The Last Meter of AI: From Intelligent Answers to Accountable Decisions

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

This manuscript introduces the “last-meter problem” of artificial intelligence: the gap between an AI-generated output and a consequential real-world action that is actually permitted to execute. It distinguishes prediction, recommendation, decision, and execution as separate system events and argues that trustworthy model performance alone is not sufficient for high-consequence AI. A specific action may still be unjustified if evidence is stale, authority is invalid, policy conditions are unmet, system health is degraded, scope is exceeded, or relevant permissions have been revoked. The paper defines decision-time assurance as a distinct systems layer that evaluates evidence, authority, context, policy, system health, scope, temporal validity, and revocation immediately before execution. It introduces a decision-gate abstraction and permission states including ALLOW, REVIEW, ABSTAIN, and BLOCK. The manuscript further presents formal invariants, a taxonomy of post-inference and pre-execution failure modes, decision-record requirements, threat models, evaluation metrics, and domain examples covering credit decisioning, healthcare, autonomous financial agents, and agentic software operations. AICOS CORE is presented as an experimental reference implementation direction for studying decision-time assurance. The manuscript explicitly separates design, implementation, validation, independent replication, and production evidence, and does not claim that the architecture has already been independently validated or proven in production.