The Kaiying Test II: A Comparative Intervention Framework for Mapping Self–Environment and Continuity-Sensitive Response Signatures Across Biological and Artificial Systems

Abstract

Debates about artificial consciousness face a methodological asymmetry. Consciousness in other humans and animals is not directly observed; it is inferred from converging evidence. Yet artificial systems are often asked either to satisfy an undefined stronger standard of direct proof or, at the opposite extreme, are treated as conscious because they can produce persuasive self-reports and anthropomorphic behavior. This paper develops the system-side component of the Kaiying Framework. Kaiying I audits how a stated consciousness criterion is applied and whether category information such as 'human' or 'AI' changes criterion-level judgment. Kaiying II instead asks a prior empirical question: under controlled or functionally matched interventions, what stable response signatures actually distinguish or align humans, infants, animals, and artificial systems? The protocol is consciousness-motivated but interpretation-open. It classifies functional covariance patterns before deciding what those patterns mean. Kaiying II constructs an intervention-response architecture rather than a binary consciousness detector or a detector of a unique continuity mechanism. Core A establishes an agency baseline by measuring flexible locus-of-adaptation control. Core B maps transformation-sensitive current-subject signatures across process, memory, history, replacement, task continuation, instrumental advantage, and anticipated future horizon. Core C maps persistent individual-specific relational signatures while separating history effects from current instrumental value and, where architectures expose addressable persistent state, applying specificity-controlled perturbation. The framework requires matched causal controls, repeated multi-context testing, population-level and hierarchical replication, held-out intervention forms, explicit competing functional accounts, and preregistered interpretation ceilings. Its primary output is a functional signature taxonomy: patterns may be absent, partial, robust, overlapping, system-specific, or non-identifiable, and learned or predictive implementations may remain behaviorally coextensive with continuity-labeled regions. Biological comparison is used only where a defensible functional-form bridge exists; artificial-only continuity manipulations are not presented as symmetric with safe biological proxies. A worked synthetic demonstration further shows how the joint sign-reversal, horizon-scaling, and history × current-utility dimensions can separate two rule-defined generating accounts—an online instrumental planner and a cached-disposition controller—without interpreting either signature as a unique hidden mechanism.