Free Pairs: Label-Free Auditing of a Borrowed Face-Match Threshold at an Open Border

A face-recognition system at a border decides "same person" with one distance threshold, usually borrowed from a benchmark and never measured on-site. We show a busy crossing supplies hundreds of free identity pairs with no labels: one unbroken track is one person, two tracks on screen at once are two people. We use these must-link and cannot-link constraints to audit a borrowed threshold in deployment and to refuse a number when the two groups overlap. On real footage from an India to Nepal post the audit did not return a better threshold, it returned a defect: 42% of same-track pairs sat at stranger distance, meaning the tracker was passing one id between people and the pipeline was averaging two faces into one identity. We fix the aggregation and state a refusal rule. We report no detection accuracy, because none was measured, and say what would have to happen for a real number to exist.