ATProto weather, as observed.

Each figure is the mean rate over the observed interval, not the current rate. The sparkline shows the per-window series it averages.

Above 300 windows a sparkline is drawn one column per rendered pixel; here each column spans 355 windows, with the line following the column maximum and the shaded range showing its minimum to maximum. Averaging would erase the spikes a sparkline exists to show, so neither bound is dropped.

latest /s is the most recent observed 60-second window — not an instantaneous reading, not a live gauge, and not an average over the run. baseline is the mean of the last 15 eligible windows before it; partial, gapped, lossy and coverage-degraded windows are skipped rather than filled.

Conditions are threshold cuts on that z-score (surging z≥3, elevated z≥1.5, quiet z≤−1.5, degrading z≤−3; unknown below 5 samples). They are not calibrated against anything and carry no statistical warrant beyond “this window looked different from the recent past”.

Every ratio is a two-body system: a/b can move because a rose, because b fell, or because both did — and it means little when b is small. The ratio is the hint; the primitive cards above are the receipts, and each ratio's numerator and denominator appear there as their own cards.

Extremes are drawn only from windows whose denominator reached 30. A ratio is a two-body system: divide by four and it will happily report 22×, which is arithmetic rather than weather. Saying so in prose is not enough — a caveat does not travel with a screenshot — so the thin windows are excluded from the min and max columns and counted in the last one instead of being quietly dropped. 30 is a legibility floor, not a statistical one: no power calculation says a ratio becomes sound at 30 events, only that a single-digit denominator should not be allowed to set a headline.

These are coarse periods in which an aggregate event rate departed from its own trailing baseline. No account identifier is published. That alone is not anonymity: precise aggregate timing and counts can be joined back to the public firehose. Publication therefore fails closed unless the existing local edge store independently observed at least 10 distinct actors performing that action during the period, and — for a rate excess — unless no single actor emitted as many events as the departure itself. Eligible times are rounded outward to 3600-second UTC periods and exact counts, rates, statistics, temporal shape, and stable episode identifiers are omitted. Covering coarse UTC periods from 2026-10-20T16:00:00Z to 2026-10-21T16:00:00Z.

Observation is not causation, and magnitude is not a verdict. Bands are provisional effect-size cuts. The actor floor is also provisional disclosure resistance, not a statistical threshold and not a claim that the remaining rows are anonymous.

Read the actor count precisely: it is how many accounts were acting in that period, not how many produced the departure. The second gate exists because the first cannot tell those apart — an interval busy with unrelated traffic clears an actor floor no matter who caused the excess. Neither gate attributes a departure to anybody, and the hour rounding is defence in depth rather than the control that bounds reconstruction: the primitive cards above publish the same metric per window.