When Does a Perturbation Model Know Enough?

Single-endpoint Perturb-seq data describe transcriptional responses without identifying the biological processes that produced them. We investigate whether layerwise response predictions provide a reproducible computational descriptor of these endpoints. Inspired by Adaptive Computation Time and PonderNet, RefineRx fits an oracle-supervised halting head and a refinement token on STATE backbones with frozen pretrained weights. All eight layers execute; expected halting depth summarizes a learned distribution over layer exits and does not save computation. Reanalysis of released four-seed arrays gives mean pairwise Spearman correlations of 0.761–0.854 across K562, HepG2, Jurkat, and RPE1. These estimates include training and validation perturbations and measure stability within fitted model–data systems. In released outputs from two CD4⁺ T-cell donors, cross-donor agreement is \(-0.110\) at rest, \(-0.097\) at 8 h, and \(0.486\) at 48 h after stimulation. At 48 h, perturbations annotated as resting-sparing suppressors have modestly lower depth than damaging or other perturbations (Cliff's \(\delta=-0.162\) and \(-0.239\)), with little separation from the inflammatory category (\(\delta=-0.011\)). These exploratory comparisons use preassigned transcriptomic labels and do not establish therapeutic effects. The results support expected layer-exit depth as a stable model-conditioned descriptor in the evaluated settings, without establishing biological complexity, causal depth, predictive superiority, or generalization to new donors.