Contextual, Layer-Resolved Language Features Predict Cortical Activity Across Divergent Developmental Sensory Experience

Congenital blindness and congenital deafness are natural experiments in altered developmental sensory experience: the brain learns to represent the world from a fundamentally different early sensory diet than typical development provides. We use this setting to ask how altered developmental experience reshapes learned neural representations, using a frozen, layer-resolved large language model as the measurement tool rather than as a model of development itself. In the 101 Dalmatians dataset (50 subjects, including congenitally blind and congenitally deaf groups who each experienced one sensory edit of the same movie), we extract per-layer Qwen3-Omni hidden states from the movie's transcript and fit per-subject ridge encoding models into cortex. A middle-layer feature set predicts cortical activity better than a static sentence embedding in every group (whole-cortex Pearson $r$ 0.047–0.085 vs. 0.029–0.058), showing that contextual language representations capture more brain-relevant structure than a fixed embedding regardless of a group's sensory history. Within pairs of groups who experienced an identical stimulus edit, two exploratory contrasts point the same way: motion energy predicts somatomotor cortex more strongly in the congenitally deaf group ($q=0.04$), and, as a trend only, language features predict visual cortex more strongly in the congenitally blind group ($q=0.10$). Only the former clears $q<0.05$, and it is a low-level visual control rather than a language feature; both are corrected within a stimulus-matched pair, and neither survives a correction family spanning all 56 feature sets tested. At $n=9$–11 per group we therefore read them as hypotheses rather than findings: suggestive that development builds representations from whichever experience is actually available rather than toward one fixed target, and that a system learning from incomplete or unusual input might do best the same way, by adapting around what is there rather than failing – a possibility this dataset can motivate but not establish.