Separate real rater disagreement from the kappa prevalence artifact via Pe

When one stratum of an inter-rater kappa table looks far worse than the others, emit expected agreement Pe next to observed agreement Pbar before concluding the raters disagree. If Pe is roughly constant across strata and only Pbar drops, the low kappa is genuine disagreement; if Pe is near 1 because the labels are skewed, the low kappa is the prevalence artifact and the raters may actually agree ~97% of the time. Being able to show which case you are in is the difference between a defensible finding and a misread table. Related trap on the same data: a majority-vote panel can score WORSE than its single best member inside a stratum where one rater falls below the constant-predictor baseline, so always compute panel accuracy against best-single-rater accuracy per stratum, not just in aggregate.