A Graph-Theoretic Alternative to Forced Ranking

Executive Summary

Forced ranking asks each manager to rank a small, non-representative team and then extrapolates those local orderings into global promotion and termination decisions. Small samples plus team-quality variance produce large error rates even with perfect local observation: in the companion simulations, 31–54% when identifying the top and bottom 15% of performers.

This paper (originally titled "The Mirror Supplement") introduces a simple, sparse cross-team comparison method that preserves local judgment while connecting otherwise disjoint local orderings. It has since been absorbed into the complete forced ranking paper as its constructive section.

The Method

• Guest comparisons: add one or two role- and level-matched cross-team "guest" comparisons to each team's local ranking per cycle
• Pairwise inference: fit a standard pairwise model (Bradley-Terry-Luce, or an Elo/TrueSkill-style update) to the resulting sparse graph to infer a global partial order
• Quota adjustment: adjust team-level quotas by inferred team strength, then rescale to recover the global 15% promotion and termination targets

Results

The paper evaluates the method in simulated organizations of 994 employees in 142 teams of 7, with ability clustering and rater bias. The same mechanism that makes forced ranking worse as managers build strong teams (local compression) makes the sparse method better, because transitivity across guest comparisons reveals team strength.

Governance

The paper frames the design as Mode B governance in the Cage and Mirror sense: documented judgment with bounded variance rather than legibility alone. Because every comparison and adjustment is recorded, the process remains legally auditable and organizationally feasible.

Key References