fidelity factorial arms (randk / noises0 / hidden-capacity) + pre-registration item 15

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

scripts/eval_loop_code.py

@@ -93,6 +93,8 @@ def main():
                     help="PerDepthAdapter (Bae-style per-iteration merges)")
     ap.add_argument("--tiedalpha", action="store_true",
                     help="TiedAlphaAdapter (learned per-dim alpha, tied B)")
+    ap.add_argument("--noises0", action="store_true")
+    ap.add_argument("--hidden", type=int, default=512)
     ap.add_argument("--halt", action="store_true",
                     help="record per-item convergence depth (free-ACT probe)")
     args = ap.parse_args()
@@ -101,13 +103,17 @@ def main():
     model, tok = load_model(dtype=torch.bfloat16)
     tok.padding_side = "left"
     looper = BandLooper(model)
-    from loop_common import PerDepthAdapter, TiedAlphaAdapter
-    cls = (TiedAlphaAdapter if args.tiedalpha
+    from loop_common import (NoisyMergeAdapter, PerDepthAdapter,
+                             TiedAlphaAdapter)
+    cls = (NoisyMergeAdapter if args.noises0
+           else TiedAlphaAdapter if args.tiedalpha
            else 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})
+    if not args.adaptive:
+        kw["hidden"] = args.hidden
     adapter = cls(d=model.config.get_text_config().hidden_size, **kw).cuda()
     if args.adapter:
         adapter.load_state_dict(torch.load(args.adapter, map_location="cuda"))