Thunderbird's adaptive spam filter learns from classification choices. Deleting an unwanted message is not the same training signal as marking it as spam. A useful review includes both unwanted messages and legitimate messages that were incorrectly classified.

Use the labels shown by your version

Mozilla documents that Thunderbird 145 and later use Spam where earlier versions use Junk. Locate the corresponding classification control in the installed version. Mark unwanted messages before deleting them if you want the adaptive filter to learn from them.

Review the spam folder for legitimate mail and mark confirmed mistakes as Not Spam or Not Junk. Thunderbird also supports training with legitimate messages elsewhere in the account. Do not open suspicious attachments or load remote images merely to decide whether an unsolicited message belongs in spam.

Check the account's classification and movement settings before interpreting a disappearing row. Marking a message can move it to a configured folder, so note the destination. Global settings and account settings can both affect the result.

Keep a false-positive example useful

For a wrongly classified project notification, record the sender, approximate date and where it was found without copying the full message into a public report. Correct its classification, then verify that the needed message is accessible in the expected folder.

Compare with the provider's webmail if the same messages are already in the server's spam folder before Thunderbird processes them. The provider may have its own filtering rules. Changing the desktop classifier does not establish that every server-side decision has changed too.

Continue reviewing a manageable sample over time. Adaptive classification is probabilistic; one corrected message does not guarantee that every future message from a similar source will be treated identically. If a critical workflow needs a specific exception, ask the mailbox owner to assess that exception's scope separately.

The useful maintenance result is fewer observed mistakes with a clear correction process. Avoid promising a spam-free inbox or treating a large number of deletions as evidence that the classifier has been trained.

Source: Thunderbird documentation.