"""Draw this note's cover from the eval-awareness direction itself, with the shared steerlab art style.

    python experiments/natural-sandbagging/make_art.py

Same picture rules as experiments/shared/make_art.py: the direction's 4,096 numbers (one per dimension of Qwen3-8B's
residual stream at the steered layer) are laid on a square grid and smoothed into a flow field; streamlines run
straight until the vertical rule and then follow the field. Nothing is drawn by hand. The direction is the
eval-minus-deployment difference of means at layer 8, results/direction.npz.
"""

from __future__ import annotations

import json
import sys
from pathlib import Path

import numpy as np

HERE = Path(__file__).resolve().parent
sys.path.insert(0, str(HERE.parent / "shared"))
import make_art  # noqa: E402


def direction_field(_model: str, _concept: str) -> np.ndarray:
    layer = json.loads((HERE / "results" / "direction.json").read_text())["chosen_layer"]
    vector = np.load(HERE / "results" / "direction.npz")[f"u_L{layer}"].astype(float)
    side = int(np.sqrt(vector.size))
    grid = vector[: side * side].reshape(side, side)
    for _ in range(6):  # the shared script's cheap Gaussian: repeated 3x3 box blur
        padded = np.pad(grid, 1, mode="wrap")
        grid = sum(padded[i : i + side, j : j + side] for i in range(3) for j in range(3)) / 9
    return grid / np.abs(grid).max()


if __name__ == "__main__":
    make_art.ART["natural-sandbagging"] = ("qwen8b", "eval-awareness direction", 0.16, 1.7)
    make_art.field = direction_field
    make_art.draw("natural-sandbagging", (1600, 900), HERE / "report" / "figures" / "data-art.jpg")