Moderation is at the AppView level, not at the protocol level. If you create an AppView you can choose to use bsky.app's moderation, or use another app's moderation, or use no moderation at all, for the same base of user posts. Far as I know the relay does do sybil attack and CSAM mitigation, but that's about it, and there are other relays going up, each of which is free to choose how it tackles CSAM, etc.. Or an AppView can pull directly from a set of PDSs without a relay in between. Or a bespoke partial relay.
With a raspberry you could spin up an AppView in a few days that shows all the posts that the bsky.app moderation team has taken down, minus presumably the illegal stuff that by law they had to deal with on the lower-infra level.
Fiatjaf's argument would be that other Kind1 type clients are incentivised to just reuse bsky.app's moderation since it's way expensive to do one's own moderation, and that argument holds weight--but again the time dimension, we need to wait and see how this plays out over time.