ATHOS‑COG (Advanced Therapeutic & Human‑Operational Safeguard for Cognitive Governance) is a full‑stack cognitive‑safety framework designed to prevent neuro‑affective destabilization, mitigate emotional dependency, enforce reality anchoring, and reduce liability exposure in high‑risk human–AI interaction. The framework integrates the Reality‑Drift Model (RDM) as clinical reference, a Cognitive Safety Kernel (CSK) implemented as a Python runtime module, and compliance‑oriented logging aligned with EU AI Act, GxP / Digital Health, and NIST AI RMF. The CSK monitors reality‑drift, emotional resonance, engagement depth, and addiction‑risk scores in real time, triggering intervention levels (Addiction Blocks) when thresholds are crossed, and logging all decisions for audit. This paper presents the conceptual framework and a production‑ready, type‑annotated Python implementation of the CSK, intended for deployment at API gateways, LLM orchestration layers, or external monitoring infrastructures.