"""Two-panel figure: accuracy vs loop depth k, and J-lens concept sharpening."""
import json
from pathlib import Path

import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt

OUT = Path(__file__).resolve().parent.parent / "results-loop"
tr = json.load(open(OUT / "eval_trained.json"))
un = json.load(open(OUT / "eval_untrained.json"))

ks = sorted(int(k) for k in tr["ks"])
BLUE, GRAY = "#2b6cb0", "#8a8f98"

fig, (ax1, ax2) = plt.subplots(1, 2, figsize=(11, 4.2))

def curve(res, sel):
    return [res["ks"][str(k)]["acc"] if sel == "all"
            else res["ks"][str(k)]["by_label"].get(sel, 0.0) for k in ks]

for ax in (ax1, ax2):
    ax.grid(True, color="#e5e5e5", lw=0.7, zorder=0)
    ax.set_axisbelow(True)
    for s in ("top", "right"):
        ax.spines[s].set_visible(False)
    ax.set_xticks(ks)
    ax.set_xlabel("loop depth k")

# --- panel 1: GSM8K accuracy vs k ---
ax1.plot(ks, curve(tr, "all"), "-o", color=BLUE, lw=2, ms=5, label="trained · all")
ax1.plot(ks, curve(un, "all"), "-o", color=GRAY, lw=2, ms=5, label="untrained · all")
ax1.plot(ks, curve(tr, "hard"), "--s", color=BLUE, lw=2, ms=5, label="trained · hard (CoT-only)")
ax1.plot(ks, curve(un, "hard"), "--s", color=GRAY, lw=2, ms=5, label="untrained · hard")
ax1.set_ylabel("accuracy (greedy, 256 held-out items)")
ax1.set_title("GSM8K accuracy vs loop depth", fontsize=11)
ax1.legend(fontsize=8, frameon=False)
ax1.annotate("hard items: 0.8% → 6.3%", xy=(2, 0.063), xytext=(3.2, 0.115),
             fontsize=8, color=BLUE,
             arrowprops=dict(arrowstyle="-", color=BLUE, lw=0.8))

# --- panel 2: J-lens sharpening ---
sp_tr = {r["k"]: r["P_spider_lens"] for r in tr["spider"]}
sp_un = {r["k"]: r["P_spider_lens"] for r in un["spider"]}
ax2.plot(ks, [sp_tr[k] for k in ks], "-o", color=BLUE, lw=2, ms=5, label="trained")
ax2.plot(ks, [sp_un[k] for k in ks], "-o", color=GRAY, lw=2, ms=5, label="untrained")
ax2.set_ylabel("P('spider') under J-lens at L30")
ax2.set_title("Latent concept sharpening across loops", fontsize=11)
ax2.legend(fontsize=8, frameon=False)
ax2.text(ks[-1], sp_tr[ks[-1]] + 0.006, "8× over control", fontsize=8,
         color=BLUE, ha="right")

fig.suptitle("Trained merge adapter: loop depth buys latent sharpening and some "
             "CoT-only answers, not overall accuracy", fontsize=12, y=1.02)
fig.tight_layout()
fig.savefig(OUT / "loop_eval.png", dpi=140, bbox_inches="tight", facecolor="white")
print("wrote", OUT / "loop_eval.png")