LaTeX tab: add TikZ carry-cascade figure as second copyable block

app.py

@@ -102,6 +102,153 @@ Across three undersized architectures and data sizes from 10K to 100K,
 as $+50$\,pp (44\% $\to$ 94\% on the smallest model at 50K examples;
 Table~\ref{tab:undersized-wins}).
 """
+
+LATEX_FIGURE_EXAMPLE = r"""% fig_arithmetic_example.tex
+% Usage in paper: \input{figures/fig_arithmetic_example/fig_arithmetic_example.tex}
+% Required packages: tikz, xcolor
+%
+% Shows 959271 + 040756 = 1000027 (4-deep carry cascade).
+% Token assignments from model add_sub_sorl_v1_abs30_K1_100K (K=1, abs30).
+
+\begin{figure}[t]
+\centering
+\begin{tikzpicture}[
+  % ── node styles ─────────────────────────────────────────────────────
+  digit/.style={
+    draw=#1!60!gray, fill=#1, rounded corners=2pt,
+    minimum width=1.05cm, minimum height=0.62cm,
+    font=\small\bfseries, inner sep=2pt, text=#1!20!black,
+  },
+  subtask/.style={
+    draw=#1!60!gray, fill=#1, rounded corners=2pt,
+    minimum width=1.05cm, minimum height=0.55cm,
+    font=\footnotesize, inner sep=2pt, text=#1!20!black,
+  },
+  token/.style={
+    draw=violet!55, fill=violet!8, rounded corners=2pt, dashed,
+    minimum width=1.05cm, minimum height=0.55cm,
+    font=\footnotesize\bfseries, inner sep=2pt, text=violet!70!black,
+  },
+  rowlabel/.style={font=\footnotesize\itshape, text=gray!70!black, anchor=east},
+  poslabel/.style={font=\scriptsize, text=gray!60!black},
+  carry/.style={->, >=stealth, thick, color=orange!70!red,
+                shorten <=3pt, shorten >=3pt},
+]
+
+% ── colours (Quirke subtask families) ───────────────────────────────────
+\colorlet{cSA}{green!22!white}
+\colorlet{cSC}{yellow!48!white}
+\colorlet{cUC}{blue!22!white}
+\colorlet{cUS}{blue!36!white}
+
+% ── column spacing ───────────────────────────────────────────────────────
+\def\cs{1.45}   % inter-column distance (cm)
+
+% ── data (d0 = MSB/overflow on left, d6 = LSB on right) ─────────────────
+%   d0    d1    d2    d3    d4    d5    d6
+%   ---   9     5     9     2     7     1     (Addend A)
+%   ---   0     4     0     7     5     6     (Addend B)
+%   1     0     0     0     0     2     7     (Answer)
+%   UC    US    US    US    US    SC    SA    (Subtask)
+%   t2    t2    t6    t2    t1    t16   t3    (SoRL token)
+
+% ── position labels ──────────────────────────────────────────────────────
+\foreach \i/\lbl in {0/$d_0$,1/$d_1$,2/$d_2$,3/$d_3$,4/$d_4$,5/$d_5$,6/$d_6$}{
+  \node[poslabel] at (\i*\cs, 3.15) {\lbl};
+}
+
+% ── row labels ───────────────────────────────────────────────────────────
+\node[rowlabel] at (-0.65, 2.5)  {Addend $A$};
+\node[rowlabel] at (-0.65, 1.8)  {Addend $B$};
+\node[rowlabel] at (-0.65, 0.85) {Answer};
+\node[rowlabel] at (-0.65, 0.1)  {Subtask};
+\node[rowlabel] at (-0.65,-0.65) {Token};
+
+% ── operand digits (d1..d6; d0 is the overflow, has no operand digits) ───
+\foreach \i/\d in {1/9,2/5,3/9,4/2,5/7,6/1}{
+  \node[font=\small, text=gray!30!black] at (\i*\cs, 2.5) {\d};
+}
+\node[font=\small\bfseries, text=gray!50!black] at (-0.4*\cs, 1.8) {$+$};
+\foreach \i/\d in {1/0,2/4,3/0,4/7,5/5,6/6}{
+  \node[font=\small, text=gray!30!black] at (\i*\cs, 1.8) {\d};
+}
+
+% ── horizontal rule ──────────────────────────────────────────────────────
+\draw[gray!50, thin] (-0.55*\cs, 1.38) -- (6.55*\cs, 1.38);
+
+% ── answer digit boxes ───────────────────────────────────────────────────
+\node[digit=cUC] (a0) at (0*\cs, 0.85) {1};
+\node[digit=cUS] (a1) at (1*\cs, 0.85) {0};
+\node[digit=cUS] (a2) at (2*\cs, 0.85) {0};
+\node[digit=cUS] (a3) at (3*\cs, 0.85) {0};
+\node[digit=cUS] (a4) at (4*\cs, 0.85) {0};
+\node[digit=cSC] (a5) at (5*\cs, 0.85) {2};
+\node[digit=cSA] (a6) at (6*\cs, 0.85) {7};
+
+% ── subtask boxes ────────────────────────────────────────────────────────
+\node[subtask=cUC] at (0*\cs, 0.1) {UC};
+\node[subtask=cUS] at (1*\cs, 0.1) {US};
+\node[subtask=cUS] at (2*\cs, 0.1) {US};
+\node[subtask=cUS] at (3*\cs, 0.1) {US};
+\node[subtask=cUS] at (4*\cs, 0.1) {US};
+\node[subtask=cSC] at (5*\cs, 0.1) {SC};
+\node[subtask=cSA] at (6*\cs, 0.1) {SA};
+
+% ── SoRL token boxes ─────────────────────────────────────────────────────
+\foreach \i/\t in {0/t2,1/t2,2/t6,3/t2,4/t1,5/t16,6/t3}{
+  \node[token] at (\i*\cs, -0.65) {\texttt{\t}};
+}
+
+% ── carry arrows (cascade flows right→left: d5→d4→d3→d2→d1→d0) ──────────
+\foreach \fr/\to in {5/4, 4/3, 3/2, 2/1, 1/0}{
+  \draw[carry] (a\fr.west) -- (a\to.east);
+}
+
+% ── cascade bracket + label ──────────────────────────────────────────────
+\draw[orange!60!red, thin]
+  (a5.north west) -- ++(0, 0.22)
+  -- (a0.north east) -- ++(0,-0.22);
+\node[font=\scriptsize\itshape, text=orange!60!red]
+  at (2.5*\cs, 1.35) {carry cascade};
+
+% ── legend ───────────────────────────────────────────────────────────────
+\matrix[
+  matrix of nodes,
+  nodes={font=\scriptsize, inner sep=2pt, anchor=west},
+  row sep=1pt, column sep=4pt,
+  anchor=south east,
+] at (6*\cs + 0.6, -1.05) {
+  \node[digit=cSA, minimum width=0.45cm, minimum height=0.3cm,
+        font=\scriptsize] {}; &
+  \node {SA --- simple add}; &
+  \node[digit=cUC, minimum width=0.45cm, minimum height=0.3cm,
+        font=\scriptsize] {}; &
+  \node {UC --- uses carry}; \\
+  \node[digit=cSC, minimum width=0.45cm, minimum height=0.3cm,
+        font=\scriptsize] {}; &
+  \node {SC --- generates carry}; &
+  \node[digit=cUS, minimum width=0.45cm, minimum height=0.3cm,
+        font=\scriptsize] {}; &
+  \node {US --- cascade}; \\
+  \node[token, minimum width=0.45cm, minimum height=0.3cm,
+        font=\scriptsize] {}; &
+  \node[text=violet!70!black] {\sorl{} token}; & & \\
+};
+
+\end{tikzpicture}
+\caption{%
+  Six-digit addition $959{,}271 + 040{,}756 = 1{,}000{,}027$, a four-deep
+  carry cascade. At each answer-digit position \sorl{} assigns one
+  abstraction token (bottom row). Tokens \texttt{t2} and \texttt{t6}
+  cluster on cascade positions (UC/US); \texttt{t16} marks the carry
+  source (SC); \texttt{t3} marks the trivial position (SA).
+  Token assignments from model \texttt{add\_sub\_sorl\_v1\_abs30\_K1\_100K}
+  (K=1, 30-token codebook).%
+}
+\label{fig:arithmetic-example}
+\end{figure}
+"""
+
 HARD_SPLITS = ["add_C4", "add_C5", "add_C6", "sub_M4", "sub_M5"]
 ALL_SPLITS = [
     "add_S0", "add_S1", "add_S2", "add_S3", "add_S4", "add_S5", "add_S6", "add_random",
@@ -719,6 +866,11 @@ hidden activations — but here it is readable directly from the token sequence.
             gr.Textbox(value=LATEX_ARITHMETIC_SETUP, label="arithmetic_setup.tex",
                        lines=30, show_copy_button=True, interactive=False)
 
+            gr.Markdown("#### § Carry-cascade example figure (TikZ)")
+            gr.Markdown("Requires: `\\usepackage{tikz}`, `\\usetikzlibrary{matrix}`, `\\usepackage{xcolor}`, and `\\providecommand{\\sorl}{\\textsc{DLR}}`.")
+            gr.Textbox(value=LATEX_FIGURE_EXAMPLE, label="fig_arithmetic_example.tex",
+                       lines=40, show_copy_button=True, interactive=False)
+
         # ── Tab 4: About ──
         with gr.TabItem("About"):
             eval_info_md = gr.Markdown("")