Dillavou通话材料v3: 机制段按RESULT 77-79改判(读出地板+零成本修复+小尺寸EP胜BP), Q2改为板上对比读出/physical epsilon, cheat sheet加破案更新节

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn

1 files changed, 82 insertions, 0 deletions

diff --git a/docs/hardware/CALL_SHEET_DILLAVOU.tex b/docs/hardware/CALL_SHEET_DILLAVOU.tex
new file mode 100644
index 0000000..5bb9ce9
--- /dev/null
+++ b/docs/hardware/CALL_SHEET_DILLAVOU.tex
@@ -0,0 +1,82 @@
+\documentclass[10pt]{article}
+\usepackage[letterpaper,margin=0.72in,top=0.6in,bottom=0.6in]{geometry}
+\usepackage[T1]{fontenc}
+\usepackage{newpxtext,newpxmath}
+\usepackage{booktabs,array,enumitem,xcolor,titlesec}
+\definecolor{accent}{RGB}{138,31,31}
+\titleformat{\section}{\normalsize\bfseries\color{accent}}{}{0pt}{}
+\titlespacing{\section}{0pt}{9pt}{4pt}
+\setlist[enumerate]{leftmargin=1.2em,itemsep=1.5pt,topsep=2pt,parsep=0pt}
+\setlist[itemize]{leftmargin=1.1em,itemsep=1.5pt,topsep=2pt,parsep=0pt}
+\pagestyle{empty}
+\setlength{\parindent}{0pt}
+\setlength{\parskip}{3pt}
+\begin{document}
+
+{\large\bfseries EP transformer program $\rightarrow$ analog board: call sheet}\\
+{\small supplements the July 12 brief and build plan \,\textperiodcentered\, Yuren Hao (UIUC) \,\textperiodcentered\, August 2026}
+
+\section{Program state since the July brief (measured)}
+\begin{itemize}
+\item \textbf{72M-parameter transformer LM sealed:} EP final CE 3.364 (2 seeds) vs.\ a tuned
+backprop twin at 3.288 (3 seeds, $\pm$0.006); gap $+$0.076 nats ($\approx$8\% perplexity). The
+symmetric two-phase estimator narrows this to $+$0.044 ($\approx$4.5\%) at 1.67$\times$ cost. Zero
+instability events in both EP seeds. To our knowledge the largest LM trained with no
+backpropagation anywhere in the loop.
+\item \textbf{135M trained end to end} (2.7B tokens), also with zero instability events. It surfaced
+a width-dependent estimator loss localized to the upper layers --- which a per-layer probe has since
+traced to the simulator's own \emph{readout floor}: at larger widths the nudge displacement falls
+below single-precision machine epsilon relative to the activations and is rounded away in the
+contrast read.
+\item \textbf{The fix is free, and it is the software version of differential readout:} read the
+contrast from the stored displacement itself, never as the difference of two large signals. Verified
+against a double-precision control and a pre-registered amplitude-shift test; the loss closes 100\%
+at zero cost. At the smaller rungs of our scaling ladder, EP now matches or beats the matched
+backprop twin (2 seeds per arm).
+\item \textbf{Random-sign nudging} (nudge polarity flipped randomly per batch) is the stability
+winner at both scales: it removes the systematic single-sided bias at zero extra cost. On a board it
+is a polarity switch.
+\end{itemize}
+
+\section{The four primitive measurements (recipe-independent)}
+\begin{small}
+\begin{tabular}{@{}p{0.28\textwidth}p{0.40\textwidth}p{0.24\textwidth}@{}}
+\toprule
+\textbf{measurement} & \textbf{simulation anchor (measured)} & \textbf{board question}\\
+\midrule
+1. contrast-update fidelity (exact difference-of-squares channel) & update tracks the \emph{faulted}
+network at cos $\approx$ 0.97 under 8-bit weights, 1\% forward noise, 10\% nudge noise & does the
+same co-adaptation law hold in physics\\[2pt]
+2. nudge operating window and its floor & ceiling tracks loop gain and sinks over training
+($\sim 1/\sigma^2$ of the output map); error channel is amplitude-indifferent across $20\times$
+(multiplicative), so the floor is set by \emph{additive} readout noise & measure the physical floor;
+track window drift over training\\[2pt]
+3. sign-only vs.\ exact vs.\ random-sign updates & sign works; random sign removes the bias at
+scale; exact channel is the reference & continuous three-way comparison on one board\\[2pt]
+4. EP current nudge vs.\ CL voltage clamp & the distinction formalized by McGinnis, Li, and Mori &
+same board, same task, same metrology\\
+\bottomrule
+\end{tabular}
+\end{small}
+
+\section{Where we most want your judgment}
+\begin{enumerate}
+\item Is this the right primitive-level list? What would you add or cut, from your
+imperfection-characterization experience?
+\item \textbf{Reading the contrast:} our width-scaling loss was ultimately a readout-floor artifact,
+and the lesson is that the contrast must be measured differentially, never as a difference of two
+large signals. How do your update cells read the contrast, and what sets the smallest resolvable
+contrast (the physical epsilon) on the board?
+\item \textbf{Additive vs.\ multiplicative separation:} in simulation we sweep nudge amplitude at a
+fixed task. What is the right bench protocol for the same separation on the board?
+\item \textbf{Settling as a gate:} in simulation the relaxation residual predicts every blowup and
+gates the weight update. What settling observable did your boards expose, and did you gate on it?
+\item \textbf{Twin replicas vs.\ a time-multiplexed single network} at 8--32 edges: matching cost
+against memory cost, your call.
+\end{enumerate}
+
+\vspace{2pt}
+{\small We are not committing the board before this conversation; rungs 1--3 remain under \$1k in
+parts, funded on our side. Results are intended for open publication.}
+
+\end{document}