This study proposes a reliability-aware framework for event-updated landslide risk assessment under extreme rainfall conditions. The prior landslide susceptibility (S0) model was developed by the machine learning method using static conditioning factors. Then, the multi-duration rainfall threshold exceedance index and rainfall-environment coupling factors were established to update the static susceptibility into event-based hazard (Se) by a Bayesian-inspired logit updated model. Bootstrap uncertainty analysis was adopted to evaluate the uncertainty propagation and reliability of updated hazard zones. The reliability-weighted landslide risk zones were finally determined by integrating exposure and vulnerability indicators, supporting post-event priority mapping and emergency risk management. A constrained large language model (LLM) prompt was designed to quickly generate the risk report based on input contributions. The results of a case study in Meizhou show that the proposed framework improved the AUC values from 0.744 (S0) to 0.967 (Se_mean) and achieved a success-rate AUC of 0.961 with all historical landslides captured in the top 20% highest-hazard area. Leave-one-out cross-validation, ablation studies, and sensitivity analysis were conducted to validate the effectiveness and robustness of the proposed model.