\section{Conclusion}\label{sec:conclusion}
The central question was whether counterfactual supervision of \emph{where} to edit a small latent workspace
improves the quality--compute frontier beyond uniform workspace computation and adaptive global recurrence.
On the evidence collected here the answer is no, and the pre-registered decision rule says so: on the locked test regimes the primary effect is $\testprismebtwovsbzero$ points against a three-point threshold. The sparse editor is learnable and helps a frozen backbone
by a few percent of loss and two points of exact match when it is applied once or twice at every position, but
a global-recurrence adapter trained with the identical recipe helps at least as much at every budget; where to
write barely matters for fixed schedules; an exhaustive oracle shows large opportunity (up to
\routingoraclebeight{}\,\% of loss) that a value controller trained on 100k executed counterfactual labels
cannot capture, because the opportunity sits in high-uncertainty states where a wrong action costs more than a
right one gains and the controller's within-state ranking is close to chance; and at small scale the
per-action dispatch overhead erases the arithmetic savings of sparsity. The exact bypass holds bitwise, the
differential readout needs a scale-aware bootstrap, and the parameter and operation accounting of the proposal
is reproduced. The scientifically justified next experiment is not a larger model but a better label: a
controller trained to rank actions within a state (pairwise or listwise, with the state's mean gain removed)
on labels collected at the scale where the editor's arithmetic is not dwarfed by dispatch, with the
global-recurrence adapter as the baseline to beat.