The Throughput Basin: Cross-Substrate Convergence and Decomposition of Serial Decoding Throughput

This study examines throughput convergence across six domains—genetics, artificial intelligence, human cognition, music, language, and engineered systems—by measuring or calculating effective information per serial decoding event for 31 systems with vocabulary sizes spanning 64,000-fold. A decomposition framework Ieff = RM(ε) + Δs + ξi partitions throughput into a universal rate-distortion floor, substrate-specific slack, and system-specific residual. Across 31 systems, 94% of observed throughput values fall between 2 and 6 bits, with a median of 4.39 bits (bootstrap 95% CI: 3.82–4.67). Monte Carlo sampling (100,000 draws) places 90.5% of bio-plausible samples in the 3–6 bit band versus 28.3% for random broad sampling (p < 10−300). Three independent evolutionary simulations converge to K ~ 19–30, and a co-evolutionary simulation discovers K = 19.76 (within 6% of the ribosome's M = 21) and ε = 6.62 × 10−3 (the same order of magnitude as biological translation error rates), starting from no biological priors. Without AI models, the remaining systems do not cluster significantly tighter than alphabet sizes predict (p ~ 0.56), indicating that cross-domain convergence statistically depends on the AI data.