Reported from the community: https://community.metabrainz.org/t/picard-3-matching-algorithm/815712

When looking up metadata for a cluster, Picard selected a wrong release, a different artist with a different tracklist, over the correct release that matched on artist, track count, title and date. Example from the report: the file cluster was Kachkin - World of Illusions (12 tracks, 2011), but Picard chose a NanoStorm album of the same title (9 tracks, 2017).

This reveals at least 2 issues:

Placeholder catalog number is treated as a real identifier. The correct release stored the MusicBrainz placeholder catalog number [none] (the value editors use to assert "no catalog number", since the field is mandatory). Because the file carried a real catalog number, [none] scored as a hard identifier mismatch (0.0), which capped the whole match low and discarded the strong album/artist/track-count/date similarity while a distractor with no  label-info escaped any catalog-number penalty entirely.

A matching album title can outweigh a wrong artist. The similarity tier is a weighted average in which the album title outweighs the artist, so a shared (often generic) title, optionally with a coincidentally equal track count, can rank a wrong-artist release close to the correct one. The matcher does detect the wrong artist, but it is averaged away.
It should be noted that the matching algorithm also tries to identify swapped title and artist.