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.
---
 protocol/report.py | 38 ++++++++++++++++++++++++++++++--------
 1 file changed, 30 insertions(+), 8 deletions(-)

(limited to 'protocol/report.py')

diff --git a/protocol/report.py b/protocol/report.py
index 15d6c34..00640eb 100644
--- a/protocol/report.py
+++ b/protocol/report.py
@@ -35,6 +35,19 @@ class DiagnosticThresholds:
     """
 
     g_norm_floor: float = 1e-7
+    # Per-block residual growth ratio threshold. The diagnostic is
+    # `max_l(||h_{l+1}|| / ||h_l||)` — the largest single-block residual
+    # amplification. We avoided `max(||h||) / ||h_0||` because it false-
+    # positives on ViT-style architectures where the cls token at layer 0
+    # is anomalously small after patch_embed.
+    #
+    # Calibration on observed data:
+    #   - BP-trained, late training: <5× per block (steady state)
+    #   - BP ViT, early training (epoch 1-5): 13-25× max (cls token still
+    #     resolving from its small init magnitude)
+    #   - DFA-trained ResMLP / ViT: 100-4000× max per block
+    # Threshold 50 sits cleanly between healthy-early-training (max 25) and
+    # failure-regime (min 100), with margin on both sides.
     h_norm_explosion_ratio: float = 50.0
     stability_drift_ceiling: float = 0.30
     frozen_acc_margin_pp: float = 2.0
@@ -58,14 +71,23 @@ class DiagnosticReport:
     # Per-diagnostic verdicts
     # ------------------------------------------------------------------ #
 
+    @property
+    def max_per_block_growth(self) -> float:
+        """max_l (||h_{l+1}|| / ||h_l||) — the largest residual-stream
+        amplification by any single block. Healthy BP/EP networks have all
+        per-block growth < 5×; pathological networks (DFA/SB/CB on pre-LN
+        residuals) have at least one block with growth > 100×."""
+        if len(self.residual_norms) < 2:
+            return 1.0
+        ratios = []
+        for i in range(len(self.residual_norms) - 1):
+            denom = max(self.residual_norms[i], 1e-30)
+            ratios.append(self.residual_norms[i + 1] / denom)
+        return max(ratios)
+
     @property
     def residual_stream_exploded(self) -> bool:
-        if not self.residual_norms:
-            return False
-        h0 = self.residual_norms[0]
-        if h0 <= 0:
-            return False
-        return (max(self.residual_norms) / h0) > self.thresholds.h_norm_explosion_ratio
+        return self.max_per_block_growth > self.thresholds.h_norm_explosion_ratio
 
     @property
     def bp_grad_at_floor(self) -> bool:
@@ -126,8 +148,8 @@ class DiagnosticReport:
             lines.append(f"      h_{l}:  {self.residual_norms[l]:.3e}")
         if self.residual_stream_exploded:
             lines.append(
-                f"    FLAG: max/min ratio "
-                f"{max(self.residual_norms)/max(self.residual_norms[0],1e-30):.2e} "
+                f"    FLAG: max per-block growth ‖h_{{l+1}}‖/‖h_l‖ = "
+                f"{self.max_per_block_growth:.2e} "
                 f"> threshold {self.thresholds.h_norm_explosion_ratio}× — "
                 "residual stream exploded."
             )
-- 
cgit v1.2.3