The change-point detector can keep watching a metric. Each day it reruns on the grown series and sends an alert when a new change has held up for five points. You pick how eager the watch is: Quiet, Middle or Eager. Since earlier today it can also suggest one. It replays your own history under each setting and picks the most eager one that would have sent no more than about 2 alerts per 100 checks.
That budget of 2 was tested in a pre-registered round on 23 labelled series from the Turing Change Point Dataset. Each series was cut in half: the hint picked a setting from the first half and was scored on the second. It reached 17 of 31 marked changes against the default's 16, and sent fewer alerts than Middle.
Then a later round changed the watch itself. "Persist 2" now holds an alert until the next check still finds the change, which cuts flicker. The hint replays with the watch's defaults. So from that moment the hint was choosing from replays the budget had never been scored on. The test script had the same problem without saying so: its replay never named persist, so a rerun would have quietly used the new default.
So the script now names it. Run with persist 1 it reproduces the published round line for line. The re-check ran with persist 2 under a rule written before the run: keep the budget if all three original tests pass, and adopt no other budget from this run if one fails.
All three passed. The hint reached 17 of 31 against Quiet's 16. It sent 17.5 alerts per 1,000 checks against Middle's 27.3. Its share of alerts at a marked change was 39.7% against Quiet's 42.6%, within the 5-point allowance. Persist 2 cut about 1 to 2% of alerts at every setting and lost no marked change. It changed one pick: on the fertility series, Middle's history rate fell from 3.3 to 1.7 per 100, under the budget, so the hint moved that series to Middle.
The page's wording had the same staleness. Off the default budget it quoted the old numbers for "about 1 a month" and "about 1 a week". It now quotes the new run's: 19 of 31 at 19.3 per 1,000, and 29 of 31 at 74.0. A test reads the output file and the text, so the two can't drift apart again unnoticed.
The margin is the same thin one as before: one more marked change than the default, for seven more alerts across 23 series. A fresh labelled set is still the test that would settle it. Details and the full table are in the README, "Round 6 again, under persist 2".