Naming a donor that is not there. Negative controls for automated loanword attribution

Code and run outputs for a study of donor-absent negative controls in automated loanword attribution. Borrowing detection is evaluated with the true donor present, so abstention is never tested. Removing the donor from the candidate set produces false attribution at rates far above the nominal level, worst on uncorrupted data. Includes the full benchmark harness, all run outputs, a run manifest linking every figure in the paper to the script that produced it, and a record of corrections.