Relational Generative Computing: From Rule–Relation–History Dynamics to Effective State Formation and Bidirectional Computational Execution

Conventional computation usually treats rules, relational structure, history and observation mappings as application-level data executed over a predefined machine state. We investigate a different computational abstraction in which rule, relation, information and retained history participate directly in the formation of computational state, persistent relational organizations may be promoted to reusable effective states only after explicit admission tests, and the same execution record supports both forward generation and reverse diagnosis. The framework distinguishes Relation from Information, instantaneous state from retained History, persistence from predictive closure, and successful forward output from mechanism-level identifiability. We define a role-relative effective-state admission vector containing persistence, prediction, intervention, history, recovery and causal-ancestry tests, and show why these criteria are non-compensatory. Four publication-safe controlled benchmarks illustrate the computational consequences: retained history improves prediction when present input is insufficient; admission prevents invalid coarse-state reuse while reducing unnecessary local updates; bidirectional traces improve localization of rule-, relation- and history-class faults relative to forward-output inspection; and versioned rule activation reduces persistent corruption compared with silent runtime mutation. These experiments are synthetic and do not establish a new physical ontology or hardware advantage. They support the narrower claim that relational, historical and admission-aware state semantics can define a falsifiable computing framework for sparse multiscale simulation and auditable scientific computation. Patent-protected production runtime, instruction-set, compiler, trust-domain and hardware-mapping details are not disclosed.