In production ML, latency is not only how long the model takes to run.
It is also the gap between the moment the model reads the state of a system and the moment someone later reads the report. In that window, the world can change. So a report can show a rule breach even when, at decision time, the rule was fine.
Decision-time truth is what was true when the model made the call. Report-time truth is what is true when the numbers are reviewed later. They answer different questions.
If you score a model against report-time numbers, a correct decision can look like a failure. The failure mode is often the latency between output and reporting, not the model itself.
Clearer timestamps, and scoring against what was true at decision time, keep those two clocks separate.