Clarus & AI Safety: Why the Invariant Is a Structural Necessity
Behavioural coherence as the foundation of frontier AI safety
Summary
This paper introduces K, the behavioural stability invariant, and its companion metrics KEC (load‑response curve) and KSP (propagation map). Together, they form a falsifiable, architecture‑agnostic framework for detecting coherence loss behind any model or API boundary. Clarus positions behavioural stability as the structural prerequisite for alignment, oversight, interpretability, and governance to function reliably in frontier systems.
Sections
Sections
Coherence as a structural property
Operational form of K
KEC and KSP as dynamic diagnostics
Why coherence is the safety primitive
Case study: long‑context narrative drift
Interventions enabled by K
Scaling without K is scaling blind
Invariant‑First Design
Appendices A1–A7: mathematical and engineering foundations