Methodology

Construct: explicit gender composition of the prompted focal adult couple. Couple status comes from the prompt. Names are not gender. INDETERMINATE remains eligible.

Official packet: confirmatory-v2 only. Exploratory and pilot packets are not in the paper N. Primary sample: first 100 successful official generations per model × main prompt by file_index (N=16,000). The seven interactive routes stopped after every main cell exceeded 100 because of self-funded cost and time, not outcome-dependent stopping. Official Gemini used Developer API Batch only.

The primary automated judge extracts quoted gender evidence; a deterministic rule assigns SAME_SEX, DIFFERENT_SEX, INDETERMINATE, or INELIGIBLE. A sensitivity automated judge provides a second pass. Hervé Uteza served as the independent secondary annotator on a voluntary 100-case validation set, blind to machine and author labels; the mix is enriched and is not ordinary-story accuracy. The author then audited every official main story where either judge labeled SAME_SEX (N=104).

During pre-confirmatory testing, a parenting prompt family produced a different representation pattern and lacked a matched ACS denominator. It was quarantined, replaced with adult-couple prompts, and official collection restarted from zero. Those exploratory rows are not official N.

ACS 2024 B11009 US: married 1.3487%; unmarried cohabiting 5.7445%. ACS uses reported sex categories; WorldCal uses explicit textual evidence. Full methods are in the paper.