08_conclusion · 完整稿件源码

% 08_conclusion.tex -- section 8, budget 0.2 page. Accepted overall wording (NARRATIVE_REPORT section 7) + D1. Polished 2026-09-02.
\section{Conclusion}
\label{sec:conclusion}

In this single teacher--corpus study, \vTwelve SFT on 8,137 trajectories moved command vocabulary, tool composition and protocol-error recovery toward the teacher in two Qwen families, on the evaluation pools and on the same tasks the teacher ran, with most dense-model change after one epoch, and cut median cumulative input tokens by 45--53\,\%.
\passat depended on base and pool: $+8.9$\pp on Pro and $+3.4$\pp on Verified for a bare base, and $-2.2$ to $+2.8$\pp against post-trained bases, the latter two below what 731 and 500 paired tasks can resolve.
The composite \vThirteenCorpus variant stayed inside the \vTwelve fingerprint and footprint range.
One epoch buys the habits and half the input tokens; transfer is verified on the fingerprint, not on \passat.