% 02_related.tex -- section 2, budget ~0.6 page. Organised by question, not paper by paper. Round-2 literature scan 2026-09-03 (lit/related_scan_2026-09-03.md).
\section{Related Work}
\label{sec:related}

\paragraph{Trajectory SFT for coding agents.}
Fine-tuning on agent trajectories began with search and tool-use agents \citep{chen2023fireact,zeng2023agenttuning}, where format following and reasoning were already seen to be learned at different speeds \citep{chen2024agentflan}.
Issue-resolution agents are now routinely trained on trajectories collected at scale \citep{pan2024training,yang2025swesmith,zeng2025skyworkswe,wang2025swedev,wang2025swemirror}, distilled from a stronger agent \citep{ahmad2026openswetraces}, curated down to a few hundred examples \citep{codearts2026yet}, or combined with skill priors and inference-time scaling \citep{yang2025kimidev,rastogi2025devstral,tao2026swelego}; reinforcement learning on software history is the alternative \citep{wei2025swerl}.
These recipes, and the study that asks which trajectories teach well \citep{yang2026what}, are compared by resolve rate; we keep the recipe fixed and ask what the student inherits, at what dose, and whether it shows in the score at all.

\paragraph{How trajectory corpora are curated.}
Trajectory- and segment-level selection with loss masks is used as an accuracy lever \citep{tao2026swelego,zheng2026sweprime,yang2026what}, failed trajectories can be kept under a signal \citep{wang2024learning}, the student is known to inherit the flaws of every retained step \citep{ma2026from}, and leakage of the upstream fix is a known contamination path \citep{prathifkumar2025does,liang2025the}.
Our object is the rule system itself and the behavioral profile of what each rule removes.

\paragraph{Behavioral fingerprints and distillation-induced similarity.}
Trajectories and pull requests identify the agent that produced them \citep{oderinwale2026agent,gupta2026dissecting,ghaleb2026fingerprinting}, models carry detectable idiosyncrasies \citep{sun2025idiosyncrasies}, distillation leaves detectable traces \citep{rawat2026referencebased} and moves a student's tool-use behavior toward its teacher \citep{yang2026when}, sometimes through signals invisible in the data \citep{cloud2025subliminal,dang2026subliminal,knig2026quantifying}.
Our extractor is in this family; we add the epoch dose, capability measured on the same arms, a same-task anchor and a vendor next generation.

\paragraph{Style versus capability in supervised fine-tuning.}
Imitating a proprietary model copies its style before its capability \citep{gudibande2023the}; small instruction sets change format more than knowledge \citep{zhou2023lima,lin2023the,raghavendra2024revisiting}; whether SFT generalizes depends on the starting model and on repetition \citep{chu2025sft,ren2026rethinking,kopiczko2026data}; long-chain-of-thought distillation transfers a reasoning \emph{form} \citep{yeo2025demystifying,muennighoff2025s,ye2025limo}; and trajectory SFT can raise an agent benchmark by memorizing its interface rather than its tools \citep{gu2026what}.
The dissociation we report is the agentic, teacher-relative instance of this pattern.

\paragraph{Evaluation validity.}
Resolve rate conceals trajectory structure \citep{shu2026what,mehtiyev2026beyond,majgaonkar2025understanding}, passes wrong patches through weak tests \citep{yu2026sweabs}, is flaky across runs \citep{wang2026contractbench,yagubyan2026how} and needs paired error bars \citep{miller2024adding}; confident endings mask silent failure \citep{advani2026from}, fewer tokens need not mean a smaller bill under prompt caching \citep{weinberger2026token}, and SWE-bench results are exposed to memorization \citep{prathifkumar2025does,liang2025the}.