Insurability of Correlated Ontological Failure: A Systemic-Risk Loss Model for Clinical-AI Pipelines

When ontological distortion in clinical AI can be induced deliberately, a liability question follows: is the resulting risk insurable? This paper is a model-and-protocol contribution, not a completed empirical study. We construct a per-pipeline expected-loss model parameterised by two inputs: a loss-severity distribution and a cross-site correlation structure of the exploitable coding channels. Using a published adversarial cascade simulation (Synthea-derived cohort, n=1,000 patients, 100 Monte Carlo iterations), we show that loss severity is regime-dependent: coordinated multi-node injection produces 1.55 times the error of single-node injection under identical magnitude and feedback, and a persistent low-magnitude stealth-ramp reaches total error 2.94 against an accidental-drift baseline of 0.31. We then argue, and pre-register as the paper's empirical arm, that the exploitable channels share a common structural driver (the coding and billing process), so that failures are correlated across insured units rather than independent. The correlation is a directional prediction, not a measured result. Conditional on it holding, correlated tail loss violates the independence assumption underlying standard actuarial pooling: the per-unit portfolio standard deviation no longer diversifies as 1/sqrt(n) but floors at sqrt(rho), so ontological-attack risk is best modelled as a systemic accumulation risk rather than an idiosyncratic per-site risk. This is the substantive content of the insurance gap for clinical AI: the loss is not uninsurable because it is small or rare, but because it is correlated.