AI Substrate Problem: Why Iterative Generation Fails, and What Any Fix Must Satisfy

This position paper argues that the failure of iterative work with generative models lies not in the model but in the substrate it operates on: a generative session keeps no lasting record of decisions, no dependency structure among them, and no way to mark a decision as finished. The underlying principle, that iterative work should preserve prior decisions rather than overwrite them, is already established in creative and engineering tools as non-destructive editing. The contribution here is not to name that problem but to unify it: the paper derives four properties any adequate non-destructive generative substrate must provide (selective rebuild, stable parts, dependency awareness, and durability across sessions and models), argues each is necessary, and uses them as a single yardstick, independent of modality, to evaluate seven families of existing system. Tools that provide three or four of the properties do so only for hand-authored content such as spreadsheets and engineering models; tools that operate on generated content provide at most one or two. No available system provides all four for generated content. This is a position paper: it offers no empirical benchmark and no demonstrated construction, and states these limitations explicitly.