\section{Limitations}
\label{sec:limitations}

\paragraph{Survivorship.} The universe requires six months of presence after the sample start and, more
importantly, 80\,\% regular-hours minute coverage over 2018--2025; only five delisted candidates passed because the
vendor's delisted one-minute archives are only partly ingested in FinData~OS. The results therefore describe liquid
instruments that survived the sample window. Nothing in the architecture depends on this choice, but the transfer
claim (Hypothesis~\ref{h2}) is a claim about unseen \emph{liquid survivors}, not about the whole market.

\paragraph{Bars are not quotes.} FirstRate bars carry no bid--ask spread, trades or order flow. Microstructure
information that is known to matter for minute-level prediction \citep{sirignano2019universal, kolm2023deep} is
absent, and the economic evaluation of Hypothesis~\ref{h5} rests on assumed costs.

\paragraph{Regular hours only.} Origins whose horizon crosses the close are excluded, extended hours are not
studied, and the closing print is excluded from context and targets. The model has never seen the first thirty
minutes as an origin (the first origin is at 10:00), although those minutes are in its context.

\paragraph{Missingness and targets.} Missing minutes are masked rather than filled; the target of a minute that
follows a gap is the return accrued since the last observed bar, so its scale is slightly larger than a one-minute
return. With coverage above 80\,\% and the 80\,\% future-presence rule this affects a small fraction of targets, and
it affects all models identically, but it is a modelling convention, not a property of the market.

\paragraph{Corporate actions.} Splits are applied explicitly; dividends are not, and the overnight gap across an
ex-dividend date is left raw. Mergers and spin-offs end a series without notice.

\paragraph{Compute and scale.} A single Apple Silicon node without mixed precision bounds the number of
optimiser steps and seeds. The brief's upper range (50--100~million parameters) was not explored; the Base model
is the largest size trained. Training-time estimates are measured, not extrapolated, but the comparison between
sizes is at matched steps rather than matched wall-clock.

\paragraph{Baselines not run.} Pretrained time-series foundation models, iTransformer and EGARCH were not evaluated
(Section~\ref{sec:design}); their absence is a limitation of this version, not evidence about them.

\paragraph{Multiplicity and overlap.} Evaluation origins overlap within a session and are cross-sectionally
dependent; all inference uses day clusters and HAC variances, and multiple-testing adjustments are reported, but the
effective sample size is far smaller than the number of origins suggests. Readers should weigh interval estimates,
not point estimates.