MarginTrace: A Protocol for Validating Near-Tie Model-Selection Reversals

Sub-point public leaderboard margins often determine which model is selected for downstream evaluation, deployment shortlists, and further investment. MarginTrace is a decision-certificate specification for establishing when that selection remains valid under task-equivalent benchmark variation. A conforming study fixes the eligible model-pair family from public scores, records a registered hidden-substitution procedure, calibrates item difficulty with a disjoint pilot panel, audits every final scored pair, and evaluates pairwise reversals with family-wise inference and independently authored recurrence. The specification separates a per-model score shift from the operational question: whether the public winner remains the winner. It also makes the evidentiary boundary explicit: a certificate is supported by linked eligibility, calibration, audit, inference, and recurrence records for the same decision. This framing turns a public near tie into a checkable scientific object and supplies a rigorous protocol for studying benchmark-mediated model selection on MATH, MMLU, and other tasks with auditable task-equivalent variants.