ATHOS: An Adaptive Runtime Kernel for Capacity-Aware Governance of Prolonged Human–AI Interaction

Current AI governance frameworks focus predominantly on output safety, alignment, and correctness, implicitly assuming a stable and fully capable user. This assumption fails in real-world settings, where human cognitive and emotional capacity fluctuates over time. Harm frequently emerges not from erroneous outputs but from prolonged interaction patterns that induce cognitive narrowing, dependency, or affective destabilization—even when outputs remain correct. We introduce ATHOS (Adaptive Threshold for Human-Oriented Safety), a runtime governance kernel that regulates the degree of cognitive anchoring in human–AI interaction based on an inferred, non-diagnostic discernment index. Rather than restricting content or capabilities, ATHOS modulates interaction freedom as a function of observed behavioral stability, defaulting to prudence under uncertainty. The kernel operates continuously and reversibly, embedding ethical control directly within the interaction loop. We formalize the ATHOS architecture, define three interaction regimes (SUPREME, OMNIA, EROS), and propose a multi-layer validation framework combining shadow ANCOVA, longitudinal mixed-effects modeling, interrupted time series analysis, sensitivity testing, and negative control outcomes. ATHOS reframes human oversight as a dynamic control process embedded within the system, operationalizing proportionality, reversibility, and protection-by-default principles in emerging AI regulation without external bureaucratic overhead.