A typical recommendation is to not do any transformations as the data lands into the bronze layer (ELT). The idea is that you want your bronze layer to be as close of a representation of your source data as possible so if there are any mistakes later, or taking your example, the existence of duplicates is a helpful indicator of some kind, it's nice to have an accurate system of record.

So in your example, raw data lands in bronze as it is and is deduplicated in the silver layer. These are not hard and fast rules - they are up to your practice.