Disruption remains a major barrier to the safe, sustained operation of tokamaks. Although current understanding of disruption physics is broadly valid across existing tokamaks, it cannot provide accurate disruption prediction rules. Data-driven methods suffer from performance degradation as future reactor-scale tokamaks cannot accumulate sufficient disruption data before risking damage. We introduce a disruption prediction framework that balances physics priors with data-driven learning. It allows known disruption-related operational limits being explicitly imposed as constraints on model training, and can also embed known physics implicitly through physics-guided representation learning. Physics can be blended with data with failure-driven active learning as well. The effect of existing tokamak data is maximized via domain adaptation. Our results reveal physics-based guidance is effective, but explicit guidance loses influence as data accumulates while implicitly guidance does not. This work outlines a progressive strategy to integrate physics and data effectively.