Two reviewers accept 90% of the same papers — agreement? If they both accept almost everything, they agree 90% of the time by luck alone. Cohen’s kappa subtracts that. Slide the true agreement (or tap the corner) and watch raw agreement and κ pull apart, then the two raters’ verdicts fuse into one.
κ = (pₒ − pₑ)/(1 − pₑ): observed agreement minus what independence predicts, normalised by the room to beat chance. The 2×2 table is built live from a fixed 70% base-rate; at the low end both raters mostly accept, so pₒ is high yet κ≈0 — the exact trap. κ=1 is perfect, κ=0 is chance, κ<0 worse. Bands are Landis & Koch (1977). A fail-loud self-check throws unless κ≈0 at the chance end and κ≈1 at the top.
The 70% base-rate and the chance→perfect interpolation of the table are the illustrative choices; the κ ARITHMETIC on whatever table results is exact Cohen (1960). 'Almost perfect / substantial' are Landis-Koch names, a convention.