PerDepthAdapter (Bae-style per-iteration merges) + convergence-halting probe (free ACT); pre-registration item 13

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

scripts/eval_loop_code.py

@@ -28,8 +28,9 @@ OUT = Path(os.environ.get("LOOP_OUT",
 
 @torch.no_grad()
 def pass1_at_k(looper, adapter, tok, items, k, batch=8, max_new=220,
-               feedforward=False, pause=0):
+               feedforward=False, pause=0, halt=False):
     codes = []
+    k_convs = []
     for i in range(0, len(items), batch):
         torch.cuda.empty_cache()
         chunk = items[i : i + batch]
@@ -42,10 +43,14 @@ def pass1_at_k(looper, adapter, tok, items, k, batch=8, max_new=220,
             enc["input_ids"] = torch.cat([enc["input_ids"], pcol], 1)
             enc["attention_mask"] = torch.cat(
                 [enc["attention_mask"], torch.ones_like(pcol)], 1)
+        cv = {} if halt else None
         gen = looper.generate_frozen_prompt(adapter, tok, enc["input_ids"], k,
                                             max_new_tokens=max_new,
                                             attention_mask=enc["attention_mask"],
-                                            feedforward=feedforward)
+                                            feedforward=feedforward,
+                                            conv_out=cv)
+        if halt:
+            k_convs.extend(cv.get("k_conv", []))
         for j in range(len(chunk)):
             txt = tok.decode(gen[j, enc["input_ids"].shape[1]:],
                              skip_special_tokens=True)
@@ -61,6 +66,9 @@ def pass1_at_k(looper, adapter, tok, items, k, batch=8, max_new=220,
         d[1] += 1
     per_item = [{"task_id": it["task_id"], "ok": bool(ok)}
                 for it, ok in zip(items, oks)]
+    if halt and k_convs:
+        for r, kc in zip(per_item, k_convs):
+            r["k_conv"] = kc
     return (hits / len(items),
             {l: c / n for l, (c, n) in per_label.items()}, per_item)
 
@@ -81,13 +89,19 @@ def main():
                     help="RecurrentAdapter (noise h0, learned A/B)")
     ap.add_argument("--parcae", action="store_true",
                     help="ParcaeAdapter (rho(A)<1 by construction)")
+    ap.add_argument("--perdepth", action="store_true",
+                    help="PerDepthAdapter (Bae-style per-iteration merges)")
+    ap.add_argument("--halt", action="store_true",
+                    help="record per-item convergence depth (free-ACT probe)")
     args = ap.parse_args()
     ks = [int(x) for x in args.ks.split(",")]
 
     model, tok = load_model(dtype=torch.bfloat16)
     tok.padding_side = "left"
     looper = BandLooper(model)
-    cls = (ParcaeAdapter if args.parcae
+    from loop_common import PerDepthAdapter
+    cls = (PerDepthAdapter if args.perdepth
+           else ParcaeAdapter if args.parcae
            else RecurrentAdapter if args.rec
            else AdaptiveMergeAdapter if args.adaptive else MergeAdapter)
     kw = ({"alpha0": args.alpha} if args.adaptive else {"alpha": args.alpha})
@@ -106,7 +120,13 @@ def main():
         t0 = time.time()
         acc, by_label, per_item = pass1_at_k(looper, adapter, tok, items, k,
                                              feedforward=args.feedforward,
-                                             pause=args.pause)
+                                             pause=args.pause, halt=args.halt)
+        if args.halt:
+            by_lbl_k = {}
+            for it, r in zip(items, per_item):
+                by_lbl_k.setdefault(it["label"], []).append(r.get("k_conv", k))
+            print("  mean k_conv:", {l: round(sum(v)/len(v), 2)
+                                     for l, v in by_lbl_k.items()}, flush=True)
         res["ks"][k] = {"acc": acc, "by_label": by_label,
                         "per_item": per_item}
         print(f"k={k}: pass@1={acc:.3f}  "