When AI systems fail in safety-critical domains, existing accountability frameworks treat those failures as unforeseeable accidents. But the mathematical structure of these systems tells a different story. Imitation-trained systems, including large language models, produce outputs through statistical pattern-matching over fixed training distributions; a design that renders certain classes of failure not merely possible but structurally derivable, entailed by the mathematical relationship between training distribution and generalization boundaries, before any system is deployed. We introduce the knowable error, a failure mode derivable from architectural properties prior to deployment, and apply it to aerospace, where existing standards such as DO-178C already formalize the relationship between software design and acceptable failure consequence, making recklessness, not negligence, the legally and ethically precise characterization of unsafe AI deployment. We argue that accountability frameworks for AI in safety-critical domains must be rebuilt around architectural foreseeability as a first-order criterion, shifting the operative standard from unforeseeable accident to knowable risk and establishing recklessness as the appropriate liability category when architecturally derivable failures are deployed without remediation.

Every operating system ever built makes the same assumption about the programs it hosts: they have no goals. The OS sees memory, CPU cycles, and file descriptors, not intentions, constraints, or objectives. For decades this was harmless. Programs were deterministic artifacts that did not have goals. Autonomous agents do. As agents become persistent infrastructure (managing resources, acting on behalf of users, coordinating with other agents), the gap between what the OS sees and what the agent intends becomes a safety and trust problem with architectural consequences. We argue that closing this gap requires a fundamental shift: the operating system must become a cognitive participant in agent reasoning. We propose that neurosymbolic architectures are uniquely positioned to enable this shift, and we define the research agenda.