Auditing policy dependent benchmarks in allocation simulations

Allocation policies change which agents remain available for later decisions, so a benchmark built from the best agent still available depends on the policy being evaluated. This note works through an audit of that measurement choice. In a two-period housing example, the moving-reference ratio rises while realised value falls from 16 to 11, and a matched-seed housing sweep reproduces the disagreement. In a separate deadline queue, value rises from 19 to 27 while the moving ratio falls from 1.00 to 0.90; 15 of 200 fixed-seed cases disagree in this direction. A reusable routine reports the changes in numerator and denominator next to a common reference. Feasibility checks tell an exact offline optimum apart from an upper bound, and a completed-cohort check shows which population an access gain refers to. Both simulations are illustrative, and the note does not measure how common the problem is in published models. Files: the manuscript (PDF and DOCX) and a package with synthetic inputs, code and saved outputs that runs offline. The model comes from The Execution-Certainty Wedge (10.5281/zenodo.22164073). See revision_notes.md for details.