Causal qualification before synthesis: a research agenda for synthetic neuroarchitecture evolution

Connectome-constrained models offer a way to test how neural organization contributes to computation, but successful simulation does not by itself identify a mechanism or justify a synthetic design. I propose Synthetic Neuroarchitecture Evolution (SNE) as a research agenda linking three questions: what an organizational relationship contributes under controlled conditions, which details can vary while preserving that contribution, and whether a compact rule generates useful new architectures. The agenda builds on existing NeuroAI, causal modelling and generative encoding methods. Its ordering concerns the evidence required to promote a design claim; exploratory synthesis remains possible throughout. I distinguish mechanism tests from engineering comparisons and show why increasingly strict matching can preserve the feature under test, why excluding silent counterfactual networks can change the question, and why a fixed readout tests preservation rather than all recoverable function. A proposed sensorimotor experiment illustrates how targeted interventions, independent confirmation and explicit failure criteria could connect model-based causal evidence to transferable organizational rules. This Perspective reports no original experimental results and makes no claim that the proposed programme has established biological causality, engineering advantage or general intelligence. Author preprint v1, prepared 23 October 2026. AI assistance is disclosed in the manuscript.