Used product-copy-context.md as intended -- checked the live public page against its own rules while the experiment is mid-fill -- rather than only annotating it. The page honours the rules that matter: it says "Backfilling network history / Live subscriptions open when the archive catches up", reports "0 listening" honestly while serving is gated, and gives build provenance without promoting the implementation language. But "up" and "live events this run" are process-local, and the process restarts at every backfill batch boundary. At 55.6 h elapsed with 16,728,092,842 events archived, the homepage read: up 12m . 289,369 live events this run . 0 listening Both correct, both useless -- understating a two-and-a-half-day archive by ~275x because the last restart was 12 minutes earlier. A visitor cannot distinguish a permanently-young instance from a brand new one. Added the rule: durable values (archived event count, elapsed since the run began, complete repository count) belong on /, process-local figures on /status where the reader is looking at this process. Noted this is the HttpConnectionClosing batch-boundary defect surfacing publicly, but that the copy is worth fixing either way, since a durable counter is the right thing to show even once the process stops dying. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>