ELIXIR-OS: An Integrated Closed-Loop Architecture for Uncertainty-Aware Online Identification, Mismatch Detection, and Representative-Scenario Finite Min-Max Predictive Control in Dynamic Systems

Managing autonomous dynamic systems under significant parametric uncertainty, unmodelled disturbances, andsudden structural shifts remains a central challenge in moderncontrol theory. This paper introduces ELIXIR-OS v2.1, a mathematically specified, uncertainty-aware closed-loop architecturethat unifies state estimation, effective covariance uncertaintypropagation, conservative outer-hull parameter identification,cumulative effective information-matrix tracking, a multi-criteriatemporal recovery state machine (Conditional Temporal Recovery Gate, CTRG), and representative-scenario finite min-maxpredictive control. By explicitly separating prior and posteriorestimation error covariances and synchronizing control inputsacross the plant, observer, and regression spaces, ELIXIR-OSpropagates estimation errors and process disturbances into anEffective Measurement Covariance (Reff = Rv + HP −t H⊤),which serves as an effective measurement-space covariance usedfor online parameter regression weighting rather than as anexact covariance representation of the full regression residual.When conservative outer-hull intersection tests detect a modelmismatch (Θold ∩ BoxHull(Et) = ∅), the architecture enforces astrict Conditional Temporal Recovery protocol, permitting thesystem to return to normal operation only when consistencyis regained, a cumulative post-mismatch parameter-informationcondition (rankϵ(Irecovery) = 10) combined with a sufficientinformation-strength condition (λmin(Irecovery) ≥ λready) isachieved, and bounded posterior state-error covariance remainssatisfied (tr(P+t) ≤ Pmax). Monte Carlo evaluations (N = 50)demonstrate no observed hard-state-constraint violations across2,000 evaluated plant steps, a 100% model-shift detection rate,and a 92% conditional recovery success rate, while empiricalMahalanobis error coverage remains close to nominal 95%levels. Wilson score intervals are reported as descriptive binomialuncertainty measures.