item 21: GSM carry-cot 57.4% (5x prior best); control isolates whiteboard to drop-bucket; E2-N/A2 planned

PLAN_SELFPACED.md

@@ -51,6 +51,56 @@ protection); (c) easy-bucket damage < carry-always's (83→45 was the
 unprotected number). Hard-bucket *gain* over carry-always is hoped for,
 not predicted.
 
+### E2-L — the internalization ladder (scratchpad → pure latent loop)
+Goal: a loop that computes internally during generation with NO pauses
+and NO visible scratchpad — reached by curriculum, never trained cold
+(cold-trained answer-only carry already failed: 9.4% overall, old carry
+arm k=2,p=0 cell — a 3-token signal can't teach the whiteboard what to
+write). Rungs, each warm-started from the previous:
+  A  loop + pauses + visible terse scratchpad, dense verified-CoT CE
+     (item 21, running 2026-07-16; control = same supervision, no
+     recurrence — the A-vs-B delta is the gate for everything below)
+  B  delete scratchpad steps one at a time, each replaced by extra
+     pauses; brief retrain per rung — visible computation forced onto
+     the pause-chain
+  C  pauses only, answer out (latent again, curriculum-reached)
+  C' no pauses either: state carries across answer tokens alone — the
+     pure internal loop
+Deliverable: the rung where accuracy breaks = measured capacity of this
+recurrence budget to absorb computation (the paper's number). Proceed
+past A only if arm A beats its control by >= 3 points overall
+(pre-registered, item 21b); A ~= B means the scratchpad text carries
+everything and internalization would only rediscover the C' failure.
+
+### E2-A2 — on-policy refresh (iterated self-distillation)
+The exposure gap = training prefixes vs deployment prefixes. Cheapest
+approximation ladder: (1) self-distilled scratchpads (stage A, done);
+(2) THIS: re-harvest scratchpads with the CURRENT adapter active each
+round, verify, retrain (STaR/ReST; DAgger at solution granularity;
+~3 min/harvest). Signature of working: verified-yield and eval accuracy
+co-improve across rounds. Gated on the A-vs-B verdict.
+
+### E2-N — lens-shaped state noise (Nils's idea, 2026-07-16 ~04:15)
+Harden the whiteboard against its own drift by injecting noise into the
+carried state during teacher-forced training — SHAPED by the J-lens
+instead of isotropic:
+  N1 sensitivity-weighted: sample noise in the span of J̄'s top-r right-
+     singular directions at the band entrance (the directions the final
+     readout depends on; isotropic noise wastes signal on the null
+     space). Cheap: jbar.pt exists; --lensnoise rank,scale flag.
+  N2 empirical-drift-matched: measure REAL exposure drift (free-run
+     state minus teacher-forced state at matched positions, few
+     rollouts), fit low-rank covariance, train under samples from it.
+     The lens diagnoses what the drift directions encode — worth running
+     as pure diagnosis regardless of verdicts (paper figure).
+  N3 concept-jitter: lens-read the carried concept (e.g. the
+     intermediate "24"), perturb toward a confusable concept in
+     embedding basis (swap machinery exists from the reproduction);
+     trains re-derivation over blind trust. Most ambitious.
+Caveats, stated in advance: J̄ is prompt-averaged (N1 directions are
+global, not per-position); noise norm-matched and magnitude-swept;
+whole line gated on arm A beating its control.
+
 ### E3 — power knob (only if E1 or E2 shows clean gating)
 Warm-start rung-2 band-LoRA under the gate; joint fine-tune. Question: do
 gate-bought iterations do MORE per iteration with a trainable band?

results-loop/PROTOCOL_UNIFIED.md

@@ -494,3 +494,28 @@ no fixed k solves but some k does.
     Honest note: arm outputs are VISIBLE tokens (~40) — this is the
     budget-CoT-with-loop hybrid, a scope change from latent planning,
     run at Nils's explicit direction ("do gsm8k and such").
+
+--- Outcome, item 21 (scored 2026-07-16 ~06:30). Harvest: 427 verified
+terse CoTs (292 hard, 135 easy; 73.5% yield, ~3 min). Grid (n=256,
+base 10.9/93.1/0.8/0.0 overall/easy/hard/drop):
+  arm A carry  2:2  53.9  (82.8 / 60.6 / 37.0)
+  arm A carry  2:6  57.4  (72.4 / 63.8 / 45.0)
+  control FF   2:2  49.6  (79.3 / 59.8 / 28.0)
+  control FF   2:6  54.7  (82.8 / 66.9 / 31.0)
+(a) CONFIRMED, dramatically: best cell 57.4% vs the previous best GSM
+arm's 12.1% — a 5x jump to full-CoT territory (~53%) at ~1/4 the
+visible tokens. Dense verified supervision was indeed the binding
+constraint; the supervision-density theory of the GSM failure is now
+POSITIVELY confirmed, not just by absence. (b) MIXED: A-vs-B delta
++4.3 at p=2 (clears the >=3 bar), +2.7 at p=6 (misses); hard/easy
+shuffle within noise between arms, BUT the whiteboard shows a
+consistent, specific signature: DROP items (unreachable by the base
+model even with full CoT at labeling) — A beats B by +9 and +14 points
+there in the two cells. Interpretation: scratchpad supervision carries
+the bulk; the carry chain specifically extends reach into previously
+unreachable problems. McNemar p-values pending (mcnemar_carrycot.py
+ready; classifier-blocked at scoring time). (c) CONFIRMED: easy damage
+much reduced vs answer-only carry (83->72-83 vs 83->45). Ladder gate:
+technically met at p=2; decision on stage B/A2/E2-N deferred to the
+morning review with the p-values in hand — the drop-bucket signature,
+if it survives pairing, is the strongest argument for continuing.

scripts/mcnemar_carrycot.py

@@ -0,0 +1,26 @@
+"""Paired McNemar: carry-cot arm vs feedforward control, both grid cells."""
+import json
+from math import comb
+from pathlib import Path
+
+OUT = Path(__file__).resolve().parent.parent / "results-loop"
+A = json.load(open(OUT / "eval_gsm_carrycot_e400.json"))
+B = json.load(open(OUT / "eval_gsm_carrycot_ff_e400.json"))
+star = {it["idx"]: it["label"]
+        for it in json.load(open(OUT / "star_data.json"))
+        if it["split"] == "test"}
+for cell in ("2:2", "2:6"):
+    a = {r["idx"]: r["ok"] for r in A["grid"][cell]["per_item"]}
+    b = {r["idx"]: r["ok"] for r in B["grid"][cell]["per_item"]}
+    ids = list(a)
+    x = sum(1 for i in ids if a[i] and not b[i])
+    y = sum(1 for i in ids if b[i] and not a[i])
+    n = x + y
+    p = (min(1, sum(comb(n, k) for k in range(max(x, y), n + 1))
+             / 2 ** n * 2) if n else 1)
+    dx = sum(1 for i in ids
+             if star.get(i) == "drop" and a[i] and not b[i])
+    dy = sum(1 for i in ids
+             if star.get(i) == "drop" and b[i] and not a[i])
+    print(f"{cell}: carry-only={x} ff-only={y}  McNemar p={p:.4f}  "
+          f"drop-bucket discordants {dx}-{dy}")

scripts/train_carry_cot.py

@@ -38,11 +38,42 @@ P_BY_LABEL = {"easy": 2, "hard": 6}
 ap = argparse.ArgumentParser()
 ap.add_argument("--feedforward", action="store_true")
 ap.add_argument("--seed", type=int, default=0)
+ap.add_argument("--lensnoise", default=None, metavar="RANK,SCALE",
+                help="E2-N1: inject noise into the carried state during "
+                     "training, shaped by the top-RANK sensitivity "
+                     "directions of jbar at the band entrance; noise norm "
+                     "= SCALE * per-position state norm (e.g. 32,0.05)")
 ARGS = ap.parse_args()
 TAG = ("carrycot_ff" if ARGS.feedforward else "carrycot") + (
+    f"_ln{ARGS.lensnoise.replace(',', '_')}" if ARGS.lensnoise else "") + (
     f"_s{ARGS.seed}" if ARGS.seed else "")
 
 
+class LensNoiseWrapper(torch.nn.Module):
+    """Perturb the carried state s (not the anchor e) before the merge,
+    within the span of the lens's top-r readout-sensitive directions.
+    Train-time only (noise_on flag); eval and checkpoints use .base."""
+
+    def __init__(self, base, jbar_layer, rank, scale):
+        super().__init__()
+        self.base = base
+        self.scale = scale
+        self.noise_on = True
+        J = jbar_layer.float()
+        _, _, Vt = torch.linalg.svd(J, full_matrices=False)
+        self.register_buffer("V", Vt[:rank].T.contiguous())  # (d, r)
+
+    def forward(self, e, s):
+        if self.noise_on and self.scale > 0:
+            z = torch.randn(*s.shape[:-1], self.V.shape[1],
+                            device=s.device, dtype=torch.float32)
+            n = z @ self.V.T
+            n = n * (s.float().norm(dim=-1, keepdim=True) * self.scale
+                     / (n.norm(dim=-1, keepdim=True) + 1e-6))
+            s = (s.float() + n).to(s.dtype)
+        return self.base(e, s)
+
+
 def lr_at(step):
     if step < WARMUP:
         return LR * (step + 1) / WARMUP
@@ -107,6 +138,15 @@ def main():
     looper = BandLooper(model)
     adapter = MergeAdapter(
         d=model.config.get_text_config().hidden_size).cuda()
+    if ARGS.lensnoise:
+        r, sc = ARGS.lensnoise.split(",")
+        jbar = torch.load(Path(__file__).resolve().parent.parent
+                          / "results/jbar.pt", map_location="cpu")["Jbar"]
+        from loop_common import BAND
+        adapter = LensNoiseWrapper(adapter, jbar[BAND[0]], int(r),
+                                   float(sc)).cuda()
+        print(f"lens-noise: rank={r} scale={sc} on jbar L{BAND[0]}",
+              flush=True)
     opt = torch.optim.AdamW(adapter.parameters(), lr=LR, weight_decay=0.01)
 
     keep = [it for it in data
@@ -146,11 +186,15 @@ def main():
             print(f"step {step:4d}  {lbl:4s} loss={loss.item():.4f}  "
                   f"({(time.time()-t0)/(step+1):.1f}s/step)", flush=True)
         if step % 200 == 199 or step == STEPS - 1:
+            if ARGS.lensnoise:
+                adapter.noise_on = False
             for l in ("easy", "hard"):
                 v = val_loss(looper, adapter, tok, val[l], P_BY_LABEL[l])
                 print(f"  val@{step}: {l}={v:.3f}", flush=True)
-            torch.save(adapter.state_dict(),
-                       OUT / f"adapter_{TAG}_e{step+1}.pt")
+            if ARGS.lensnoise:
+                adapter.noise_on = True
+            sd = (adapter.base if ARGS.lensnoise else adapter).state_dict()
+            torch.save(sd, OUT / f"adapter_{TAG}_e{step+1}.pt")
             json.dump(log, open(OUT / f"train_{TAG}_log.json", "w"))
     print("done", flush=True)