Trustworthy Agentic Digital Twins for Autonomous Mining: An ECMS Framework for Calibrated Self-Awareness and Dynamic Decision Authority

Abstract

Agentic digital twins can extend mining systems from prediction to closed-loop decision making, but autonomy should contract when the twin's state or uncertainty becomes unreliable. This study develops an Eco-Cognitive Mining Systems (ECMS) architecture in which calibrated operational self-awareness governs decision authority at runtime. A goal-directed grinding-classification agent is supervised by an independent assurance layer that evaluates state alignment A(t), uncertainty calibration C(t), calibration direction CDI(t), and the composite Calibrated Self-Awareness Index (C-SAI). These signals drive a four-level authority policy ranging from full autonomy to safeguarded fallback and escalation requests. The framework was evaluated using a fully matched common-random-number benchmark comprising 4,536 simulation episodes across nine operating/challenge conditions, three severity levels, seven policies, and 24 replications per cell. Across the balanced benchmark, C-SAI+CDI gating reduced mean decision regret from 0.02757 for ungated autonomy to 0.00847 (69.3%; paired mean difference = −0.01911; stratified 95% bootstrap CI: −0.01919 to −0.01903; Holm-adjusted randomization p < 0.001) and reduced the observed high-regret decision rate from 4.95% to 0%. Compared with an always-safe reference controller, it reduced regret by 25.3% while increasing mean reward from 0.13644 to 0.15252. The ablation analysis indicated that calibration provided most of the protection, while C-SAI+CDI achieved the lowest balanced-grid regret. The benefits were challenge-dependent, with the ore-hardness regime shift remaining a boundary condition. The results support calibrated self-awareness as a runtime governance signal for selective mining autonomy, while the findings remain simulation-based and should not be interpreted as evidence of industrial safety certification.