P99 is the tail

The 99th percentile of per-position divergence — the worst one percent of places a representation damaged, which a mean happily hides. Recorded: under uniform NVFP4, granite's mean KL is 0.2778 bits but its p99 is 4.6224 — the damage lives in the tail, and the tail is where a model says something wrong with confidence. That is why the Record reports both, and why representation decisions answer to p99, not to the average.

SUPPORTED · recorded — granite-4.1-3b uniform NVFP4: mean 0.2778 · p99 4.6224