The Missing Layer: Specification Infrastructure for AI Oversight

AI oversight currently lacks standardized infrastructure: Organizations deploying AI systems independently build their own "audit schema, policy dialect, monitoring stack, and escalation path," representing wasteful reinvention of patterns that could be shared across the industry—this is a coordination problem rather than a research gap.

Specification is the critical missing layer: The paper identifies "Layer 2 (Specification)," where humans translate intent into machine-checkable rules, as "the connective tissue every layer depends on," yet it lacks the four hallmarks of mature engineering disciplines: shared vocabulary, design principles, composability standards, and governance practices.

A practical framework exists but remains fragmented: Existing systems like Cedar, Constitutional AI, and Open Policy Agent each address parts of the specification problem well, but none comprehensively cover the full matrix of concerns spanning alignment, robustness, adversarial defense, security, governance, and accountability.

Six design principles can guide better specifications: The paper proposes concrete principles for specification infrastructure including elicitability (humans can express their intent), composability (specifications work together), adversary-awareness (anticipating misuse), traceability (auditable decisions), and governability (enabling oversight)—moving from ad-hoc approaches to engineered systems.

Standardized specifications enable better oversight at runtime: A reference architecture connecting specifications to runtime enforcement, evaluation, and escalation could allow organizations to move beyond custom solutions toward deployable, auditable oversight systems.