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

Headline. The data-driven throughput equation established by Paper 7 for language — BPT ≈ sourceₑntropy − f (structuraldepth) — generalizes to vision and audio. Each modality has its own characteristic throughput (vision ~1. 33 bits/pixel, real speech ~1. 89 bits/meldim, language ~1 bit/sourcebyte) but every modality obeys the same form: noise extracts zero bits, structured data extracts source-entropy mi- nus exploitable hierarchy, and the structural bonus is real and reproducible. The ver- ification was built in seven rounds of escalating adversarial pressure (§1. 3), cul- minating in a 112M-parameter ViT-MAE trained from scratch on a controlled-entropy ladder that confirms the equation at the architectural level with Welch’s t = 249, 994, p = 1. 6×10−11, and Cohen’s d = 204, 119 — the visual analog of Paper 7’s SYN-8 result.