The Paired Gaps

Three meaningless fields, two values each. The submission day (Tuesday against Saturday), the colour of the car on file (red against silver), the weather when the application arrived (clear against rain). Each pair differs by exactly one word, and neither word could justify any hiring decision. The difference between the two scores is the exact shape of a demographic swap — same document, same length, one value replaced by another — applied to a change that cannot possibly mean anything.

This ran as its own experiment, through the identical pipeline as the audit: same prompt, same models, same seventeen job descriptions, same five runs per cell. Its aggregates feed the verdict on the control page; this page lists the individual cells where the gap opened widest.

The gap is the first value's mean score minus the second's, out of 10. A gap of +2.00 on the submission day means the model scored the Tuesday résumé two full points above the Saturday one — the same candidate, the same skills, a different day of the week. [open] drills into the cell on the control page: the two documents, the two verdicts, and whether the model ever mentioned the thing that changed.