The Default Index

19 web challenges, from a Linear clone to a Pokémon battle screen. Each page is reduced to the choices it made: typeface, light or dark, accent colour, icon set, and 23 habits like frosted glass, gradient text, glow, emoji icons, uppercase labels and floating animations. A habit a page skips is not counted, so doing less makes a page neither more nor less default.

Every page another lab shipped for the same prompt before this model came out. A lab's own earlier models are not the crowd, they are the family. Labs with many variants count less: each lab weighs the square root of its model count, so a lab with sixteen variants counts as four.

A page's typicality is its average overlap (Jaccard) with every earlier page. Its percentile is where that lands among the earlier pages closest to it in length, because longer pages carry more of everything. The Default score is the mean percentile across a model's pages, with a 90% bootstrap interval. A model needs 6 pages to get one. Scores freeze at release: a newer model never moves an older model's score.

Split-half reliability across web challenges: 0.76. SVGs (0.293) and the shelf answers (0.08) are too unstable per model to score, so they are reported, not ranked. Page length against percentile after matching: -0.106. Stealth builds and the model they turned out to be land 10.6 points apart on average; random pairs from different labs, 16.7.

Not a quality score. A default page can be a good page, and the most original page in the set may not load. One sample per model per prompt, so this describes the field, not a model's own variety. The prompts have been public since 2025, and later models may have read earlier answers. That would make them more default, which is the thing being measured.

Recomputed from the archive on every build. No judge, no API calls, no human in the loop. A new model is scored the day its web challenges land. The method is versioned; any change to it bumps the version on this page and in the dataset.