Worked on stabilizing time-based checkpointing in the ecmwf/anemoi-core repository to enhance training reliability for long-running machine learning experiments. Addressed a bug where checkpointing based on elapsed time failed to trigger by ensuring the timer was correctly initialized in the on_train_start method, resolving an issue caused by an overridden method that omitted a necessary call to its superclass. This fix, implemented using Python and leveraging skills in software debugging and unit testing, restored reliable checkpointing without affecting epoch- or step-based mechanisms. The change improved recoverability and reproducibility for multi-GPU training runs, directly aligning with tracked issues and pull requests.