1 files changed, 38 insertions, 0 deletions
diff --git a/docs/audit/EP_READOUT_FLOOR_AUDIT.md b/docs/audit/EP_READOUT_FLOOR_AUDIT.md
index e925be3..7786b87 100644
--- a/docs/audit/EP_READOUT_FLOOR_AUDIT.md
+++ b/docs/audit/EP_READOUT_FLOOR_AUDIT.md
@@ -85,3 +85,41 @@ Hughes et al. Optica 2018、Kendall 2006.01981、2503.22810、2602.03670、Stoch
 **同一个算术，两种处方**。位移以独立张量存在时（dualprop 与我们），修复免费且彻底；两个状态来自
 独立弛豫时（其余所有人），任何重写都救不回来（已验证），只剩 fp64 累加对比量或给 β 设下限两条路。
 这一点必须写清楚，否则读者会望文生义地按"灾难性抵消"去找数值稳定的重写形式。
+
+---
+
+## 关键人际连接（审计本身未点出，本人核对）
+**`Rasmuskh` = Rasmus Kjær Høier**（仓库 commit 署名 `Rasmus Kjær Høier`，README 指向 ICML 2024
+"Two Tales of Single-Phase Contrastive Hebbian Learning", arXiv:2402.08573），**即 KHS ImageNet PCN
+论文 (arXiv:2606.03584, Kerjan/Høier/Scellier, 2026-06-02) 的第二作者**。
+论文已公开发表；**代码未发布**（arXiv 页与检索均无仓库链接，2026-08-02 复核）。
+
+⟹ **策略推论**：一封发给 Ben Scellier 的信即可同时覆盖三个目标，且走已开启的暖渠道，无需三封冷邮件：
+(1) 他自己 Rain 的 `energy-based-learning/training/sgd.py:357`；(2) 合著的 ImageNet PCN（β 标称 0.05,
+敏感性扫描下探 2e-4, 余量薄 250×）；(3) 合作者 Høier 的 dualprop（唯一出厂即危险者, 输出层可 1 行修复）。
+时机：等 135M 修复版完赛后一并发出（结果本身是这封信的由头）。
+
+## 给 Ben 的信中技术段（预写, 不依赖 135M 结果）
+> While tracing a width dependent quality loss in our own language model runs, we found the cause was
+> arithmetic rather than the learning rule. We form the nudged state as z = o + d and then read the
+> contrast as the squared norm of z minus o. Those agree algebraically and not in float32: forming z
+> rounds away every component of d below machine epsilon times the local activation, and the later
+> subtraction returns the damaged copy with no error signal. Wider models push more layers below that
+> threshold, so the loss grew with width and imitated a scaling wall. Reading the contrast from the
+> stored d instead, which has the same derivative, costs nothing and removed the effect completely.
+>
+> We then checked whether the pattern appears elsewhere, and it is common. The relative error of a
+> float32 contrast estimator follows roughly epsilon over r, where r is the displacement to activation
+> ratio, so r near 1e-5 costs a few percent and r near 1e-7 leaves noise. Most published work is
+> protected by a large nudge or by float64 in numpy, and the margin is real but undocumented: none of
+> the implementations we read assert on r or accumulate the contrast in double precision. Two specific
+> notes, offered privately and only because they are close to you. The estimator in
+> energy-based-learning at training/sgd.py:357 is unguarded at small nudging, and the beta sweep in the
+> ImageNet PCN paper reaches 2e-4, where the margin is a couple of hundred times thinner than at the
+> headline setting. Separately, Rasmus's dualprop code forms the states as act(ff plus or minus fb) and
+> then recovers the difference by subtraction, in float32 at beta 0.01 on VGG16; at the output layer
+> there is no activation in between, so the difference is exactly fb and the retained tensor can be
+> substituted directly. In the hidden layers the difference passes through the activation, so the
+> remedy there is to form it elementwise in higher precision rather than to substitute.
+>
+> A standalone checker is at [tools/contrast_floor_check.py], and it runs against any implementation.