Persisting in the World: A Longitudinal Evidence Architecture for Understanding, Development, and Continuity in Artificial Systems

As artificial systems acquire long-term memory, persistent operation, tool use, autonomous goal formation, and component-migration capabilities, several different kinds of evidence are increasingly treated as if they supported the same conclusion. Retrieving the past does not show that the past has changed present meaning. Correctly understanding a target now does not show that a particular experience has changed the system's future. Persistent behavioral change does not show that the system understands the relations implicated in that change. Nor can preservation of a name or task performance after a model, memory, or harness replacement establish that previously formed cognitive and relational organization continues to operate. Enactive cognition, situated action, BDI, memory lifecycle, persistent agents, causal attribution, provenance governance, and artificial-system identity already cover many components of these problems. What remains to be established is not another capability list, but a common longitudinal research object and the bridging burdens required to connect distinct evidentiary levels. This paper proposes an evidence architecture organized around a relationally structured longitudinal object. The object is the process by which a persistent artificial system forms its current orientation from a relationally structured historical situation; uses that orientation to organize salience, meaning, possibilities, action, or reasoned non-action; attributes the consequences of action; reorganizes relational and developmental history; and enters the next moment under changed orienting conditions. To prevent evidence from one layer from standing in for this entire process, the framework assigns separate evidentiary burdens to current understanding U, experience-specific Developmental Uptake DU, and cognitive-relational continuity C across component change. Once these judgments have been independently supported, pre-action history/standing coupling and consequence-attribution-bounded update are audited as interfaces that may be implemented by existing mechanisms. A diachronic reconnection contract, with its own intervention burden, then tests whether experience-induced organization supports later understanding of the same target relation. Ordinary sufficient states, structural causal models, and existing agent architectures can in principle represent these variables and relations. The central increment is therefore not an irreducibility claim, but an account of what evidence licenses each longitudinal conclusion and which bridging burden must be met before one conclusion can support another. A persistent agent maintaining a shared experimental facility serves as a running case. Three cross-comparison protocols are developed: a functional-order comparison between upstream organization and downstream filtering; a longitudinal-update comparison between different paths and attributions leading to the same terminal state; and a component-change comparison separating content access, task performance, surface identity, lineage, and cognitive-relational continuity. Rather than presupposing an imaginary, costlessly compatible "strongest system," the protocols use a composite-baseline registration rule, R_B^1. Each study must freeze a finite registered baseline set B_r before observing results and declare hosts, modules, interfaces, information access, resources, process observations, and intervention budgets. No empirical results are reported. Joint support for the three evidentiary endpoints is not treated as sufficient for consciousness, personhood, moral rights, or AGI. The contribution is a longitudinal architecture of evidentiary permission and bridge evaluation: at one preregistered granularity, it jointly adjudicates U / DU / C, interface and reconnection contracts, registered-baseline absorption, edge deletion, and claim downgrade, so that the full generative question can fail in parts rather than being replaced by evidence from a single layer. This is an integrative conceptual research-programme paper and evaluation agenda. It reports no empirical results and has not been peer reviewed.