Topology‑aware augmentation of language‑derived semantic embeddings improves neural predictivity in higher visual cortex: a registered analysis plan

Recent work reports that higher‑level visual representations in the human brain align with embeddingsfrom large language models (LLMs), as assessed by representational similarity analysis (RSA) and voxel‑wiseencoding. We outline a conservative, falsifiable study that (i) replicates those findings using public fMRIdatasets and (ii) tests whether augmenting caption‑based LLM embeddings with topology‑aware featuresimproves neural predictivity in high‑level visual cortex. The topological features summarize coarse‑grainedorientation‑reversal structure of stimuli (a proxy for non‑orientability intensity ) via persistent‑homologydescriptors and texture‑flow parity cues. We define hypotheses, metrics (RSA, noise‑ceiling‑correctedcorrelations, CKA), preregistered statistics, and a small synthetic Klein‑illusion stimulus set to probemonotonic effects of . The goal is a short, reproducible Zenodo paper plus open code and stimuli.