A model predicts at 09:00. A feature event occurred at 08:58 but arrived at 09:03. A backfill later rewrites its timestamp. How should the historical training row be built?

Use the information available at the actual decision cutoff, including arrival time and source version. Exclude the late event for the 09:00 prediction unless the production decision is deliberately delayed under the same contract. Preserve original availability evidence so backfills cannot silently create historical knowledge.