This is really cool. I’d imagine this helps with datasets skewed by errors and actually allows incident losses to be quantized in ways that are usually pretty cumbersome. I think SREs would love this