The zero-drift property demonstrated for diffusion models in VDR-26 — where arithmetic error does not accumulate across sequential computation chains — applies to any domain where each step's output feeds the next step's input. This paper maps VDR exact arithmetic to twelve computational domains beyond language models: autoregressive generation (speech, music, protein), normalizing flows, Kalman filtering and state estimation, cryptographic protocols, financial computation, control systems, physics simulation, blockchain and consensus, geodesy and navigation, game theory and mechanism design, digital signal processing, and quantum computing primitives. In every domain, the structural problem is the same: float arithmetic introduces per-step error that compounds through the chain. VDR eliminates the per-step error entirely. The remaining errors — model approximation, measurement noise, basis set truncation — are the domain's problems, not the arithmetic's. No prior reading is required. VDR arithmetic concepts are summarized where first used; full specifications are in VDR-1 and VDR-14.
Geoffrey Howland (Fri,) studied this question.