From 96ff889ed214b68142daa0abf00456c733c879c6 Mon Sep 17 00:00:00 2001
From: Yuren Hao <yurenh2@illinois.edu>
Date: Thu, 30 Jul 2026 17:48:35 -0500
Subject: =?UTF-8?q?RESULT=2078:=20read=5Flin=E6=A2=AF=E5=BA=A6=E7=BA=A7?=
 =?UTF-8?q?=E9=AA=8C=E8=AF=81=3D=3Dfp64(fp32=E4=B8=8B=E6=B3=84=E6=BC=8F?=
 =?UTF-8?q?=E5=A1=8C=E9=9B=B6,=E8=AF=BB=E5=87=BA=E8=88=8D=E5=85=A5=3D?=
 =?UTF-8?q?=E5=85=A8=E9=83=A8=E7=97=85=E5=9B=A0);=20=E9=99=84=E5=B1=9E?=
 =?UTF-8?q?=E8=B0=B1=E4=BB=AA:=E5=89=8D=E6=B2=BF=E9=9A=8F=E8=AE=AD?=
 =?UTF-8?q?=E7=BB=83=E4=B8=8B=E5=8E=8B+seed=E8=BD=B4=E8=AF=81'=E6=AF=94?=
 =?UTF-8?q?=E5=80=BC=E5=86=B3=E5=AE=9A=E9=9D=9E=E5=B1=82=E5=8F=B7'+?=
 =?UTF-8?q?=E9=98=88=E5=80=BC=E5=B8=A62-5e-7;=20=E4=BF=AE=E5=A4=8D?=
 =?UTF-8?q?=E5=85=A5=E8=AE=AD=E7=BB=83=E8=B7=AF=E5=BE=84;=20CE=E7=BA=A7?=
 =?UTF-8?q?=E4=BB=AA=E5=99=A8=E7=BB=88=E5=AE=A1=E5=9C=A8=E9=A3=9E(?=
 =?UTF-8?q?=E9=A2=84=E6=B3=A8=E5=86=8C~+100%=E5=85=B3=E9=97=AD)?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
---
 docs/campaign/CASCADE_ABLATION_PLAN.md | 14 ++++++++++++++
 ep_run/casc_eq_train.py                |  9 ++++++++-
 2 files changed, 22 insertions(+), 1 deletion(-)

diff --git a/docs/campaign/CASCADE_ABLATION_PLAN.md b/docs/campaign/CASCADE_ABLATION_PLAN.md
index 503a321..e0e5cac 100644
--- a/docs/campaign/CASCADE_ABLATION_PLAN.md
+++ b/docs/campaign/CASCADE_ABLATION_PLAN.md
@@ -1932,3 +1932,17 @@ Matched 三口径（同窗 434-440k, n=61/臂）:
 - **硬件含义(为 Dillavou 通话)**: 这就是"读出必须差分"的数字版 — 对比读数要直接测差(d), 不能减两个
   大数((o+d)−o); 板上对应相关双采样/差分读出。基底加性地板设计律获得数字孪生实证。
 - 1B 迁移: 逐层 disp_ratio(一次自由相可测)+基底地板 → 逐层安全裕度表, 零训练。
+
+## RESULT 78 (2026-07-30): read_lin 梯度级验证成功 == fp64; 附属谱仪三发现; 修复入训练路径
+- **rlin 判决(fp32, C768 旗舰 ckpt, NB256)**: 上半 proj −0.006/+0.003/+0.003/+0.000/−0.000/+0.001,
+  与 fp64 臂逐块吻合(rel 同降至 0.013-0.020)。**读出侧 (z−o) 舍入 = 全部病因**; 位移的传输通道
+  (vjp 链的独立 d 张量)从来是满精度的 — 这也解释了为何 bf16-amp 与 fp32 泄漏相同(两者读出均 fp32)。
+- **附属谱仪(5 剖面)**: (1) 病灶前沿随训练下压 — C768 块6: 23%时净(ratio 5.2e-7)→57% −0.70(2.0e-7)
+  →91% −0.75(1.5e-7), 同块在自家比值穿线时转病; (2) seed 轴证律 — C512 s2 块7(3.8e-7)病 −0.43 而
+  s1 块7(6.9e-7)净: 决定者=测得比值, 非层号(条令级律形); (3) 阈值是带(RMS 2-5e-7)非刀刃 — 逐元素
+  |d_i|<eps 比例才是真判据, 1B 裕度表须用逐元素分位数。
+- 91% vs 57% 泄漏仅微深(−0.70→−0.75) → dg128 的 63%/97% 缺口主因=续跑欠账(候选a), 剂量漂移(b)次要。
+- **--read_lin 已入训练路径**(single 估计器; centered 留 TODO), 训练冒烟 gate cos 0.9999。
+- **CE 级终审在飞**: leak_rlin 仪器臂(battery-8 同源配置+read_lin, GPU3), 预注册预测: 尾窗≈allbp 杆
+  3.5206(物理泄漏 0.0462 全关)。确认后议程: 72M read_lin 从头重测真 gap(C512 块8-11 带 −0.40..−0.06
+  病训完了整个 crown, 封卷 +0.076 内含伪影贡献) > C768 从头 > ladder EP 臂重跑。
diff --git a/ep_run/casc_eq_train.py b/ep_run/casc_eq_train.py
index 3cb9494..ed164ce 100644
--- a/ep_run/casc_eq_train.py
+++ b/ep_run/casc_eq_train.py
@@ -699,7 +699,14 @@ def ep_step(x, y):
         gs = torch.autograd.grad(obj, all_params, allow_unused=True)
     elif EST == 'single':
         E = 0.0
-        for z, o in zip(zp, last_outs): E = E + 0.5 * ((z.detach().float() - o.float()) ** 2).sum()   # fp32 accumulation (bf16-safe; no-op in fp32)
+        if args.read_lin:
+            # FP-FLOOR FIX (RESULT 77/78): the read identity z - o = d holds algebraically but not in
+            # fp32 — (z - o) is a ROUNDED copy of d (components below eps·|o| are erased; the C768
+            # top-half leak). The d tensors themselves are full-precision; the linear form -<d,o> has
+            # the identical theta-derivative and bypasses the rounding. Probe-verified == fp64.
+            for dl, o in zip(GOV['_last_d'], last_outs): E = E - (dl.detach().float() * o.float()).sum()
+        else:
+            for z, o in zip(zp, last_outs): E = E + 0.5 * ((z.detach().float() - o.float()) ** 2).sum()   # fp32 accumulation (bf16-safe; no-op in fp32)
         obj = E / (NBT * beta_t) + obj_loss(readout(zp[-1].detach()).reshape(-1, vocab), y.reshape(-1))
         gs = torch.autograd.grad(obj, all_params, allow_unused=True)
     else:
-- 
cgit v1.2.3