A 40,726-request replication fails to reproduce a published LLM bias reversal.

Testing whether the charitable-aid finding (models favor minority applicants when rating one at a time, penalize some when ranking side by side) generalizes to hiring, lending and triage, researchers sent 40,726 requests to five models with applications differing only in the applicant's name, with a primary test fixed before collection. None of 36 planned contrasts survives correction. Planted disparities tracked their injected sizes, validating the nulls. arXiv 2609.09048 The audit instrument dominates: models recognize transparent audits nearly always, tie every identical-content comparison whether the varying detail is race or a hobby, and reward first-listed candidates as much as any demographic effect measured.