The Vision Basin: Cross-Modal Throughput Measurement Reveals Modality-Specific Information Extraction Rates

Paper 7 established that the language-model throughput basin reflects training-data entropy minus exploitable structure: BPT ≈ source entropy − f(structural depth). If this equation is general, different modalities should sit at different throughput levels. We test this across language, vision, and audio using 12 pretrained models, from-scratch training, source-entropy measurement via lossless compression, and structural-destruction cascades. Each modality has its own characteristic throughput: language at 0.85–1.30 bits per source byte (9 models, 2 corpora), generative vision at 1.33–1.36 bits per pixel (MAE reconstruction), and real speech at 1.89 bits per mel-spectrogram dimension (LJ Speech). Controls validate the framework: the structural bonus vanishes for noise across all modalities. A visual shuffling cascade on STL-10 reveals a spatial structural bonus of 0.69 bits/pixel—confirming that visual hierarchy is exploitable, though smaller than language's ~6.7 bits/token. Music (synthetic) extracts 2.69 bits per mel-dimension, more than speech, consistent with richer harmonic structure. The throughput-basin equation holds qualitatively across modalities, with modality-specific extraction rates set by each domain's exploitable structure; the per-modality units are not directly comparable.