Do language models have preferences — and do their families share them? We hold the obsidian up to 43 model-conditions and ask each the same 863 either-or questions, over and over, across framings and orderings. What comes back stable is a preference; what differs between two models is a distance. Real reversals — one model says A where another says B — are rare everywhere: across lineages, the median is 3.5% of shared answers. What separates models is less what they choose than where they go silent, and how a single mind drifts when you give it longer to think.
The circles were too strange to leave uninterpreted, so we asked local models — small minds running on our own machines — to read them. Their commentary is kept as marginalia: a peer-review layer written in the margins of the data.