J-lens workspace reproduction + loop retrofit: lens, band looping, adapters, controls, multi-task evals

Reproduction of the 2026 workspace/J-lens paper on gemma-4 (E2B/12B/26B),
plus the workspace-loop retrofit line: merge adapter, prompt-only latent
planning (MBPP), carry variant, attribution controls (FF/pause/untrained),
band-location ablation, Blocksworld harness, 12B replication scripts.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

scripts/plot_loop_eval.py

@@ -0,0 +1,57 @@
+"""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")