>..."This is the problem - no wait, we already proved it can't be this - but actually ..."

Ditto. Has me wondering why there isn't a reconciliation pass somewhere on the final output.

At least it's a decent signal for when model confidence is low.