<feed xmlns='http://www.w3.org/2005/Atom'>
<title>ept.git/docs/campaign/CASCADE_ABLATION_PLAN.md, branch master</title>
<subtitle>Unnamed repository; edit this file 'description' to name the repository.
</subtitle>
<link rel='alternate' type='text/html' href='https://git.blackhao.com/ept.git/'/>
<entry>
<title>种子恢复: rlin_s2精确续@80k(~48h), bp_s3续@260k(~14h, weights-only冷动量星号记档)</title>
<updated>2026-08-05T18:06:55+00:00</updated>
<author>
<name>Yuren Hao</name>
<email>yurenh2@illinois.edu</email>
</author>
<published>2026-08-05T18:06:55+00:00</published>
<link rel='alternate' type='text/html' href='https://git.blackhao.com/ept.git/commit/?id=0df61c1d36faf3e985394c943853b5bab5adaa8c'/>
<id>0df61c1d36faf3e985394c943853b5bab5adaa8c</id>
<content type='text'>
Co-Authored-By: Claude Fable 5 &lt;noreply@anthropic.com&gt;
Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
</content>
<content type='xhtml'>
<div xmlns='http://www.w3.org/1999/xhtml'>
<pre>
Co-Authored-By: Claude Fable 5 &lt;noreply@anthropic.com&gt;
Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
</pre>
</div>
</content>
</entry>
<entry>
<title>筛选正式闭卷(11优化器33臂): AdamW收口≈5.01; 增益钳位测试设计入档(逐元件vs逐模式饱和=唯一的坎, 预注册两结局)</title>
<updated>2026-08-05T15:35:27+00:00</updated>
<author>
<name>Yuren Hao</name>
<email>yurenh2@illinois.edu</email>
</author>
<published>2026-08-05T15:35:27+00:00</published>
<link rel='alternate' type='text/html' href='https://git.blackhao.com/ept.git/commit/?id=a7c6403027cc9297adc607e702f1c8290e50d146'/>
<id>a7c6403027cc9297adc607e702f1c8290e50d146</id>
<content type='text'>
Co-Authored-By: Claude Fable 5 &lt;noreply@anthropic.com&gt;
Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
</content>
<content type='xhtml'>
<div xmlns='http://www.w3.org/1999/xhtml'>
<pre>
Co-Authored-By: Claude Fable 5 &lt;noreply@anthropic.com&gt;
Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
</pre>
</div>
</content>
</entry>
<entry>
<title>RESULT 90终局: NS价值曲线k=3以60%成本买回99%(质量门槛远低于符号复刻,廉价正交化器规格=k3等效); muon 4.71卫冕, olion 4.75; 全程赛推荐muon/olion-k3/sgdm</title>
<updated>2026-08-05T14:29:42+00:00</updated>
<author>
<name>Yuren Hao</name>
<email>yurenh2@illinois.edu</email>
</author>
<published>2026-08-05T14:29:42+00:00</published>
<link rel='alternate' type='text/html' href='https://git.blackhao.com/ept.git/commit/?id=d462ce44240e0016a1e90098c759eba1ac2d06da'/>
<id>d462ce44240e0016a1e90098c759eba1ac2d06da</id>
<content type='text'>
Co-Authored-By: Claude Fable 5 &lt;noreply@anthropic.com&gt;
Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
</content>
<content type='xhtml'>
<div xmlns='http://www.w3.org/1999/xhtml'>
<pre>
Co-Authored-By: Claude Fable 5 &lt;noreply@anthropic.com&gt;
Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
</pre>
</div>
</content>
</entry>
<entry>
<title>RESULT 90: 优化器筛选收敛 — Muon 4.7227 领跑, NS消融判决(摘NS代价0.45, nanoGPT版未调用NS之谜归他们), sign家族聚在5.13-5.25; k价值曲线在跑; 曲率通道死刑+偏置公理入档</title>
<updated>2026-08-05T14:09:14+00:00</updated>
<author>
<name>Yuren Hao</name>
<email>yurenh2@illinois.edu</email>
</author>
<published>2026-08-05T14:09:14+00:00</published>
<link rel='alternate' type='text/html' href='https://git.blackhao.com/ept.git/commit/?id=807c6ba3bab8d9afe18ca7d8fb67038bc2cfaf73'/>
<id>807c6ba3bab8d9afe18ca7d8fb67038bc2cfaf73</id>
<content type='text'>
Co-Authored-By: Claude Fable 5 &lt;noreply@anthropic.com&gt;
Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
</content>
<content type='xhtml'>
<div xmlns='http://www.w3.org/1999/xhtml'>
<pre>
Co-Authored-By: Claude Fable 5 &lt;noreply@anthropic.com&gt;
Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
</pre>
</div>
</content>
</entry>
<entry>
<title>RESULT 89终局 + 交付物: 135M完赛 EP 3.2087 vs BP 3.2074(+0.1% ppl), 旧读出+45.2%</title>
<updated>2026-08-04T10:46:57+00:00</updated>
<author>
<name>Yuren Hao</name>
<email>yurenh2@illinois.edu</email>
</author>
<published>2026-08-04T10:46:57+00:00</published>
<link rel='alternate' type='text/html' href='https://git.blackhao.com/ept.git/commit/?id=f3ade7674328c55a98d32fd332a3a1cf79fd40a6'/>
<id>f3ade7674328c55a98d32fd332a3a1cf79fd40a6</id>
<content type='text'>
窗口敏感性诚实记录: 97%窗给-0.0030(seed间距0.0059), 完赛大窗给+0.0013(间距0.0005, cosine
衰减后BP收拢); 以大窗为准且不得称"统计不可分"(2.6倍间距), 正确表述=差0.1%且处分辨极限,
seed数(BP n=2/EP n=1)不足以给区间。
交付: --gen模式(复用模型定义, EP/BP同代码路径)、fig_135m_curves(四曲线/横轴token/BP细线在上)、
EPT_135M_samples.ipynb(同prompt同种子并排, 输出已烤入, 代码真可跑)。

Co-Authored-By: Claude Fable 5 &lt;noreply@anthropic.com&gt;
Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
</content>
<content type='xhtml'>
<div xmlns='http://www.w3.org/1999/xhtml'>
<pre>
窗口敏感性诚实记录: 97%窗给-0.0030(seed间距0.0059), 完赛大窗给+0.0013(间距0.0005, cosine
衰减后BP收拢); 以大窗为准且不得称"统计不可分"(2.6倍间距), 正确表述=差0.1%且处分辨极限,
seed数(BP n=2/EP n=1)不足以给区间。
交付: --gen模式(复用模型定义, EP/BP同代码路径)、fig_135m_curves(四曲线/横轴token/BP细线在上)、
EPT_135M_samples.ipynb(同prompt同种子并排, 输出已烤入, 代码真可跑)。

Co-Authored-By: Claude Fable 5 &lt;noreply@anthropic.com&gt;
Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
</pre>
</div>
</content>
</entry>
<entry>
<title>RESULT 89: 135M判决 — 修复版3.2024 vs BP双seed均值3.2054(gap -0.0030=seed间距一半), 统计不可分</title>
<updated>2026-08-04T09:13:43+00:00</updated>
<author>
<name>Yuren Hao</name>
<email>yurenh2@illinois.edu</email>
</author>
<published>2026-08-04T09:13:43+00:00</published>
<link rel='alternate' type='text/html' href='https://git.blackhao.com/ept.git/commit/?id=3962588a667d0d8945044b105dedda75e3fa3a41'/>
<id>3962588a667d0d8945044b105dedda75e3fa3a41</id>
<content type='text'>
旧读出同点+0.3703=+44.8%ppl, 即"宽度墙"真实体量, 全部由一行读出改写消除。五尺寸(11M-135M)
修复后全部落入种子噪声。BP s2完赛best 3.0968。EP剩11.4k步出终局。

Co-Authored-By: Claude Fable 5 &lt;noreply@anthropic.com&gt;
Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
</content>
<content type='xhtml'>
<div xmlns='http://www.w3.org/1999/xhtml'>
<pre>
旧读出同点+0.3703=+44.8%ppl, 即"宽度墙"真实体量, 全部由一行读出改写消除。五尺寸(11M-135M)
修复后全部落入种子噪声。BP s2完赛best 3.0968。EP剩11.4k步出终局。

Co-Authored-By: Claude Fable 5 &lt;noreply@anthropic.com&gt;
Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
</pre>
</div>
</content>
</entry>
<entry>
<title>RESULT 88: K电池判决 — K=1/2/3在CE上不可分(配对差3e-5, 比seed散布小两个数量级), 成本可降至1.67×BP</title>
<updated>2026-08-03T22:47:13+00:00</updated>
<author>
<name>Yuren Hao</name>
<email>yurenh2@illinois.edu</email>
</author>
<published>2026-08-03T22:47:13+00:00</published>
<link rel='alternate' type='text/html' href='https://git.blackhao.com/ept.git/commit/?id=1e1d266db41f3acdcd647df71070982b2d1b057b'/>
<id>1e1d266db41f3acdcd647df71070982b2d1b057b</id>
<content type='text'>
三臂共种子=val批次相同=噪声完全相关, 配对差分为正统计量。先疑bit-identical(bf16空转)经逐行
比对否决: K有真实但可忽略的数值效应。诚实定位: K=1即BP+仅存于θ-雅可比的O(β)修正, 迭代在本
架构不做实质工作(级联自由相按构造为不动点, 与KHS同); 贡献在局部性/有限β偏差长程可控/硬件约束,
不在"梯度不同"。待决: 主线是否改K=1(倾向旗舰保K=3可比, 另开K=1成本臂)。

Co-Authored-By: Claude Fable 5 &lt;noreply@anthropic.com&gt;
Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
</content>
<content type='xhtml'>
<div xmlns='http://www.w3.org/1999/xhtml'>
<pre>
三臂共种子=val批次相同=噪声完全相关, 配对差分为正统计量。先疑bit-identical(bf16空转)经逐行
比对否决: K有真实但可忽略的数值效应。诚实定位: K=1即BP+仅存于θ-雅可比的O(β)修正, 迭代在本
架构不做实质工作(级联自由相按构造为不动点, 与KHS同); 贡献在局部性/有限β偏差长程可控/硬件约束,
不在"梯度不同"。待决: 主线是否改K=1(倾向旗舰保K=3可比, 另开K=1成本臂)。

Co-Authored-By: Claude Fable 5 &lt;noreply@anthropic.com&gt;
Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
</pre>
</div>
</content>
</entry>
<entry>
<title>RESULT 87: KHS论文精读 — 同性质获发明人背书, 但迭代计数对比口径不当已停用</title>
<updated>2026-08-03T14:20:07+00:00</updated>
<author>
<name>Yuren Hao</name>
<email>yurenh2@illinois.edu</email>
</author>
<published>2026-08-03T14:20:07+00:00</published>
<link rel='alternate' type='text/html' href='https://git.blackhao.com/ept.git/commit/?id=0918c65cc07b3d4a06be56a46f80a7c5f143471a'/>
<id>0918c65cc07b3d4a06be56a46f80a7c5f143471a</id>
<content type='text'>
关键: 他们自由相亦为一次前向; BP等价性由他们主动写出(cos≈1是正确性依据非弱点); 他们未测余弦
(此轴我方更严格)。不利: 其扫描结论K≥4且"最优K≈层数", 我方K=3在线下一格(K电池落点极佳)。
口径修正: 其迭代为异步局部更新(一次约走一层), 我方一次走完整链 ⟹ "3次vs20次"是不同物相比且
偏向我方, 停用; 改用实测墙钟2.7× vs其6.7-12×。项目页proof卡与技术摘要已同步。

Co-Authored-By: Claude Fable 5 &lt;noreply@anthropic.com&gt;
Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
</content>
<content type='xhtml'>
<div xmlns='http://www.w3.org/1999/xhtml'>
<pre>
关键: 他们自由相亦为一次前向; BP等价性由他们主动写出(cos≈1是正确性依据非弱点); 他们未测余弦
(此轴我方更严格)。不利: 其扫描结论K≥4且"最优K≈层数", 我方K=3在线下一格(K电池落点极佳)。
口径修正: 其迭代为异步局部更新(一次约走一层), 我方一次走完整链 ⟹ "3次vs20次"是不同物相比且
偏向我方, 停用; 改用实测墙钟2.7× vs其6.7-12×。项目页proof卡与技术摘要已同步。

Co-Authored-By: Claude Fable 5 &lt;noreply@anthropic.com&gt;
Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
</pre>
</div>
</content>
</entry>
<entry>
<title>RESULT 86: K=1梯度余弦亦0.9999 — 记录对我方不利的证据</title>
<updated>2026-08-03T14:16:45+00:00</updated>
<author>
<name>Yuren Hao</name>
<email>yurenh2@illinois.edu</email>
</author>
<published>2026-08-03T14:16:45+00:00</published>
<link rel='alternate' type='text/html' href='https://git.blackhao.com/ept.git/commit/?id=f0552201cfea2521052bdd7faa6227e73bd1118b'/>
<id>f0552201cfea2521052bdd7faa6227e73bd1118b</id>
<content type='text'>
三点连平线(K=1/3/8)=梯度保真对K不敏感; 代码层归因(forces的vjp链即逐层反传, 级联自由相
按构造为精确不动点); 诚实定位(cos≈1是EP定理不是丑闻, 区分点在更新局部性与基底操作);
真正脆弱点=J^T非K(Stern/Liu不可能性, 答案在硬件线); 成本含义K=1→1.67×BP待CE裁决;
对外口径即刻调整(不得再拿"3次弛豫"当卖点, 主动公布K=1余弦)。

Co-Authored-By: Claude Fable 5 &lt;noreply@anthropic.com&gt;
Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
</content>
<content type='xhtml'>
<div xmlns='http://www.w3.org/1999/xhtml'>
<pre>
三点连平线(K=1/3/8)=梯度保真对K不敏感; 代码层归因(forces的vjp链即逐层反传, 级联自由相
按构造为精确不动点); 诚实定位(cos≈1是EP定理不是丑闻, 区分点在更新局部性与基底操作);
真正脆弱点=J^T非K(Stern/Liu不可能性, 答案在硬件线); 成本含义K=1→1.67×BP待CE裁决;
对外口径即刻调整(不得再拿"3次弛豫"当卖点, 主动公布K=1余弦)。

Co-Authored-By: Claude Fable 5 &lt;noreply@anthropic.com&gt;
Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
</pre>
</div>
</content>
</entry>
<entry>
<title>RESULT 85: EP成本定案 — 受控实测wall-clock 2.70×(纯训练步2.77×), 解析FLOPs 9F/3F=3.0×互证</title>
<updated>2026-08-03T14:04:08+00:00</updated>
<author>
<name>Yuren Hao</name>
<email>yurenh2@illinois.edu</email>
</author>
<published>2026-08-03T14:04:08+00:00</published>
<link rel='alternate' type='text/html' href='https://git.blackhao.com/ept.git/commit/?id=2ec65555120378de3ff8298bed97852b143bd802'/>
<id>2ec65555120378de3ff8298bed97852b143bd802</id>
<content type='text'>
口径清理: 项目页/call sheet的"2.0×"与给Suhas的"1.0-1.4×"均作废(后者疑将holofast+sdpa对EP
自身的~1.5×加速误作对BP比值); 七月brief的3.2× FLOPs一直正确。KHS对比(6.7-12×)不受影响,
EP家族内仍有2.5-4.4×成本优势。方法: SIGSTOP静音同卡, 同模型同batch各400步, 两窗口复核。

Co-Authored-By: Claude Fable 5 &lt;noreply@anthropic.com&gt;
Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
</content>
<content type='xhtml'>
<div xmlns='http://www.w3.org/1999/xhtml'>
<pre>
口径清理: 项目页/call sheet的"2.0×"与给Suhas的"1.0-1.4×"均作废(后者疑将holofast+sdpa对EP
自身的~1.5×加速误作对BP比值); 七月brief的3.2× FLOPs一直正确。KHS对比(6.7-12×)不受影响,
EP家族内仍有2.5-4.4×成本优势。方法: SIGSTOP静音同卡, 同模型同batch各400步, 两窗口复核。

Co-Authored-By: Claude Fable 5 &lt;noreply@anthropic.com&gt;
Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
</pre>
</div>
</content>
</entry>
</feed>
