SARA Scale: A Governance-First Taxonomy for Artificial Intelligence in Critical Systems

Artificial intelligence systems are increasingly deployed in critical, regulated domains where decisions have significant consequences. Existing taxonomies prioritize capability and autonomy over governance, leaving architects and regulators without clear criteria for high-stakes contexts. We present the SARA Scale (Structured Autonomous Reasoning Architecture), a taxonomy that classifies AI systems by governance architecture rather than capability. The SARA Scale is grounded in one principle: operational intelligence in critical systems is defined by what a system is explicitly forbidden to do, not by what it can do. We formalize this through six architectural invariants and define ten technical levels (SARA-0–SARA-10) plus a legal frontier (SARA-11). The scale distinguishes a Pre-Governed Zone (SARA-0–SARA-4: stochastic and deterministic systems without architectural governance) from a Governed Zone (SARA-5–SARA-10: explicit world models, deterministic evaluation, architectural constraints). An orthogonal Execution Regime axis clarifies hard limits: stochastic systems cannot exceed SARA-4 regardless of capability improvements. The scale is taxonomic, not teleological—higher levels are not inherently “better,” and SARA-11 (legal recognition) is achieved through regulation, not engineering. The SARA Scale was not designed top-down as a theoretical framework. It was derived bottom-up from an operational system in the energy domain, demonstrating that fully governed autonomous intelligence is empirically achievable. We provide an operational classification procedure with verification tests and evidence requirements, enabling auditable assessment. By reframing progress in critical AI from more autonomy to more governance, the SARA Scale offers a foundation for designing and auditing AI systems where accountability is non-negotiable.