arXiv:2611.25082v1 Announce Type: cross Abstract: Digital twins of cyber-physical systems must operate with scarce data, on assets that change after deployment, and for humans whose needs differ, yet existing implementations fix fidelity at design time and do not match the representation to the human or the task. This paper proposes a human-centric adaptive multi-fidelity digital twin framework that models the evolving human, the physical twin, and the digital twin core, and adds a human inference and adaptation function that estimates the human state, predicts the response, and adapts the representation. The framework defines low-, medium-, and high-fidelity tiers distinguished by adaptivity and by coupling, quantified as the ratio of the digital twin latency to the duration of the process step it follows. Three proof-of-concept instantiations are evaluated: a virtual reality tutor, a computer numerical control milling model, and an industrial training station. The low tier estimated learner sentiment with an accuracy of 0.86. The forecasters of the milling model and the training station, both medium-tier, showed a normalized error decline of 0.39 to 0.89 and a human adaptation gain of 1.16 to 1.51, but their latency exceeded the physical process step by 25 to 87 times. Only the station's vision model met the physical step. The tiers are separated by which functions adapt and where the latency constraint lies, not by a monotone increase in gain or coupling.