Generally, I think you want to set very strict criteria on a bot that automatically reverts vandalism.

E.g. with this query, you can see that setting a confidence threshold at 0.96 "probability" of damaging, you can expect to get 90% precision (wrong 1/10th of the time) and 17.7% recall (catches 17.7% of the most damaging edits):

https://ores.wikimedia.org/v3/scores/eswiki?models=damaging&model_info=statistics.thresholds.true.%27maximum%20recall%20@%20precision%20%3E=%200.9%27

Patrubot is no longer performing antivandalism tasks (it's been replaced by SEROBot from @Superzerocool). Perhaps this task is no longer relevant and could be closed.