From 98bfda74f7ee0457799b86465c867b6943c7be91 Mon Sep 17 00:00:00 2001
From: YurenHao0426 <Blackhao0426@gmail.com>
Date: Thu, 6 Aug 2026 14:39:14 -0500
Subject: experiment: freeze stagewise causal capture runner

---
 experiments/analyze_oral_a_v6_calibration.py | 103 +++++++++++++++++++++++++++
 1 file changed, 103 insertions(+)
 create mode 100644 experiments/analyze_oral_a_v6_calibration.py

(limited to 'experiments/analyze_oral_a_v6_calibration.py')

diff --git a/experiments/analyze_oral_a_v6_calibration.py b/experiments/analyze_oral_a_v6_calibration.py
new file mode 100644
index 0000000..8972138
--- /dev/null
+++ b/experiments/analyze_oral_a_v6_calibration.py
@@ -0,0 +1,103 @@
+#!/usr/bin/env python3
+"""Validate and gate stagewise causally whitened feedback capture."""
+import argparse
+import json
+import math
+import os
+
+
+SPLIT_HASH = "8328b206a97c420e49e54e3eca4abe3274c4756b084355784ea3fb8059e4515b"
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--input", default="results/oral_a_v6_calibration/result.json")
+    parser.add_argument(
+        "--out", default="results/oral_a_v6_calibration_gate.json")
+    args = parser.parse_args()
+    with open(args.input) as handle:
+        record = json.load(handle)
+    if record.get("protocol") != "oral_a_v6_stagewise_whitened_causal_capture_v1":
+        raise ValueError("unexpected V6 protocol")
+    expected = {
+        "depth": 20, "width": 16, "seed": 0, "loader_seed": 0,
+        "batch_size": 128, "train_limit": 10000,
+        "val_examples": 5000, "split_seed": 2027,
+        "normalization": "batchnorm", "residual_scale": 1.0,
+        "feedback_scale": 1.0, "sigma": 0.01,
+        "perturb_seed": 5000, "events_per_stage": 20,
+        "readout_relative_ridge": 1e-6,
+        "conv_diagonal_relative_ridge": 1e-3,
+        "alignment_probe": 64, "calibration_augmentation": False,
+    }
+    if record.get("settings") != expected:
+        raise ValueError("V6 settings drift")
+    if record["provenance"]["git_tracked_dirty"]:
+        raise ValueError("V6 result came from a tracked-dirty tree")
+    if record["split"]["validation_index_sha256"] != SPLIT_HASH:
+        raise ValueError("V6 split drift")
+    if record["test_examples_touched"] or record["validation_endpoints_observed"]:
+        raise ValueError("V6 touched a held-out endpoint")
+    work = record["work"]
+    audit = record["method_audit"]
+    fixed = record["fixed_hfa"]
+    learned = record["learned_scib"]
+    finite_metrics = [
+        fixed["early_third_alignment"], fixed["all_layer_alignment"],
+        learned["early_third_alignment"], learned["all_layer_alignment"],
+        learned["min_feedback_forward_norm_ratio"],
+        learned["max_feedback_forward_norm_ratio"],
+    ]
+    checks = {
+        "finite": bool(record["finite"])
+                  and all(math.isfinite(value) for value in finite_metrics),
+        "exactly_19_stages": work["stages"] == 19,
+        "exactly_380_edge_events": work["edge_events"] == 380,
+        "exactly_760_batch_loss_queries": (
+            work["logical_batch_loss_queries"] == 760),
+        "exactly_48640_per_example_observations": (
+            work["per_example_causal_observations"] == 48640),
+        "stage_order_readout_then_18_to_1": (
+            audit["stage_order"] == ["readout"] + list(range(18, 0, -1))),
+        "forward_state_bitwise_fixed": (
+            audit["forward_state_max_absolute_difference"] == 0.0),
+        "zero_forward_weight_reads_in_fit": (
+            audit["forward_weight_reads_in_feedback_fit"] == 0),
+        "zero_reverse_mode_learning_operations": (
+            audit["reverse_mode_learning_operations"] == 0),
+        "early_third_at_least_0.10": (
+            learned["early_third_alignment"] >= 0.10),
+        "all_layer_at_least_0.20": (
+            learned["all_layer_alignment"] >= 0.20),
+        "early_gain_over_fixed_hfa_at_least_0.08": (
+            learned["early_third_alignment"]
+            - fixed["early_third_alignment"] >= 0.08),
+        "feedback_norm_ratios_in_0.1_to_3": (
+            learned["min_feedback_forward_norm_ratio"] >= 0.1
+            and learned["max_feedback_forward_norm_ratio"] <= 3.0),
+    }
+    output = {
+        "protocol": "oral_a_v6_stagewise_whitened_causal_capture_gate_v1",
+        "status": "passed" if all(checks.values()) else "failed",
+        "checks": checks, "fixed_hfa": fixed, "learned_scib": learned,
+        "work": work, "source_commit": record["provenance"]["git_commit"],
+        "source_result": args.input,
+        "conditional_short_task_gate_open": all(checks.values()),
+        "confirmation_test_seeds_touched": False,
+        "review_score_before": 5, "review_score_after": 5,
+        "score_change_rule": "causal capture alone cannot raise score",
+    }
+    os.makedirs(os.path.dirname(os.path.abspath(args.out)), exist_ok=True)
+    with open(args.out, "w") as handle:
+        json.dump(output, handle, indent=2, sort_keys=True)
+        handle.write("\n")
+    print(json.dumps({
+        "status": output["status"], "checks": checks,
+        "fixed_hfa": fixed, "learned_scib": learned,
+    }, indent=2))
+
+
+if __name__ == "__main__":
+    main()
+
-- 
cgit v1.2.3