This is an idea to get data about the frequency of scroll anchoring feedback loops / adjustments causing further adjustments.

Run an A/B experiment:

In A, scroll anchoring is disabled but we measure the number of times we would have made an adjustment if it were enabled.

In B, scroll anchoring is enabled and we measure the number of adjustments made.  (We already have an UMA metric for this.)

By comparing A and B we can estimate how often the intervention feeds into itself triggering more intervention (which is likely to be a negative experience).