From 31ddecc9eb646b15c4ac5960c7de9346c8f7be68 Mon Sep 17 00:00:00 2001
From: YurenHao0426 <Blackhao0426@gmail.com>
Date: Tue, 7 Apr 2026 23:00:54 -0500
Subject: Protocol diagnostic (a): use max per-block growth, not max/min ratio
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Old metric: max(||h||) / max(||h_0||, eps). False-positives on ViT-style
architectures because the cls token at layer 0 (right after patch_embed)
has anomalously small magnitude (~0.3-1.5), inflating the ratio even on
healthy BP-trained ViTs.

New metric: max_l(||h_{l+1}|| / ||h_l||) — the largest single-block
residual amplification. Architecture-invariant.

Calibration:
  - BP-trained, late training: <5x per block
  - BP ViT, early epochs (cls token resolving): 13-25x max
  - DFA-trained ResMLP/ViT: 100-4000x per block
Threshold raised from 10 to 50 to sit cleanly between healthy-early-
training (max 25) and failure-regime (min 100).

Re-verifications:
  - smoke test (BP/DFA/EP): all 3 verdicts unchanged
  - random init (3 seeds): trustworthy on all 3
  - 5-method audit table single-seed: identical verdicts
  - decision-utility ablation: identical (still 0/5 by S1, 3/5 by S_full)
  - temporal evolution 3-seed: (b) now fires first at ep 3-4, (a) at ep
    8-11. Both well before training ends. The 'protocol fires ~92 epochs
    early' story still holds.
  - ViT temporal evolution: BP no longer false-fires; DFA fires (a) ep 1,
    (b) ep 3 — protocol works on the second architecture.
---
 results/protocol_audit/audit_table_s42.json | 15 ++++++++++-----
 1 file changed, 10 insertions(+), 5 deletions(-)

(limited to 'results/protocol_audit/audit_table_s42.json')

diff --git a/results/protocol_audit/audit_table_s42.json b/results/protocol_audit/audit_table_s42.json
index d1c1f84..901ac85 100644
--- a/results/protocol_audit/audit_table_s42.json
+++ b/results/protocol_audit/audit_table_s42.json
@@ -1,6 +1,6 @@
 {
   "reports": {
-    "bp": {
+    "bp_s42": {
       "method_name": "BP",
       "notes": "4-block d=256 ResMLP, CIFAR-10, seed 42",
       "residual_norms": [
@@ -29,7 +29,7 @@
         "frozen_acc_margin_pp": 2.0
       }
     },
-    "dfa": {
+    "dfa_s42": {
       "method_name": "DFA",
       "notes": "4-block d=256 ResMLP, CIFAR-10, seed 42",
       "residual_norms": [
@@ -58,7 +58,7 @@
         "frozen_acc_margin_pp": 2.0
       }
     },
-    "state_bridge": {
+    "state_bridge_s42": {
       "method_name": "STATE_BRIDGE",
       "notes": "4-block d=256 ResMLP, CIFAR-10, seed 42",
       "residual_norms": [
@@ -87,7 +87,7 @@
         "frozen_acc_margin_pp": 2.0
       }
     },
-    "credit_bridge": {
+    "credit_bridge_s42": {
       "method_name": "CREDIT_BRIDGE",
       "notes": "4-block d=256 ResMLP, CIFAR-10, seed 42",
       "residual_norms": [
@@ -116,7 +116,7 @@
         "frozen_acc_margin_pp": 2.0
       }
     },
-    "ep": {
+    "ep_s42": {
       "method_name": "EP",
       "notes": "4-block d=256 ResMLP, CIFAR-10, seed 42",
       "residual_norms": [
@@ -149,6 +149,7 @@
   "summary": [
     {
       "method": "bp",
+      "seed": 42,
       "acc": 0.6149,
       "h_L": 205.75946044921875,
       "g_L": 0.0003701267414726317,
@@ -158,6 +159,7 @@
     },
     {
       "method": "dfa",
+      "seed": 42,
       "acc": 0.3107,
       "h_L": 435299520.0,
       "g_L": 4.174704582027289e-09,
@@ -167,6 +169,7 @@
     },
     {
       "method": "state_bridge",
+      "seed": 42,
       "acc": 0.1695,
       "h_L": 228665568.0,
       "g_L": 1.8411722146893794e-09,
@@ -176,6 +179,7 @@
     },
     {
       "method": "credit_bridge",
+      "seed": 42,
       "acc": 0.2562,
       "h_L": 606231552.0,
       "g_L": 9.011226209665324e-10,
@@ -185,6 +189,7 @@
     },
     {
       "method": "ep",
+      "seed": 42,
       "acc": 0.359,
       "h_L": 3286.841064453125,
       "g_L": 0.00016422539192717522,
-- 
cgit v1.2.3