I just pick a recent submission from the AcousticBrainz front page
This has nine submissions, the key for the first one is F Minor, the key for the remaining eight is C Major. So it seems likely that the key for the first one is incorrect, and I vaguely remember @alastairp saying something the key being incorrectly calculated. If this is the case why don’t we remove these old records when they have more than one submission ?
More generally there is always going to be some variation for all the metrics, but where one metric is totally out should it not be dropped. Its not really viable for end users of this data to do this, not least because too many api calls would have to be made to get each submission. The data would be more useful if AcousticBrainz was cleaned up a bit or if AcousticBrainz offered a combined dataset for a recordingId that represented the average value of each metric. (Exactly what is meant by average of course requires some thought which is why it would be better if this was done by AcousticBrainz team who hopefully have a better understanding of the data then I do)