Contract-Preserving Lower-to-Upper Recalibration in Hierarchical Agent Systems: An Integrated Framework

Abstract

This paper addresses hierarchical agent systems with persistent state, tool authority, upper maps, and an auditable runtime, in which verified lower-level recovery reveals errors in upper-level models or cross-level translation. It proposes an integrated rule that, under a fixed external contract, restricts revisions to upper maps, compasses, translators, or observation models according to what the residual evidence identifies. Approval binds the proposed change to current state, evidence, authority, and execution semantics. Soft and hard interventions along multiple axes, restricted-authority probes, disclosure obligations, lineage, compensation, and conditional re-entry govern this process. Conditional results include tracking bounds for fixed and moving targets with explicit approval-cadence assumptions, safety and paired-improvement certificates over nonempty response-model sets, and approval integrity within a declared transaction boundary. A restricted representation bottleneck is extended to a finite-feature first-proposal ceiling and an expected-cost condition that includes the initial repair cost. Feedback structure, descriptive audit gaps, and causal harm are distinguished. The paper specifies competing explanations and measurable falsifiers; it does not establish component novelty, necessity of the full architecture, universal convergence, or empirical superiority. Scope and status. Version 1.0 working paper, dated 19 September 2026, about 9,200 words; unsubmitted. This is the flagship of the Integrated Framework series on contract-preserving lower-to-upper recalibration (SHDA). The upload also contains the Technical Supplement (about 12,500 words), which formalizes the typed state, contracts, commit gates, conditional theorems and execution schemas that the flagship cites as P1-P8. Three companion papers are archived separately: Companion A on residual genesis and dynamic feedback route attribution, Companion B on intervention-disclosure visibility and time-bounded selective nondisclosure, and Companion C on family-scoped capability control, typed lineage and atomic re-entry. The results are conditional arguments and measurement specifications; no empirical study, deployment result, or safety guarantee is reported.