Each employee has a hidden true performance. Raters report it with leniency bias and random noise. Add raters and watch the ratings line up with the truth.
The science
Classical test theory: a rating is y = t + b + e with true score t (variance 1), rater leniency b (SD σ_b) and noise e (SD σ_e). Averaging m independent raters leaves the error variance (σ_e² + σ_b²)/m, so the correlation with the truth is r = 1/√(1 + (σ_e² + σ_b²)/m).
The same result follows from the Spearman-Brown prophecy formula: single-rater reliability ρ₁ = 1/(1 + σ_e² + σ_b²) and the reliability of the mean of m raters ρ_m = mρ₁/(1 + (m−1)ρ₁), with r = √ρ_m. The page accumulates thousands of simulated rounds and tests the measured correlation against both routes, with a z-score.
Simplifications: Gaussian errors, raters independent and drawn fresh for each employee (in real firms the same manager rates a whole team, which turns leniency into a shared offset), no halo effect, no real goal-setting or feedback conversation. It illustrates measurement, not how to manage people.