Compute the page's composite and overall percentiles against the full English-language ranked cohort, not the top 100 rendered rows. The payload carries compact factor vectors and warning counts for every ranked publication, so both the default view and weight experiments use the same full-cohort denominator.
The top 100 remains the rendered subset. The daily job already performs the full crawl and replaces the snapshot; this removes the top-100 renormalisation from the browser.
🤖 Generated with Claude Code