Earthquakes are among the least predictable natural disasters, mainly due to our insufficient understanding of their complex mechanical processes. The ground shaking that is generally recognized as an earthquake is caused by a sudden slip on a fault between two blocks of the Earth's crust, which is termed an earthquake rupture. The mechanical processes of an earthquake rupture involves initiation, propagation, and termination. Each of these three phases involves physics at various length scales, in which the termination is the integrated product that finalizes the eventual magnitude of an earthquake. While analytical analyses, laboratory experiments, and numerical simulations have made great contributions in deepening our understanding of earthquake rupture initiation and propagation, our knowledge of how natural earthquakes arrest is almost exclusively based on remote measurements. The fault length required for an earthquake rupture to arrest spontaneously is significantly larger than most laboratory setups and, therefore, rarely investigated. Here we present a large-scale (3 m) experiment that generates realistic laboratory earthquake ruptures, where the ruptures spontaneously nucleate, propagate, and arrest before reaching either ends of the simulated fault. On-fault slip and close-fault strains are measured along the simulated fault at high spatio-temporal resolution. The slip profile of these laboratory earthquakes are found to be similar to natural earthquakes, which implies comparable mechanical processes. We showed that linear elastic fracture mechanics is capable of predicting the extent of an imminent rupture event if it were to happen under the present conditions. An analytical crack model is proposed to accurately represent the spatial distribution of the slip and stress change measurements. The deviation between fracture energies estimated through crack models and through the dynamic rupture propagation uncovers the misrepresentation of stress changes to absolute stress in previous studies. Further, the seismologically estimated fracture energy of natural earthquakes are observed to be scale-dependent, while experimental studies highlighted it to be a scale-invariant material/interfacial constant. Informed by our experimental observations, we developed a suite of scaling numerical models that extend the length scale beyond our experiments. These numerical models share identical material and interfacial properties, which results in identical fracture energy. However, the fracture energy derived from extracted seismological parameters still shows the same scale-dependency observed in natural earthquakes. We found that the seismologically observed scaling in earthquake fracture energy is likely dominated by the stress overshoot, which is embedded in the formula and invalidly assumed to be zero. Through the experimental, analytical, and numerical works in this study, the proposed earthquake models and scaling scheme will aid in our understanding of fault mechanics and the forthcoming developments in physics-based earthquake prediction frameworks.