feat: add superderive child and open phase-2 routing

Attach the stepwise derivation toolkit as a submodule and default SUPERPAPER_PHASE to 2 so rewrite-figure can use superderive. Notes still keep align for short expansions without cancel or subst.
2026-08-17 10:06:39 +08:00
parent 3a322aa8dc
commit e6918ffa35
24 changed files with 659 additions and 12 deletions
@@ -0,0 +1,46 @@
+schema: superpaper.ledger/v1
+retired_ids: []
+paper:
+  id: derive-delegate
+  title: Scaled attention
+  authors: ["Fixture"]
+  notes_language: zh
+  source: {kind: excerpt}
+coverage:
+  mode: excerpt
+questions:
+  - {id: Q1, text: "缩放从哪来？", source: "excerpt"}
+claims:
+  - id: C1
+    text: "1/√d_k 是改写，不是新算子"
+    kind: theoretical
+    status: core
+    supports: [Q1]
+definitions: []
+assumptions: []
+lemmas: []
+symbols:
+  - {name: Q, latex: "Q", meaning: "query", kind: value}
+  - {name: K, latex: "K", meaning: "key", kind: value}
+  - {name: dk, latex: "d_k", meaning: "key 维", kind: "shape parameter"}
+derivations:
+  - id: DER1
+    claim: C1
+    title: "缩放"
+    expand: true
+    figure: F3
+    steps:
+      - {id: S1, from: "QK^{\\top}", to: "QK^{\\top}/\\sqrt{d_k}", rule: scale}
+figures:
+  - id: F3
+    claim: C1
+    title: "缩放来自方差"
+    grammar: derivation
+    toolkit: superderive
+    signals: [rewrite-figure]
+    request: figures/F3/F3.request.md
+    include: figures/F3/build/F3.pdf
+    status: included
+evidence: []
+terms: []
+source_assets: []
@@ -0,0 +1 @@
+$QK^\top \to QK^\top/\sqrt{d_k}$。缩放压方差，避免 softmax 饱和。
@@ -0,0 +1,14 @@
+# Figure request F3
+toolkit: superderive
+language: cjk
+claim: 1/√d_k 是改写，不是新算子
+grammar: derivation
+work_rel_dir: figures/F3
+
+## Flow
+stage: {name: D1, text: "缩放来自方差，不是装饰"}
+row: {name: R1}
+in_row:
+  - {macro: sdstep, name: s1, role: rewrite,
+     lhs: "QK^{\\top}", rhs: "QK^{\\top}/\\sqrt{d_k}"}
+  - {macro: sdreason, name: s1, text: "§3.2.1，避免 softmax 饱和"}
@@ -0,0 +1,20 @@
+% Golden example -- a scale rewrite that must be seen, not just aligned.
+%   ../scripts/build.sh rewrite-cancel.tex
+\documentclass[border=10pt]{standalone}
+\usepackage[cjk]{superderive}
+
+\begin{document}
+\begin{tikzpicture}
+\sdstage{D1}{缩放来自方差，不是装饰}
+\sdrow{R1}
+  \sdstep[role=rewrite]{s1}{QK^{\top}}{QK^{\top}/\sqrt{d_k}}
+  \sdreason{s1}{§3.2.1，避免 softmax 饱和}
+\sdrowend
+\sdbbox{all}
+\sdtopformula{F}{$\mathrm{Attention}(Q,K,V)=\mathrm{softmax}(QK^{\top}/\sqrt{d_k})V$}
+\sdmeaningbox{mb}{120mm}{all}
+  {点积方差随 $d_k$ 线性涨}
+  {缩放是改写，不是新算子}
+  {没有缩放，softmax 进饱和区，梯度消失}
+\end{tikzpicture}
+\end{document}
@@ -0,0 +1,14 @@
+\documentclass[a4paper]{article}
+\input{notes-macros}
+\renewcommand{\notetitle}{缩放来自方差}
+\renewcommand{\notepaper}{Scaled attention}
+\begin{document}
+\tableofcontents
+\newpage
+\input{sections/sec-01.tex}
+\input{sections/sec-02.tex}
+\input{sections/sec-03.tex}
+\appendix
+\section{符号表}
+\input{sections/symbols.tex}
+\end{document}
@@ -0,0 +1,2 @@
+\section{这篇论文在问什么}
+注意力里的 $1/\sqrt{d_k}$ 是装饰，还是一次必须看见的改写？
@@ -0,0 +1,8 @@
+\section{主张与贡献}
+\splabel{C1}
+点积方差随 $d_k$ 涨。除掉 $\sqrt{d_k}$ 是改写，不是新算子。
+
+\spfig{F3}{缩放来自方差。}{fixture；toolkit: \texttt{superderive}；ledger id: F3。}
+
+\subsection{本章小结}
+这一步必须被看见，所以走 superderive，而不是笔记里的 \texttt{align}。
@@ -0,0 +1,2 @@
+\section{总结与延伸}
+四步以内、没有消项或代入要画出来时，仍然用 \texttt{align}。
@@ -0,0 +1,14 @@
+# Outline: Scaled attention
+
+## Lecture map
+
+| file | lecture_title | paper_sections | ledger_ids |
+|---|---|---|---|
+| sec-01.tex | 这篇论文在问什么 | 1 | Q1 |
+| sec-02.tex | 主张与贡献 | 1 | C1, F3 |
+| sec-03.tex | 总结与延伸 | 1 | C1 |
+| sec-app-a.tex | 符号表 | — | |
+
+## Locked
+- 首节标题必须是「这篇论文在问什么」
+- 末节（appendix 前）必须是「总结与延伸」
				`@@ -0,0 +1 @@`
				`$QK^\top \to QK^\top/\sqrt{d_k}$。缩放压方差，避免 softmax 饱和。`