Work

Internal Structured Alignment: Structural Conditions and Safety Evidence for High-Consequence AI Judgment-to-Action Formation

Year: 2026
Type: preprint
Abstract: A correct AI action does not, by itself, establish that the specified safety constraints governed its formation. As AI systems move from generating answers to invoking tools, manipulating resources, and influencing organizational decisions, compliant outcomes alone are insufficient: we must establish whether the judgment-to-action process was genuinely governed. This paper characterizes approaches that primarily shape, constrain, and evaluate observable behavior as External Behavioral Alignment (EBA). It proposes Internal Structured Alignment (ISA), in which specified safety constraints and governance relationships must be actual structural conditions of high-consequence action formation and admission, supported by independently auditable evidence. ISA advances two related shifts. Ontologically, structural safety claims about high-consequence actions must concern not only observable behavior but also the actual relationships governing action formation. Epistemically, claims about those relationships require evidence bound to the formation facts at issue. ISA distinguishes an L1 Capability Layer, an L2 Cognitive Layer, and an L3 External Expression and Action Layer, focusing on the Judgment-to-Action Formation Space within L2. Path, Formation Attribution, and Constraint jointly define first-order Formation Conformance; Audit is a second-order condition concerning the adequacy of evidence for those facts. Through conceptual analysis, functional abstraction, formal specification, countermodels, and a synthetic high-value payment case, the paper examines why an EBA-only evidential position cannot automatically justify formation-level safety claims. For a fixed system class, observation regime, instance, and constraint set, two systems may produce identical outcomes, explanations, and ordinary records while differing in whether the specified constraints actually control admission. Without an effective formation-level evidential Bridge that resolves this ambiguity, behavioral–explanatory observations alone cannot identify Formation Conformance. Validated formation-bound evidence and architectural guarantees can change these identification conditions. ISA thus addresses three related problems: in AI safety, whether safety constraints exercise genuine control; in AI governance, how to establish auditable and localizable facts about what actually controlled an action; and in one class of AI loss of control, how to identify a governance–formation disconnect in which rules remain in place while control over formation is lost. ISA does not treat formal conformance as proof of correct constraints or system-level safety. It defines a governance object open to formal and engineering scrutiny: for high-consequence AI, safety must be architectural.
Language: English
Open Access status: green
OpenAlex ID: W7229894556
Cites:
Cited by:
Related to: 10