\section{Related work and novelty positioning}\label{sec:related}

We re-verified every reference at execution time from primary sources (arXiv metadata and full texts, official
repositories and licences; Appendix~\ref{app:literature} and \texttt{literature/novelty\_audit.md}). Four strands
are directly relevant.

\paragraph{Adaptive depth and latent recurrence.}
Universal Transformers \citep{dehghani2019universal} and ACT \citep{graves2016act} introduced weight-shared
recurrence over depth with learned per-position halting; PonderNet \citep{banino2021pondernet} made halting
probabilistic. Mixture-of-Depths routes tokens around blocks with a fixed capacity \citep{raposo2024mixtureofdepths},
LayerSkip trains early exits \citep{elhoushi2024layerskip}, and CALM calibrates confidence-based exits
\citep{schuster2022calm}. The recurrent-depth line scaled latent recurrence to language model pretraining:
Huginn \citep{geiping2025recurrentdepth}, Ouro \citep{zhu2025ouro}, Mixture-of-Recursions \citep{bae2025mixtureofrecursions},
looped Transformers \citep{giannou2023looped,saunshi2025reasoning}, and in 2026 RecurTrace
\citep{wang2026recurtrace}, Elastic Latent Recurrence \citep{ladhak2026elastic}, Think-at-Hard
\citep{fu2026thinkathard,you2026tah2} and several contemporaries listed in Appendix~\ref{app:literature}. All of
these refine the \emph{entire} hidden state of a position (globally), and route depth per token or per sequence.

\paragraph{Slots, workspaces and sparse writes.}
Recurrent Independent Mechanisms \citep{goyal2019rims} and the Shared Global Workspace
\citep{goyal2022workspace} update a small number of competing modules per step; Slot Attention
\citep{locatello2020slot} and Perceiver / Perceiver AR \citep{jaegle2021perceiver,hawthorne2022perceiverar} maintain
a small bank of latents that cross-attend to a long input; Sparse Access Memory \citep{rae2016sparse} writes
sparsely into an external memory. Most recently ReFlux \citep{ye2026reflux} writes router-selected sparse
deep-to-shallow deltas into a frozen backbone, with the backbone unchanged when no route is active. \prisme{}'s
workspace differs in being \emph{ephemeral} (reset per token), separate from a read-only cached bank, and edited by
discrete 1--2-slot actions rather than by competition among modules.

\paragraph{Loss-gain supervision of routing.}
The supervision signal we use---the realised change in next-token loss of an executed extra branch---has become
standard for \emph{binary} continue/stop decisions in 2025--2026: Ouro's second stage fits its exit gate to observed
per-token loss improvements \citep{zhu2025ouro}; TaH2 calls the same construction ``lookahead depth supervision''
\citep{you2026tah2}; RecurTrace trains a halting head on an oracle ``more depth still helps'' label
\citep{wang2026recurtrace}. Earlier, fitted value iteration and counterfactual risk minimisation established the
use of logged propensities for off-policy action evaluation \citep{ernst2005tree,dudik2011doublyrobust,swaminathan2015crm}.
\prisme{} makes the label \emph{action-conditional} over 24 spatial actions, logs propensities, and separates gain,
cost, uncertainty and harm heads.

\paragraph{Zero-initialised identity insertion.}
Adding a zero-initialised branch so that a new module starts as the identity is a well-established device
\citep{bachlechner2020rezero,alayrac2022flamingo,zhang2023controlnet}; exact bypass of unrouted tokens exists in
Mixture-of-Depths and in TaH/TaH2, and ReFlux leaves the frozen backbone unchanged with no active route. The
differential readout of Section~\ref{sec:readout} is a modest variant of this device and is \emph{not} claimed as a
contribution.

\paragraph{What is and is not new.}
After this audit, none of the three ingredients of \prisme{} is new on its own. What has no direct predecessor is
their combination: a per-token slot workspace with selected local writes, trained with action-conditional
counterfactual gains over a 25-action catalogue with logged propensities, behind an exact zero-budget bypass.
Accordingly, the contribution of this paper is empirical and methodological---a controlled test of whether spatial
selection supervised by counterfactual gains buys anything over uniform, dense or global alternatives at matched
cost---rather than architectural priority. Table~\ref{tab:novelty} in Appendix~\ref{app:literature} gives the
claim-by-claim comparison.