Cohen’s κ handles TWO reviewers. But a program committee has many — and Fleiss’ κ measures how much a whole panel agrees, corrected for the agreement you’d get if they all just guessed at the same base rate. Slide the true agreement and watch raw agreement and κ pull apart.
Fleiss’ κ = (P̄ − P̄ₑ)/(1 − P̄ₑ) generalises Cohen to any number of raters per subject: P̄ is the mean per-subject agreement across all rater pairs, P̄ₑ the agreement expected from the marginal category rates. κ=1 perfect, 0 chance, <0 worse. Here the rater table interpolates from a pure-chance split to unanimous as you slide, so κ climbs 0→1 while raw agreement stays deceptively high at the chance end. A fail-loud self-check throws unless κ≈0 at chance and κ≈1 at unanimity. ◆ real statistics, node-verified.
The interpolated rater table is the illustrative choice; the Fleiss ARITHMETIC on whatever table results is exact. Landis–Koch band names are a convention. THE PEER reads the panel, not one reviewer.