax.tick_params(colors=INK, length=3, pad=6) ax.set_axisbelow(True) def save(fig, directory, stem): paths = [directory / f"{stem}.png", directory / f"{stem}.pdf"] fig.savefig(paths[0], dpi=DPI, facecolor="white") # Stable PDF metadata makes repeated generation byte-identical in one environment. fig.savefig(paths[1], facecolor="white", metadata={"Creator": "generate_figures.py", "CreationDate": None, "ModDate": None}) plt.close(fig) return paths def toy_panel(data, directory): fig, ax = plt.subplots(figsize=DIMENSIONS["toy-phase"]) fig.subplots_adjust(left=.15, right=.96, bottom=.19, top=.79) for record, color, label in zip(data["toy"]["trajectories"], (BLUE, ORANGE), ("μ = −0.05: decays", "μ = +0.05: persists")): xy = np.asarray(record["xy"], dtype=np.float64) ax.plot(xy[:, 0], xy[:, 1], color=color, lw=1.8, label=label) ax.scatter(xy[-1, 0], xy[-1, 1], s=20, color=color, zorder=3) ax.set_aspect("equal", adjustable="box") ax.set_xlim(-.25, .25) ax.set_ylim(-.25, .25) ax.set_xticks([-.2, 0, .2]) ax.set_yticks([-.2, 0, .2]) ax.set_xlabel("State coordinate $z_1$", labelpad=7) ax.set_ylabel("State coordinate $z_2$", labelpad=7) style_axis(ax) ax.grid(alpha=.12, color=MUTED) fig.suptitle("Toy system—not PANDA", y=.98, fontsize=19, color=INK, fontweight="bold") handles, labels = ax.get_legend_handles_labels() fig.legend(handles, labels, loc="upper center", bbox_to_anchor=(.52, .915), ncol=2, frameon=False, fontsize=13, columnspacing=1.1, handlelength=1.3) return save(fig, directory, "toy-phase") def cycle_panel(data, directory): values = data["panda_cycle"] fig, ax = plt.subplots(figsize=DIMENSIONS["panda-cycle"]) fig.subplots_adjust(left=.15, right=.94, bottom=.19, top=.76) ax.plot(np.arange(-128, 0), values["input_tail"], color=INK, lw=1, label="_nolegend_") ax.plot(np.arange(128), values["unedited_forecast"], color=MUTED, lw=1.7, label="Unedited") ax.plot(np.arange(128), values["edited_forecast"], color=BLUE, lw=2, label="Edited") ax.axvline(0, color=INK, lw=1, ls=(0, (2, 3)), alpha=.6) ax.text(.035, .98, "Input", transform=ax.transAxes, ha="left", va="top", fontsize=14, color=INK, bbox=dict(facecolor="white", edgecolor="none", pad=2)) ax.text(.52, .98, "Forecast", transform=ax.transAxes, ha="left", va="top", fontsize=14, color=MUTED) ax.set_xlim(-128, 127) ax.set_xticks([-128, 0, 127]) ax.set_xlabel("Sample relative to forecast", labelpad=9, fontsize=16) ax.set_ylabel("Signal value", labelpad=8, fontsize=16) style_axis(ax) ax.grid(axis="y", color=MUTED, alpha=.12) ax.tick_params(labelsize=14) fig.suptitle("PANDA · linear output-head edit", y=.98, fontsize=18, color=INK, fontweight="bold") handles, labels = ax.get_legend_handles_labels() fig.legend(handles, labels, loc="upper center", bbox_to_anchor=(.60, .90), ncol=2, frameon=False, fontsize=16, columnspacing=1.5, handlelength=1.8) return save(fig, directory, "panda-cycle") def control_panel(data, directory, objective): stem = f"{objective}-control" fig, ax = plt.subplots(figsize=DIMENSIONS[stem]) fig.subplots_adjust(left=.23, right=.96, bottom=.22, top=.82) rows = [row for row in data["control"]["series"] if row["objective"] == objective] for index, method in enumerate(("smwm", "pca", "direct")): selected = sorted((row for row in rows if row["method"] == method), key=lambda row: row["seed"]) y = [float(np.median(np.asarray(row["values"], dtype=np.float32))) for row in selected] ax.scatter(index + np.array([-.09, 0, .09]), y, color=PALETTE[method], s=46, zorder=3) ax.axhline(1, color=MUTED, ls=(0, (3, 3)), lw=1.1) ax.text(2.34, 1.11, "No edit", color=MUTED, ha="right", va="bottom", fontsize=11) ax.set_yscale("log") ax.set_xlim(-.45, 2.45) # Both objectives share the physical no-edit normalization and axis range. ax.set_ylim(1e-4, 2) ax.yaxis.set_major_locator(FixedLocator([1e-4, 1e-3, 1e-2, 1e-1, 1])) ax.yaxis.set_major_formatter(FuncFormatter(lambda value, _: f"{value:g}")) ax.yaxis.set_minor_locator(NullLocator()) ax.set_xticks([0, 1, 2], ["SMWM", "PCA", "Direct\nPANDA"]) ax.set_ylabel("Median error / no edit", labelpad=10) style_axis(ax) ax.grid(axis="y", color=MUTED, alpha=.12) title = "Reachable activation target" if objective == "activation" else "Requested waveform target" fig.suptitle(title, y=.97, fontsize=18, color=INK, fontweight="bold") fig.text(.57, .865, "Actual PANDA outcomes · lower is better", ha="center", fontsize=12, color=MUTED) fig.text(.57, .035, "3 seeds · the same 128 held-out inputs", ha="center", fontsize=12, color=MUTED) return save(fig, directory, stem) def generate(data_path=HERE / "figure-data.json", output_dir=HERE): output_dir = Path(output_dir) output_dir.mkdir(parents=True, exist_ok=True) data = load_data(data_path) matplotlib.rcParams.update({"font.family": "DejaVu Sans", "font.size": 13, "axes.labelsize": 14, "xtick.labelsize": 13, "ytick.labelsize": 12, "axes.labelcolor": INK, "text.color": INK, "pdf.fonttype": 42, "path.simplify": False}) paths = toy_panel(data, output_dir) + cycle_panel(data, output_dir) paths += control_panel(data, output_dir, "activation") paths += control_panel(data, output_dir, "waveform") return paths if __name__ == "__main__": parser = argparse.ArgumentParser(description=__doc__) parser.add_argument("--data", type=Path, default=HERE / "figure-data.json") parser.add_argument("--output-dir", type=Path, default=HERE) args = parser.parse_args() for path in generate(args.data, args.output_dir): print(f"{path.name} {hashlib.sha256(path.read_bytes()).hexdigest()}")