experiments: freeze C2 calibration quality recovery

2 files changed, 178 insertions, 0 deletions

diff --git a/experiments/analyze_c2_calibration_quality_pilot.py b/experiments/analyze_c2_calibration_quality_pilot.py
new file mode 100644
index 0000000..3233064
--- /dev/null
+++ b/experiments/analyze_c2_calibration_quality_pilot.py
@@ -0,0 +1,129 @@
+"""Audit and apply the frozen C2 calibration-quality development rule."""
+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_calibdev_v1_"
+DIRECTIONS = (4, 16)
+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",
+        "epochs": 80, "batch_size": 256, "eta": 0.03, "momentum": 0.9,
+        "eta_A": 0.01, "eta_P": 0.002, "pert_sigma": 0.01,
+        "pert_every": 4, "pert_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}
+    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 args.get("pert_ndirs") not in DIRECTIONS:
+        mismatches["pert_ndirs"] = (args.get("pert_ndirs"), DIRECTIONS)
+    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 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["pert_ndirs"], 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 = {(directions, depth, seed) for directions in DIRECTIONS
+                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("| K | depth | validation (%) | depth gain | d4 lesion | d4 work/ordinary |")
+    print("|---:|---:|---:|---:|---:|---:|")
+    for directions in DIRECTIONS:
+        accs = {depth: [100 * rows[(directions, 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[(directions, depth, seed)]["final"]
+                                          .get("val_loss", math.nan))]
+        lesions = [100 * rows[(directions, 4, seed)]["final"]["residual_lesion"]
+                   ["lesion_acc_drop"] for seed in MODEL_SEEDS]
+        work_ratios = [rows[(directions, 4, seed)]["cost"]
+                       ["training_forward_equivalent_examples"]
+                       / rows[(directions, 4, seed)]["cost"]
+                       ["ordinary_training_forward_examples"] for seed in MODEL_SEEDS]
+        summaries[directions] = {
+            "d4_mean": statistics.mean(accs[4]),
+            "finite": not nonfinite,
+        }
+        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"| {directions} | {depth} | {mean:.3f} +/- {sd:.3f} | "
+                  f"{gain_text} | {lesion_text} | {work_text} |")
+        print(f"K={directions} 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 = [directions for directions in DIRECTIONS if summaries[directions]["finite"]]
+    if not eligible:
+        raise RuntimeError("no finite calibration candidate")
+    best_mean = max(summaries[directions]["d4_mean"] for directions in eligible)
+    selected = min(directions for directions in eligible
+                   if summaries[directions]["d4_mean"] >= best_mean - 1.0)
+    print(f"C2 calibration-quality development selection: K={selected} layerwise/e4 "
+          f"(best d4={best_mean:.3f}%, selected d4={summaries[selected]['d4_mean']:.3f}%)")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/experiments/c2_calibration_quality_pilot.sh b/experiments/c2_calibration_quality_pilot.sh
new file mode 100755
index 0000000..d1c9291
--- /dev/null
+++ b/experiments/c2_calibration_quality_pilot.sh
@@ -0,0 +1,49 @@
+#!/usr/bin/env bash
+# Development-only C2 screen after the direct-NP diagnosis confirmed that
+# feedback amortization is the bottleneck.  Task seed 0 is reused strictly for
+# development; task seeds 6--8 are reserved for any independent confirmation.
+#
+# Frozen choices: context-gated vectorizer, eta=0.03, eta_A=0.01, layerwise
+# antithetic calibration every four steps.  Compare K4/K16 at d1/d4 over model
+# seeds 0--2.  The analyzer chooses the higher mean d4 endpoint, preferring K4
+# whenever it is within one point of the best finite candidate.
+# Usage: c2_calibration_quality_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 directions in 4 16; do
+  for model_seed in $MODEL_SEEDS; do
+    for depth in 1 4; do
+      tag="c2_calibdev_v1_tent_l2_w8_context_k${directions}_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 \
+        --epochs 80 --batch_size 256 --eta 0.03 --momentum 0.9 \
+        --eta_A 0.01 --eta_P 0.002 \
+        --pert_sigma 0.01 --pert_every 4 --pert_ndirs "$directions" \
+        --pert_mode layerwise \
+        --traffic_mode none --nuis_rho 0 \
+        --use_residual 1 --learn_A 1 --learn_P 1 --p_neutral 1 \
+        --task_train_examples 10000 --task_test_examples 5000 \
+        --task_levels 2 --task_n_in 1 --task_seed 0 \
+        --val_examples 2000 --split_seed 2027 --eval_split validation --eval_every 0 \
+        --diagnostics alignment --diagnostics_schedule final --probe_bs 512 \
+        --seed "$model_seed" --log_every 100000 --tag "$tag"
+    done
+  done
+done