IAS-AI is a machine-readable, auditable self-constraint protocol derived from Zero Leap Theory (ZLT) for AI outputs that function as interventions in human and institutional systems with memory (hysteresis), low observability, and non-linear fragility. It formalizes five necessary admissibility gates—Consent/Permeability (C), Stability (η), Alignment/Legitimacy (Φ), Memory/Hysteresis (H), and Observability (O)—and states a Non-Compensability Theorem: if any gate is closed or unknown, increasing intervention intensity cannot reliably produce the intended outcome and instead increases dissipation and tail risk. The paper contributes a dual-layer format (human-readable argument + machine spec), mandatory audit logging and a Gate Evidence Schema to prevent gate-washing, an adversarial threat model, canonical minimal cases with ordinal predictions, and a standardized citable refusal template.
DANNY YUBI DAGOGLIANO (Thu,) studied this question.