Because the allowlist and blocklist operations work absolutely based on the sender, right? Is this before the Bayesian filters? Please let me know if I’m understanding this incorrectly.
Yes, the allowlist and blocklist have higher priority than the Bayesian classifier.
If the sender is a real person, and you like some of their messages but not others, we discourage you from training the unwanted ones as spam. Some people do this anyway, and SpamSieve can indeed help you sort through the messages, but it can also reduce the filtering accuracy (for other senders), as the unwanted messages are not actually unsolicited bulk messages.
The other way this can happen is if a spammer is pretending to be a legitimate sender, and so you have genuine good messages and true spam messages coming in from the same address. In this case, you should train the spam ones as spam (and the good ones as good).
What will happen here is that, if you end up with rules for this sender in the allowlist and/or blocklist (e.g. that SpamSieve auto-creates when you train messages), it will realize that this name or address is sending both spam and good messages and disable the allowlist and blocklist rules. Then, going forward, SpamSieve will analyze the full contents of these messages to determine whether they’re spam.
It’s important not to delete the allowlist and blocklist rules because you want SpamSieve to remember that they’re disabled. You don’t want it to go back to accepting or blocking all the messages.