Same as the current labeling/moderation service: any participant can verify any other participant. Which verifiers gets a check to appear is a property of the AppView.
If Bluesky becomes evil, you just configure your AppView not to trust their verifications.
Of course, that's the problem: right now we mostly have one AppView (bsky.app), which is the current SPOF in the mitigation plan against the "Bsky becomes the baddies" scenario.