How do slices prevent aggregate AI metrics from hiding regressions?

Slice definitions are versioned and tied to product risk, because arbitrary post-hoc slicing can create noisy findings. Release gates protect critical slices even when their traffic share is small, while exploratory slices generate hypotheses rather than automatic failures. I correlate offline slice regressions with production traces, model versions, prompt routes, and data changes to locate the cause. When traffic distribution shifts, I update weights without discarding the stable slice history needed for comparison.