There can be multiple strategies to solve this.
This depends on both data structure, content (e.g. error messages vs useful payload), what transformation are applied (e.g. where extra items come from) and what exactly you want to have as an output (data structure, valid items requirements, items compounding/grouping strategy).
From what I understand, you may want to strip items that actually reflect errors. But that’s only a hypothesis. If you could share some sample data to illustrate the case and to play around with, it would be nice. I cannot even imagine how to emulate your data.
One of the approaches to provide the sample data is to pin outputs from HTTP Requests and/or Google Sheets (data read nodes) in workflow editor, and then paste the WF with that pinned data.
On posting a workflow with some data pinned see this short guide:
You can also edit pinned data to strip sensitive information.
Also it is always good to have a broader context around the WF fragment in question (i.e. a couple of critical nodes before the fragment and one or two nodes after).
Underlying data is critical where workflow works technically (no failures) and solution depends on logical decisions and data.
For e.g. this request was resolved succesfully thanks to the data and even data store made available: Append or Update Row does not work
And this one got stuck in endless hypothetising: Google sheet: data is replaced in the existing row instead of adding new row
I failed on the latter one being unable to guess what could have possibly gone wrong.