Bad merges have been mentioned a couple of times recently and I think one of the contexts with Ben's gene/protein work.
I think there are two general issues here which could be improved:
1. Merging is too easy. Because splitting/unmerging is much harder than merging, particularly after additional edits, the process should be biased to mark merging more difficult.
2. The impedance mismatch between Wikidata and Wikipedias tempts wikipedians who are new to wikidata to do the wrong thing.
The second is a community education issue which will hopefully improve over time, but the first could be improved, in my opinion, by requiring more than one person to approve a merge. The Freebase scheme was that duplicate topics could be flagged for merge by anyone, but instead of merging, they'd be placed in a queue for voting. Unanimous votes would cause merges to be automatically processed. Conflicting votes would get bumped to a second level queue for manual handling. This wasn't foolproof, but caught a lot of the naive "these two things have the same name, so they must be the same thing" merge proposals by newbies. There are lots of variations that could be implemented, but the general idea is to get more than one pair of eyes involved.
A specific instance of the structural impedance mismatch is enwiki's handling of genes & proteins. Sometimes they have a page for each, but often they have a single page that deals with both or, worse, a page who's text says its about the protein, but where the page includes a gene infobox.
This unanswered RFC from Oct 2015 asks whether protein & gene should be merged: https://www.wikidata.org/wiki/Wikidata:Requests_for_comment/Oxytocin_and_OXT...
I recently ran across a similar situation where this Wikidata gene SPATA5 https://www.wikidata.org/wiki/Q18052679 is linked to an enwiki page about the associated protein https://en.wikipedia.org/wiki/SPATA5, while the Wikidata protein is not linked to any wikis https://www.wikidata.org/wiki/Q21207860
These differences in handling make the reconciliation process very difficult and the resulting errors encourage erroneous merges. The gene/protein case probably needs multiple fixes, but many mergers harder would help.
Tom