06_costs · 完整稿件源码

% 06_costs.tex -- section 6, budget ~1 page (Table 1 is placed in section 4 and referenced here). Polished 2026-09-02.
% Numbers: NARRATIVE_REPORT.md H2, H3, D2; data/report_excerpt.json; figures/fig5_data.json.
\section{What it costs and what it buys}
\label{sec:costs}

\subsection{Token footprint}
The input-token footprint is part of the fingerprint.
Median cumulative input per \SWEPro task falls from 3.27\,M tokens (\nTwoSeven) to 1.52\,M (\sTwoFiveSix), $-53$\,\%: 32\,\% fewer steps times 29\,\% less context per step, with whole-file reads carrying 74\,\% of the drop in tool output (Figure~\ref{fig:waterfall}, Appendix~\ref{app:extras}); rewriting file reading in the harness reaches 9--50\,\% \citep{zheng2026corvus,zhang2026fastcontext}.
The post-trained MoE line moves the same way (3.27\,M to 2.05\,M) and so does the bare base (3.40\,M to 2.44\,M); every \vTwelve and \vThirteenCorpus checkpoint of the dense line lands at 1.52--1.80\,M regardless of its score.
Output tokens rise by 16--49\,\%, reflecting the thinking overshoot of \S\ref{sec:inherit} (Table~\ref{tab:C6}).
These are input-token counts, not bills: under prompt caching the two diverge \citep{weinberger2026token}.

\subsection{\passat}
The corpus on a bare base improves \SWEPro by $+8.9$\pp (\aTwoEp against \aBase; Table~\ref{tab:arms}); its $+3.4$\pp on \SWEVerified is below the detectable effect.
The bare-base gain exceeds the post-trained MoE gain by $6.2$\pp (standard error 2.3\pp), a post-hoc contrast that also crosses generation, dose and learning rate (\S\ref{sec:setup}), so it resolves that the gains differ, not why.
Against post-trained bases the differences span $-2.2$ to $+2.8$\pp in both pools and both families, all below the detectable effect; excluding the 76 teleport tasks changes no decision (Appendix~\ref{app:noise}).
The generation jump (\qThreeEight against \nTwoSeven) gives $+5.6$\pp on Pro and $+5.8$\pp on Verified, more than any checkpoint on the post-trained dense line, at 4.41\,M input tokens per task ($1.35\times$ the native model); against the two-epoch student the Pro margin is $+2.9$\pp at $2.9\times$ the input.
The bare-base gain is consistent with the corpus supplying part of the agentic post-training transition on this base, the diminishing return on stronger starting models that curated-trajectory SFT also shows \citep{codearts2026yet}.
The composite \vThirteenCorpus arm gives $+1.3$\pp on Pro and $-5.0$\pp on Verified against \sThreeEightOne, the one composition signal at the design's detectable effect (Appendix~\ref{app:noise}); as a composite change it is not attributable, and the arm's fingerprint and footprint stay within the \vTwelve range.