experiments: freeze normalized C2 recovery

3 files changed, 196 insertions, 0 deletions

diff --git a/ROADMAP.md b/ROADMAP.md
index 75efa60..a28f9af 100644
--- a/ROADMAP.md
+++ b/ROADMAP.md
@@ -201,6 +201,16 @@ mean d4 validation from `78.633 ± 24.866%` to `81.867 ± 8.779%` and `82.450 ±
 advances and task seeds 6--8 remain unused. Timescale separation reduces basin variance but does
 not make plain SGD regression of the moving apical map accurate enough.
 
+One final task-seed-0 regression-mechanism screen is allowed before stopping C2 development.
+Normalized LMS divides each local A update by the squared norm of its complete instructional
+feature `[c, c outer tanh(h_top)]`, addressing the angle/gain separation in C5 without enlarging
+the vectorizer. It fixes the A-only prefix to 100 layerwise-K4 steps, keeps the joint phase at
+simultaneous K1/e4, and compares only `eta_A in {0.01, 0.05}` over d1/d4 and model seeds 0--2.
+Eligibility is identical to the feedback-first screen: all losses finite, d4 mean at least 90%, at
+least 2/3 positive depth gains, and lesion mean at least 2 points with at least 2/3 positive. Among
+eligible candidates select the highest d4 mean, preferring `eta_A=0.01` within one point. Failure
+closes further tent-map tuning; success permits one frozen confirmation on task seeds 6--8.
+
 ### C3. Causal calibration is genuinely amortized
 
 **Status: passed on 2026-07-22.** The frozen five-seed confirmation retained 112.9% of the K16/e4
diff --git a/experiments/analyze_c2_nlms_pilot.py b/experiments/analyze_c2_nlms_pilot.py
new file mode 100644
index 0000000..f7ed39a
--- /dev/null
+++ b/experiments/analyze_c2_nlms_pilot.py
@@ -0,0 +1,136 @@
+"""Audit and select the frozen normalized-regression C2 development screen."""
+import glob
+import json
+import math
+import os
+import statistics
+
+
+ROOT = os.path.join(os.path.dirname(os.path.dirname(os.path.abspath(__file__))), "results")
+PREFIX = "c2_nlmsdev_v1_"
+ETA_AS = (0.01, 0.05)
+DEPTHS = (1, 4)
+MODEL_SEEDS = (0, 1, 2)
+
+
+def audit_row(path, row):
+    args = row["args"]
+    required = {
+        "mode": "sdil", "dataset": "tentmap", "width": 8,
+        "act": "relu", "residual": 1, "vectorizer_mode": "context_gated",
+        "vectorizer_optimizer": "nlms", "vectorizer_eps": 1e-6,
+        "epochs": 80, "batch_size": 256, "eta": 0.03, "momentum": 0.9,
+        "eta_P": 0.002, "pert_sigma": 0.01, "pert_every": 4,
+        "pert_ndirs": 1, "pert_mode": "simultaneous",
+        "a_warmup_steps": 100, "a_warmup_ndirs": 4,
+        "a_warmup_mode": "layerwise", "traffic_mode": "none",
+        "nuis_rho": 0.0, "use_residual": 1, "learn_A": 1, "learn_P": 1,
+        "p_neutral": 1, "task_train_examples": 10000,
+        "task_test_examples": 5000, "task_levels": 2, "task_n_in": 1,
+        "task_seed": 0, "val_examples": 2000, "split_seed": 2027,
+        "eval_split": "validation", "eval_every": 0,
+        "diagnostics": "alignment", "diagnostics_schedule": "final",
+        "probe_bs": 512,
+    }
+    mismatches = {key: (args.get(key), expected) for key, expected in required.items()
+                  if args.get(key) != expected}
+    if args.get("eta_A") not in ETA_AS:
+        mismatches["eta_A"] = (args.get("eta_A"), ETA_AS)
+    expected_lesion = 1.0 / 3.0 if args["depth"] == 4 else 0.0
+    if abs(args.get("residual_lesion_fraction", 0.0) - expected_lesion) > 1e-12:
+        mismatches["residual_lesion_fraction"] = (
+            args.get("residual_lesion_fraction"), expected_lesion)
+    if mismatches:
+        raise RuntimeError(f"protocol mismatch {path}: {mismatches}")
+    if row["final"].get("eval_split") != "validation":
+        raise RuntimeError(f"test-contaminated development row: {path}")
+    if any("eval_acc" in step or "cos_r_negg" in step for step in row.get("steps", [])):
+        raise RuntimeError(f"intermediate held-out metric/diagnostic: {path}")
+    split = row.get("split", {})
+    if (not split.get("split_from_training_only")
+            or split.get("validation_examples") != 2000
+            or split.get("evaluation_split") != "validation"):
+        raise RuntimeError(f"invalid validation split {path}: {split}")
+    if row.get("provenance", {}).get("git_dirty") is not False:
+        raise RuntimeError(f"dirty or unknown source provenance: {path}")
+    if row.get("cost", {}).get("feedback_warmup_perturbation_events") != 100:
+        raise RuntimeError(f"warmup cost mismatch {path}: {row.get('cost')}")
+    if args["depth"] == 4:
+        lesion = row["final"].get("residual_lesion")
+        if not lesion or lesion.get("lesioned_layers") != [3]:
+            raise RuntimeError(f"incorrect d4 lesion: {path}")
+
+
+def mean_sd(values):
+    return statistics.mean(values), statistics.stdev(values)
+
+
+def main():
+    paths = sorted(glob.glob(os.path.join(ROOT, PREFIX + "*.json")))
+    rows = {}
+    commits = set()
+    split_hashes = set()
+    for path in paths:
+        with open(path) as handle:
+            row = json.load(handle)
+        audit_row(path, row)
+        args = row["args"]
+        key = (args["eta_A"], args["depth"], args["seed"])
+        if key in rows:
+            raise RuntimeError(f"duplicate row: {key}")
+        rows[key] = row
+        commits.add(row["provenance"]["git_commit"])
+        split_hashes.add(row["split"]["validation_index_sha256"])
+    expected = {(eta_a, depth, seed) for eta_a in ETA_AS
+                for depth in DEPTHS for seed in MODEL_SEEDS}
+    if set(rows) != expected or len(commits) != 1 or len(split_hashes) != 1:
+        raise RuntimeError(f"incomplete/mixed pilot: rows={len(rows)}, "
+                           f"missing={expected - set(rows)}, extra={set(rows) - expected}, "
+                           f"commits={commits}, split_hashes={split_hashes}")
+
+    summaries = {}
+    print(f"commit={next(iter(commits))} rows={len(rows)} task_seed=0")
+    print("| eta_A | depth | validation (%) | depth gain | d4 lesion | d4 work/ordinary |")
+    print("|---:|---:|---:|---:|---:|---:|")
+    for eta_a in ETA_AS:
+        accs = {depth: [100 * rows[(eta_a, depth, seed)]["final"]["val_acc"]
+                        for seed in MODEL_SEEDS] for depth in DEPTHS}
+        gains = [deep - shallow for shallow, deep in zip(accs[1], accs[4])]
+        nonfinite = [seed for depth in DEPTHS for seed in MODEL_SEEDS
+                     if not math.isfinite(rows[(eta_a, depth, seed)]["final"]
+                                          .get("val_loss", math.nan))]
+        lesions = [100 * rows[(eta_a, 4, seed)]["final"]["residual_lesion"]
+                   ["lesion_acc_drop"] for seed in MODEL_SEEDS]
+        work_ratios = [rows[(eta_a, 4, seed)]["cost"]
+                       ["training_forward_equivalent_examples"]
+                       / rows[(eta_a, 4, seed)]["cost"]
+                       ["ordinary_training_forward_examples"] for seed in MODEL_SEEDS]
+        eligible = (not nonfinite and statistics.mean(accs[4]) >= 90.0
+                    and sum(value > 0 for value in gains) >= 2
+                    and statistics.mean(lesions) >= 2.0
+                    and sum(value > 0 for value in lesions) >= 2)
+        summaries[eta_a] = {"d4_mean": statistics.mean(accs[4]), "eligible": eligible}
+        for depth in DEPTHS:
+            mean, sd = mean_sd(accs[depth])
+            gain_text = "--" if depth == 1 else f"{statistics.mean(gains):+.3f}"
+            lesion_text = "--" if depth == 1 else f"{statistics.mean(lesions):+.3f}"
+            work_text = "--" if depth == 1 else f"{statistics.mean(work_ratios):.1f}x"
+            print(f"| {eta_a:.2f} | {depth} | {mean:.3f} +/- {sd:.3f} | "
+                  f"{gain_text} | {lesion_text} | {work_text} |")
+        print(f"eta_A={eta_a:.2f} positive gains={sum(value > 0 for value in gains)}/3; "
+              f"positive lesions={sum(value > 0 for value in lesions)}/3; "
+              f"nonfinite={nonfinite}; eligible={eligible}")
+
+    eligible = [eta_a for eta_a in ETA_AS if summaries[eta_a]["eligible"]]
+    if not eligible:
+        print("C2 normalized-regression development: NO ELIGIBLE CANDIDATE")
+        raise SystemExit(1)
+    best_mean = max(summaries[eta_a]["d4_mean"] for eta_a in eligible)
+    selected = min(eta_a for eta_a in eligible
+                   if summaries[eta_a]["d4_mean"] >= best_mean - 1.0)
+    print(f"C2 normalized-regression development selection: eta_A={selected:.2f} "
+          f"(best d4={best_mean:.3f}%, selected d4={summaries[selected]['d4_mean']:.3f}%)")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/experiments/c2_nlms_pilot.sh b/experiments/c2_nlms_pilot.sh
new file mode 100755
index 0000000..43a3d4a
--- /dev/null
+++ b/experiments/c2_nlms_pilot.sh
@@ -0,0 +1,50 @@
+#!/usr/bin/env bash
+# Final bounded task-seed-0 C2 recovery screen. The context vectorizer uses
+# normalized LMS in both a 100-step A-only layerwise-K4 prefix and the original
+# simultaneous-K1/e4 joint phase. Compare eta_A 0.01/0.05 only.
+# Usage: c2_nlms_pilot.sh <gpu> "<model seeds>"
+set -eu
+
+cd "$(dirname "$0")/.."
+GPU="${1:?GPU index required}"
+MODEL_SEEDS="${2:-0 1 2}"
+PYTHON="${PYTHON:-/home/yurenh2/miniconda3/envs/ep_pascal/bin/python3}"
+
+for eta_a in 0.01 0.05; do
+  case "$eta_a" in
+    0.01) eta_tag=0p01 ;;
+    0.05) eta_tag=0p05 ;;
+    *) echo "unexpected eta_A: $eta_a" >&2; exit 2 ;;
+  esac
+  for model_seed in $MODEL_SEEDS; do
+    for depth in 1 4; do
+      tag="c2_nlmsdev_v1_tent_l2_w8_context_e${eta_tag}_d${depth}_t0_s${model_seed}"
+      out="results/${tag}.json"
+      if [ -s "$out" ]; then
+        echo "[$tag] exists; skipping"
+        continue
+      fi
+      lesion=0
+      if [ "$depth" -eq 4 ]; then
+        lesion=0.3333333333333333
+      fi
+      CUDA_VISIBLE_DEVICES="$GPU" "$PYTHON" experiments/run.py \
+        --mode sdil --dataset tentmap --device cuda \
+        --depth "$depth" --width 8 --act relu --residual 1 \
+        --residual_lesion_fraction "$lesion" \
+        --vectorizer_mode context_gated --vectorizer_optimizer nlms \
+        --epochs 80 --batch_size 256 --eta 0.03 --momentum 0.9 \
+        --eta_A "$eta_a" --eta_P 0.002 \
+        --pert_sigma 0.01 --pert_every 4 --pert_ndirs 1 \
+        --pert_mode simultaneous \
+        --a_warmup_steps 100 --a_warmup_ndirs 4 --a_warmup_mode layerwise \
+        --traffic_mode none --nuis_rho 0 \
+        --use_residual 1 --learn_A 1 --learn_P 1 --p_neutral 1 \
+        --task_train_examples 10000 --task_test_examples 5000 \
+        --task_levels 2 --task_n_in 1 --task_seed 0 \
+        --val_examples 2000 --split_seed 2027 --eval_split validation --eval_every 0 \
+        --diagnostics alignment --diagnostics_schedule final --probe_bs 512 \
+        --seed "$model_seed" --log_every 100000 --tag "$tag"
+    done
+  done
+done