Drill: Clone PostHog analytics dashboard: build, critique against real PostHog, iterate. Log divergences.

Interpretation: Rebuild of the 2026-10-22 funnel surface that published at 2.5/5. Rather than reskin the same template, the dataset was redesigned first so the worst leak (78.8% at step 4-5) is 4.3x larger than any other step's loss, giving the page one true subject. The finding is then stated in the first content band, the loss bands are drawn to scale between bars, and the global CTA is demoted so a page-level action never out-shouts a page-level conclusion. One accent means selected/active/data only; red means loss alone. Dark-mode native, depth by luminance stepping, no shadows or gradients.