Physical AI Lifecycle Management: A Derivation Framework from Generic AI Lifecycle Baselines

Physical AI systems interact with their physical environments through closed loops of perception, decision-making, and actuation. Because the term Physical AI has no universally accepted definition, a lifecycle management (LCM) framework must make its conceptual assumptions explicit. This paper develops a testable framework through analytical derivation and addresses four research questions: To what extent can generic AI lifecycle stages be applied to Physical AI (RQ1)? Can four analytical dimensions explain the adjustments needed when those stages are applied to physical systems (RQ2)? How can change-triggered re-entry and evidence inheritance be operationalized across lifecycle stages (RQ3)? How can the framework be evaluated, and what findings would call its claims into question (RQ4)? Using generic AI lifecycle stages as an organizing baseline, this paper develops a Physical AI-specific framework linking change classification, selective lifecycle re-entry, evidence applicability, and release decisions across system configurations. Four analytical dimensions—P1 physical closed-loop coupling, P2 behavioral evolution, P3 capability–carrier decoupling, and P4 population-scale effects—inform the derivation of Physical AI-specific activities, cross-stage dependencies, change re-entry rules, and evidence management mechanisms. The paper's central contribution is to move beyond a static sequence of lifecycle stages to a dynamic approach based on change impact, evidence applicability domains, and selective re-entry. It specifies criteria, rationales, and exceptions for classifying changes; a cross-stage decision table linking change events, affected baselines, re-entry stages, evidence disposition, and release conditions; four evidence-disposition states, each defined relative to a specific capability–carrier configuration and claim rather than as an inherent property of the evidence; four distinct levels of claim, from component operability to a capability supported by sufficient evidence; the M1–M8 coverage criteria; and potential grounds for refuting the framework. The proposal is compared directly with ISO/IEC/IEEE 15288:2023, prior work on safety-case maintenance and evidence reuse, research on autonomous-system safety assurance, and existing Physical AI lifecycle studies. A 15-item expert-review questionnaire, with explicit rating anchors, is provided as an instrument for subsequent empirical evaluation.