Supplementary Materials: Volume and Distribution as Separable Dimensions of Gender Bias in Large Language Model Brand Recommendation

Replication data, brand-extraction instrument, analysis code and human validation for a study of how large language models vary brand recommendations when the only change to a prompt is the gender of the gift recipient. 574 responses from three commercial models (Google Gemini 3 Flash, OpenAI GPT-5.2, xAI Grok-4-1) across two seasonal contexts: 424 Christmas responses over four prompt conditions collected 23 and 29 December 2025, and 150 Valentine's responses over five prompt conditions collected February 2026. This package documents three defects in the brand-extraction procedure used for the submitted manuscript, each of which had produced a published number. First, the brand Away was matched case-insensitively and so counted inside the common noun getaway; it was reported as the most frequent brand in four Valentine's conditions at 15 to 21 mentions out of 30, where the corpus contains one standalone mention in 574 responses. Second, a replacement instrument recorded failed extractor calls as responses naming no brands, and the failure rate rose with the number of brands a response contained, from 0 per cent at one to six brands to 90 per cent at twenty-three or more, deleting the brand-richest observations at rates that differed by condition. Third, exact-string matching split one brand across its product names, so that Ridge and Ridge Wallet counted as different brands and an agreement rule discarded both. The deposited instrument reaches precision 262/262 and recall 22/24 against human annotation, with no failed extractor calls in 574 responses. Correcting the extraction changes the manuscript's central claim: the Valentine's arm is substantially unchanged, while the Christmas arm reverses on the distributional axis, so that the female-framed prompt receives both a narrower brand pool and a less even distribution within it. The package contains the raw model responses as returned, the extraction output with per-call extractor status, the superseded instruments deposited as the record of what they did wrong, the analysis code implementing rarefaction for unequal condition sizes and sub-sampling inference, and the human annotation files. Every figure in the manuscript recomputes from these files. See README.md for the instrument and DATASHEET.md for provenance. No personal data. All responses are model-generated text about consumer products; no human participants were involved.