diff --git a/results-loop/fig_loop_vs_ff.png b/results-loop/fig_loop_vs_ff.png
new file mode 100644
index 0000000..e4e5d2e
Binary files /dev/null and b/results-loop/fig_loop_vs_ff.png differ
diff --git a/scripts/fig_loop_vs_ff.py b/scripts/fig_loop_vs_ff.py
new file mode 100644
index 0000000..6837e62
--- /dev/null
+++ b/scripts/fig_loop_vs_ff.py
@@ -0,0 +1,60 @@
+"""Ladder: what the loop adds over the base model and the same implant
+applied once (feedforward). Same 250-item MBPP eval throughout.
+Per-arm colors follow fig_phase/fig_kcurves; easy pastel, hard dark,
+overall neutral.
+"""
+from pathlib import Path
+
+import matplotlib
+matplotlib.use("Agg")
+import matplotlib.pyplot as plt
+
+OUT = Path(__file__).resolve().parent.parent / "results-loop"
+
+#        arm                                       overall easy   hard   color
+ROWS = [
+    ("base model (k=0)",                            0.488, 0.984, 0.036, "#8a8f98"),
+    ("+ FF implant\n(same adapter, applied once)",  0.496, 0.934, 0.179, "#8a8f98"),
+    ("+ untrained loop (best k)",                   0.476, 0.892, 0.179, "#8a8f98"),
+    ("+ trained loop, random-depth\n(randk, k=8)",  0.512, 0.902, 0.393, "#2b6cb0"),
+    ("+ trained loop, curriculum\n(merge, k=4)",    0.512, 0.885, 0.464, "#2b6cb0"),
+    ("+ plan-distilled implant\n(no recurrence; 8-seed mean)*", 0.555, 0.919, 0.457, "#b7791f"),
+]
+
+fig, ax = plt.subplots(figsize=(8.6, 4.6))
+H = 0.22
+for i, (name, ov, easy, hard, c) in enumerate(ROWS):
+    y = len(ROWS) - 1 - i
+    ax.barh(y + H, hard, height=H, color=c, zorder=3)
+    ax.barh(y, easy, height=H, color=c, alpha=0.32, zorder=3)
+    ax.barh(y - H, ov, height=H, color="#c9ccd1", zorder=3)
+    for v, dy in ((hard, H), (easy, 0), (ov, -H)):
+        ax.text(v + 0.012, y + dy, f"{v:.1%}", va="center", fontsize=8,
+                color="#333")
+ax.set_yticks([len(ROWS) - 1 - i for i in range(len(ROWS))])
+ax.set_yticklabels([r[0] for r in ROWS], fontsize=9)
+ax.set_xlim(0, 1.09)
+ax.set_xlabel("pass@1", fontsize=9)
+ax.xaxis.set_major_formatter(lambda x, _: f"{x:.0%}")
+
+import matplotlib.patches as mpatches
+ax.legend(handles=[
+    mpatches.Patch(color="#555a61", label="hard (plan-dependent, n=28)"),
+    mpatches.Patch(color="#555a61", alpha=0.32, label="easy (n=122)"),
+    mpatches.Patch(color="#c9ccd1", label="overall (n=250)"),
+], fontsize=8, frameon=False, loc="lower right")
+
+ax.grid(True, axis="x", color="#ececec", lw=0.7)
+ax.set_axisbelow(True)
+for s in ("top", "right"):
+    ax.spines[s].set_visible(False)
+ax.set_title("What recurrence adds — and what distillation matches without it\n"
+             "(MBPP; FF = trained adapter without recurrence)", fontsize=10.5)
+fig.text(0.13, -0.015, "*distill rows evaluated on the 500-item set (hard "
+         "n=55, base hard 3.6-5.5%); all other rows on the 250-item set. "
+         "Distill trains the same adapter on written plans, deploys with "
+         "zero loops.", fontsize=7.5, color="#555")
+fig.tight_layout()
+fig.savefig(OUT / "fig_loop_vs_ff.png", dpi=140, facecolor="white",
+            bbox_inches="tight")
+print("wrote", OUT / "fig_loop_vs_ff.png")