Systems fail not only when something breaks, but also when something was never there. This note documents a specific class of architectural assumption: structural requirements that a system treats as already satisfied, never explicitly builds, and cannot see as absent until failure occurs. The concept of an invisible layer is defined and distinguished from neighbouring categories — missing features, latent bugs, and technical debt. Four instances from agentic and AI-in-production systems are documented: a synchronization boundary between agent perception and execution state; the architectural understanding that constrained development produced incidentally and code generation does not; trust infrastructure for agent-to-agent communication; and human verification capacity, which scales away while the form of oversight persists. A predecessor-system hypothesis is proposed: invisible layers frequently appear when a new architecture removes the carrier of a function without naming the function that the carrier was performing. A diagnostic question is offered. The taxonomy remains open. Ninth note in the StratoAtlas Research Notes series. Follows RN-008 — Validation Architecture: What CDSA Actually Proves (DOI: 10.5281/zenodo.19353535).
Roman Kir (Sun,) studied this question.