Information Physics of Generative Systems: A Unified Framework for Scheduling, Weighting, Tokenization, Inference, and Training Dynamics

The design choices of modern generative systems-noise schedules, EMA weights, VQ codebooks, classifier-free guidance, sampling temperature, learning-rate schedules-are generally treated as independent engineering decisions. This paper shows that they all realize a single information-physics structure with two axes: a density axis (where information flows) and a retention axis (how long information persists). Machine-verified analyses of public data show: (1) noise schedules live on a decay-shape spectrum (classical DDPM linear = exponential, R^2=0.979; scaled-linear tilts power-law); (2) EMA weights form a two-timescale retention spectrum with tau_ema*lr ~ 1.00 landing precisely on the layer-matching sweet spot; (3) VQ codebooks are density clusters (utilization 2-6% implies Zipf alpha~1.6-1.8, vs alpha~0.9 for language-image information is more cluster-concentrated); (4) inference retention differs by paradigm (autoregressive R~12 transfer tax vs diffusion R~0.15 in-place); (5) CFG is a convergence-force knob (U-shaped FID curves, stage-dependent optimum = four-state lifecycle optimal control); (6) training loss is adsorption dynamics: the universal 1/3 time-scaling of LLM training loss (2026 theoretical derivation + measurements) equals our framework's slow-asymptotic relaxation exponent 1/3, and the two derivations are the same ODE in two parametrizations (beta ~ 1/delta duality), with softmax being the Boltzmann form of density clustering. The framework unifies scheduling, weighting, tokenization, inference, guidance, sampling, and training under one relaxation universality class spanning physical damping, learning, and memory decay.