Notes from the field.

Confidence is the share of a tracked asset's pools whose normalized price falls inside a tolerance band around the reference — the same ±10bp-style convention used for liquidity depth. It's a summary statistic, not a probability of anything happening next. High confidence means most pools are reading the same state right now, quotes have converged; low confidence usually shows up during a fast-moving read, before the slower pools have caught up to whichever venue moved first.

Reading changes in confidence

Confidence is most informative watched over a sequence rather than at a single instant, which is why Replay tracks it alongside dispersion. A representative run: confidence opens calm around 96% several blocks before a divergence event, drops to 78% at the exact block where AMD's dispersion peaks at 84bp, then recovers to 92% within a handful of blocks as pools re-converge. The dispersion spike and the confidence dip are the same underlying event described two different ways — one says how wide the disagreement got, the other says how many pools were part of it.

Why it's deliberately coarse

Confidence is meant to be glanced at, not optimized. A wide dispersion with high confidence and a wide dispersion with low confidence are meaningfully different situations — one venue quoting far from an otherwise-agreeing market, versus the whole market genuinely split — and the score exists specifically to tell those two apart quickly, without requiring a read of every individual pool.