Inheritance Architecture: Recursion, Cost, and the Safety of Development in Long-Horizon HAI

Long-horizon human–AI interaction (HAI) emerges when the outcomes of joint work retain functional force across changes in sessions, tasks, context windows, and technical conditions, and continue to participate in the production of subsequent actions. This preprint introduces inheritance architecture as a framework for analysing how consequences of prior interactions gain, retain, transform, or lose functional weight in a system’s subsequent trajectory. Long-horizon HAI is treated as a distributed recursive system. The framework distinguishes regulatory, trajectory-level, and organizational recursion; separates thresholds from developmental Gate events; and operationalizes inheritance through the unit “move — distinguishable consequence — carrier — next functionally comparable opportunity — altered choice,” an L0–L5 scale of inheritance depth, and a separate V axis for Gate events. Longitudinal observation of one long-horizon HAI is supplemented by pilot coding of 12 external reports of discontinuities following platform changes. The analysis distinguishes memory, local application, and trajectory-level inheritance; parasitic return cost from functional cost; and locally capable output from compatibility with an already developed human–AI system. It argues that the safety potential of recursion depends on the distributed architectures of attention, presence, and return, together with preserved model-side variability, human-side selection, consequence checking, and the ability to stop, refuse, and revise. The accompanying codebook and pilot coding matrix make the framework available for independent application and further validation. The deposit contains the English-language preprint, the authoritative Russian-language original, the Recursive Dynamics Codebook (Supplementary Material S1), and the pilot coding of external cases (Supplementary Material S2). In the event of interpretive divergence between the language versions, the Russian-language text is authoritative. Contact: emitter.resonance@gmail.com