For ratings on a NUMBER scale (not just accept/reject), the ICC asks: how much of the total variation is real difference BETWEEN the things rated, versus disagreement WITHIN each thing’s raters? All signal → ICC 1; all noise → 0. Slide the rater noise.
The one-way ICC = (MSᵇ − MSᶠ)/(MSᵇ + (k−1)MSᶠ): between-subject mean-square minus within-subject, over their weighted sum. It is the share of total variance due to real between-subject differences — 1 when raters agree exactly, 0 when rater noise swamps the signal. The standard reliability index for continuous ratings. A fail-loud self-check throws unless identical raters give ICC=1. ◆ real statistics, node-verified.
One-way ICC(1) on a small synthetic panel is the illustration; the MS ratio is the exact formula. Which ICC form (1,k / 2,1 / 3,1) to use depends on the design — a real choice a peer must state.