experiments: freeze FashionMNIST C1 recovery

1 files changed, 155 insertions, 0 deletions

diff --git a/experiments/analyze_c1_fmnist_validation.py b/experiments/analyze_c1_fmnist_validation.py
new file mode 100644
index 0000000..af90543
--- /dev/null
+++ b/experiments/analyze_c1_fmnist_validation.py
@@ -0,0 +1,155 @@
+"""Audit and select the frozen validation-only FashionMNIST C1 recovery."""
+import glob
+import json
+import math
+import os
+
+
+ROOT = os.path.join(os.path.dirname(os.path.dirname(os.path.abspath(__file__))), "results")
+PREFIX = "c1_fmnist_val_v1_"
+TRAFFIC_SEEDS = (1234, 5678)
+SIGNALS = ("raw", "matched", "residual", "taskfit")
+
+
+def finite_mean(values):
+    finite = [value for value in values if value is not None and math.isfinite(value)]
+    if not finite:
+        raise RuntimeError("expected at least one finite diagnostic")
+    return sum(finite) / len(finite)
+
+
+def signal_name(args):
+    if not args["use_residual"]:
+        return ("matched" if args["raw_scale_control"] == "match_innovation_norm"
+                else "raw")
+    return "residual" if args["p_neutral"] else "taskfit"
+
+
+def audit_row(path, row):
+    args = row["args"]
+    required = {
+        "mode": "sdil", "dataset": "fmnist", "depth": 3, "width": 256,
+        "residual": 1, "epochs": 15, "batch_size": 128, "eta": 0.05,
+        "momentum": 0.9, "eta_A": 0.02, "eta_P": 0.05,
+        "pert_sigma": 0.01, "pert_every": 4, "pert_ndirs": 1,
+        "pert_mode": "simultaneous", "learn_A": 1, "learn_P": 1,
+        "p_warmup_steps": 200, "p_warmup_eta": 0.05,
+        "val_examples": 5000, "split_seed": 2027,
+        "eval_split": "validation", "eval_every": 0,
+        "diagnostics": "alignment", "diagnostics_schedule": "final",
+        "probe_bs": 512, "seed": 0,
+    }
+    mismatches = {key: (args.get(key), value) for key, value in required.items()
+                  if args.get(key) != value}
+    if mismatches:
+        raise RuntimeError(f"protocol mismatch {path}: {mismatches}")
+    if row.get("final", {}).get("eval_split") != "validation":
+        raise RuntimeError(f"non-validation 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") != 5000
+            or split.get("evaluation_split") != "validation"):
+        raise RuntimeError(f"invalid validation split {path}: {split}")
+    protocol = row.get("diagnostic_protocol", {})
+    if protocol != {"probe_source": "training_prefix", "probe_examples": 512,
+                    "schedule": "final"}:
+        raise RuntimeError(f"diagnostic protocol mismatch {path}: {protocol}")
+    if row.get("provenance", {}).get("git_dirty") is not False:
+        raise RuntimeError(f"dirty or unknown source provenance: {path}")
+
+
+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)
+        rows.append(row)
+        commits.add(row["provenance"]["git_commit"])
+        split_hashes.add(row["split"]["validation_index_sha256"])
+    if len(commits) != 1 or len(split_hashes) != 1:
+        raise RuntimeError(f"expected one source and split, got {commits}, {split_hashes}")
+
+    groups = {}
+    for row in rows:
+        args = row["args"]
+        key = (args["traffic_mode"], args["nuis_rho"], args["traffic_seed"],
+               signal_name(args))
+        if key in groups:
+            raise RuntimeError(f"duplicate validation row: {key}")
+        groups[key] = row
+
+    expected = {("none", 0.0, 1234, signal) for signal in SIGNALS[:3]}
+    expected |= {("soma", 0.5, traffic_seed, signal)
+                 for traffic_seed in TRAFFIC_SEEDS for signal in SIGNALS}
+    expected |= {("topdown", rho, traffic_seed, signal)
+                 for rho in (0.2, 0.5) for traffic_seed in TRAFFIC_SEEDS
+                 for signal in SIGNALS}
+    if set(groups) != expected or len(rows) != 27:
+        raise RuntimeError(f"incomplete/unexpected panel: rows={len(rows)}, "
+                           f"missing={expected - set(groups)}, extra={set(groups) - expected}")
+
+    print(f"commit={next(iter(commits))} split={next(iter(split_hashes))} rows={len(rows)}")
+    print("| traffic | rho | projection seed | signal | validation (%) | cos(r,-g) | R2 |")
+    print("|:---|---:|---:|:---|---:|---:|---:|")
+    order = [("none", 0.0, 1234)]
+    order += [("soma", 0.5, seed) for seed in TRAFFIC_SEEDS]
+    order += [("topdown", rho, seed) for rho in (0.2, 0.5) for seed in TRAFFIC_SEEDS]
+    for family, rho, traffic_seed in order:
+        signals = SIGNALS[:3] if family == "none" else SIGNALS
+        for signal in signals:
+            row = groups[(family, rho, traffic_seed, signal)]
+            final = row["final"]
+            r2 = "--" if family == "none" else f"{finite_mean(final['traffic_r2']):.3f}"
+            print(f"| {family} | {rho:g} | {traffic_seed} | {signal} | "
+                  f"{100 * final['val_acc']:.3f} | {finite_mean(final['cos_r_negg']):+.3f} | {r2} |")
+
+    no_key = ("none", 0.0, 1234)
+    no_acc = 100 * groups[no_key + ("residual",)]["final"]["val_acc"]
+    no_changes = {
+        signal: no_acc - 100 * groups[no_key + (signal,)]["final"]["val_acc"]
+        for signal in ("raw", "matched")
+    }
+    no_ok = min(no_changes.values()) >= -0.5
+    print(f"no-traffic residual-minus controls={no_changes}: {'PASS' if no_ok else 'FAIL'}")
+
+    checks = {}
+    for family, rho in (("soma", 0.5), ("topdown", 0.2), ("topdown", 0.5)):
+        realization_checks = []
+        gains = []
+        for traffic_seed in TRAFFIC_SEEDS:
+            prefix = (family, rho, traffic_seed)
+            residual = groups[prefix + ("residual",)]
+            matched = groups[prefix + ("matched",)]
+            gain = 100 * (residual["final"]["val_acc"] - matched["final"]["val_acc"])
+            residual_acc = 100 * residual["final"]["val_acc"]
+            r2 = finite_mean(residual["final"]["traffic_r2"])
+            alignment = finite_mean(residual["final"]["cos_r_negg"])
+            passed = (gain >= 2.0 and residual_acc >= no_acc - 3.0
+                      and r2 >= 0.5 and alignment >= 0.1)
+            gains.append(gain)
+            realization_checks.append(passed)
+            print(f"{prefix}: gain={gain:+.3f}, residual={residual_acc:.3f}, "
+                  f"no-traffic gap={residual_acc - no_acc:+.3f}, R2={r2:.3f}, "
+                  f"alignment={alignment:+.3f}: {'PASS' if passed else 'FAIL'}")
+        checks[(family, rho)] = (all(realization_checks), min(gains))
+
+    soma_ok = checks[("soma", 0.5)][0]
+    eligible = [(rho, checks[("topdown", rho)][1]) for rho in (0.2, 0.5)
+                if checks[("topdown", rho)][0]]
+    if soma_ok and no_ok and eligible:
+        selected_rho, worst_gain = max(eligible, key=lambda item: (item[1], -item[0]))
+        print(f"C1 FashionMNIST validation gate: PASS; selected topdown rho={selected_rho:g} "
+              f"by highest worst-projection gain ({worst_gain:+.3f} points)")
+        return
+    print("C1 FashionMNIST validation gate: FAIL; do not evaluate the FashionMNIST test set")
+    raise SystemExit(1)
+
+
+if __name__ == "__main__":
+    main()