<p>The improved tangential angle method is widely used in rock slope landslide early-warning level assessment. However, its capability is often compromised by the difficulty in real-time estimation of the secondary creep velocity. To address this issue, three distributional indicators (deviation coefficient, skewness, and kurtosis) are extracted from historical landslide cases to characterize rockmass creep evolution, and are fused using an adaptive weighted-average model to detect the onset of the acceleration creep stage. Based on the detected onset, the secondary creep velocity is updated automatically via an inverse-search algorithm, enabling continuous calculation of improved tangential angles and the corresponding early-warning level. The path-dependent effect is incorporated by introducing reduction and amplification coefficients to quantify how historical early-warning states influence the real-time assessment of early-warning level, thereby correcting the early-warning level. Validation on six held-out cases shows that the proposed model effectively captures the transition of rockmass deformation from secondary creep to acceleration creep, and can better identify the onset of the acceleration creep stage. Building on this detection, the secondary creep velocity can be estimated automatically, with high agreement with the reference secondary creep velocity. After the path-dependent effect is incorporated, the estimated early-warning levels become more stable and generally more conservative. The abnormal oscillations that are often observed in conventional approaches and may mislead decision-makers are effectively suppressed. Overall, this study provides a robust and practical method for real-time early-warning level assessment and proactive management of rock-slope landslides.</p>