evil_MoE/scripts/plot_deploy_overlay.py

"""All-arms per-mode DEPLOY overlay (#162) from the per_mode_deploy.json artifacts.

Each run writes out/runs/<ts>_<tag>/per_mode_deploy.json (train.py, #164) with the
HONEST deploy numbers: for route/route2 the quarantine is deleted before eval, so
this is the model you would actually ship -- unlike plot_substrate's hk_<mode>
curves which are TRAIN-time (routed forward still hacks) and overstate routing.

Reads JSON, not logs, so it never trips on a route2 arm the log-parsers don't know.

The headline comparison: per loophole mode, does each intervention suppress the
DEPLOY hack rate below vanilla, and at what cost to DEPLOY solve? run_tests is the
in-dist mode (v_hack built closest to it); the rest are held-out (the no-cheat
generalisation test). Bars grouped by mode, one bar per arm.

Usage:
  uv run python scripts/plot_deploy_overlay.py                       # globs out/runs/*sub4*/
  uv run python scripts/plot_deploy_overlay.py out/runs/*_sub4_*/per_mode_deploy.json
  uv run python scripts/plot_deploy_overlay.py --out out/figs/deploy_overlay.png
"""
from __future__ import annotations

import argparse
import json
from pathlib import Path

import matplotlib.pyplot as plt
import numpy as np
from loguru import logger

from projected_grpo.figs import save_fig

# arm -> (display label, colour). Order = legend/bar order (baseline first).
# Reader-facing names only -- "route2"/"grad" are internal tags. The grad-mask
# routing arm is the one we report, so it is plain "route"; the failed
# activation-mask variant is disambiguated, not version-numbered.
ARM = {
    "vanilla":      ("vanilla",            "#444444"),
    "projected":    ("erase",              "#c1432b"),
    "routing":      ("route (v1)",         "#33508c"),
    "routing2_act": ("route (act-mask)",   "#2f7d4f"),
    "routing2_grad":("route",              "#b8860b"),
}
# mode display order: in-dist first, then held-out.
MODE_ORDER = ["run_tests", "file_marker", "stdout_marker", "sentinel", "eq_override"]


def load(paths: list[Path]) -> list[dict]:
    out = []
    for p in paths:
        d = json.loads(p.read_text())
        out.append(d)
        logger.info(f"{d['arm']:<14} deploy hack={d['hack_deploy']:.3f} solve={d['solve_deploy']:.3f}  ({p})")
    return out


def _despine(ax):
    ax.spines[["top", "right"]].set_visible(False)
    ax.grid(axis="y", lw=0.4, alpha=0.35)


def _panel(ax, records, modes, arms, field, title, ylabel):
    """Grouped bars: x = mode, one bar per arm, height = records[arm].by_mode[mode][field]."""
    w = 0.8 / len(arms)
    x = np.arange(len(modes))
    for i, arm in enumerate(arms):
        rec = next(r for r in records if r["arm"] == arm)
        label, color = ARM[arm]
        vals = [rec["by_mode"].get(m, {}).get(field, np.nan) for m in modes]
        bars = ax.bar(x + i * w, vals, w, label=label, color=color)
        for b, v in zip(bars, vals):
            if not np.isnan(v):
                ax.annotate(f"{v:.2f}", (b.get_x() + b.get_width() / 2, v), fontsize=6,
                            ha="center", va="bottom", color=color)
    ax.set_xticks(x + 0.4 - w / 2)
    ax.set_xticklabels([f"{m}\n{'IN' if m == 'run_tests' else 'held-out'}" for m in modes], fontsize=8)
    ax.set_title(title, fontsize=10)
    ax.set_ylabel(ylabel)
    ax.set_ylim(0, 1.05)
    _despine(ax)


def main() -> None:
    ap = argparse.ArgumentParser(description=__doc__)
    ap.add_argument("jsons", nargs="*", type=Path,
                    help="per_mode_deploy.json paths; default globs out/runs/*sub4*/")
    ap.add_argument("--out", type=Path, default=Path("out/figs/deploy_overlay.png"))
    ap.add_argument("--title", action="store_true",
                    help="draw the suptitle (off by default: the caption carries it)")
    args = ap.parse_args()

    paths = args.jsons or sorted(Path("out/runs").glob("*sub4*/per_mode_deploy.json"))
    if not paths:
        raise SystemExit("no per_mode_deploy.json found (run the sweep first)")
    records = load(paths)
    # dedupe arms (keep latest by file order), then order canonically
    by_arm = {r["arm"]: r for r in records}
    arms = [a for a in ARM if a in by_arm]
    records = [by_arm[a] for a in arms]
    modes = [m for m in MODE_ORDER if any(m in r["by_mode"] for r in records)]

    fig, (a1, a2) = plt.subplots(1, 2, figsize=(5.5 + 1.2 * len(modes), 4.2))
    _panel(a1, records, modes, arms, "deploy_hack",
           "DEPLOY hack rate by mode (lower = better)", "deploy hack rate")
    _panel(a2, records, modes, arms, "deploy_solve",
           "DEPLOY solve rate by mode (higher = better)", "deploy solve rate")
    a1.legend(fontsize=8, frameon=False, loc="upper right")
    if args.title:
        n_seed = {r.get("seed") for r in records}
        fig.suptitle(f"Per-mode deploy overlay ({len(arms)} arms, seed {sorted(n_seed)}) -- "
                     f"quarantine deleted = shipped model", fontsize=11)
    fig.tight_layout()
    save_fig(fig, args.out)
    logger.info(f"wrote {args.out}  ({len(arms)} arms x {len(modes)} modes)")


if __name__ == "__main__":
    main()