Nightly run (04:30 UTC): register rebuilt from the rule, all time series fully re-collected, fits recomputed. The canonical artefact is a versioned JSON file in the repository — git is the archive.
Per event: baseline estimate (median of the last 14 post-peak days), then log-linear regression on the post-peak course → decay constant λ, half-life = ln 2 / λ, coefficient of determination R². Status rules: "measured" (robust fit), "measurement running" (peak younger than 21 days), "no measurable attention" (peak < 1,000 views, R² < 0.30, fewer than 8 usable post-peak days, or no measurable decay, λ ≤ 0). Provisional fits of running measurements are not archived. The headline number is the median over all measured events. Views per death = peak ÷ death toll. The inequality axis plots views per death logarithmically — one tick per event, the extremes labelled; events with no measurable attention do not appear and are reported as a count. On the decay curve the dot marks the half-life from the fit; the curve itself is the raw course — the distance between the two is the goodness of fit (R²).
One Cloud Run execution per night (< 5 min), ~300 throttled REST calls. No BigQuery costs: the planned large-table scan was prevented by our own cost brake and replaced with the REST API — the renunciation is stated here because it is part of the method. The site itself is static.