Thanks. I found that allowing a large relative tolerance didn't help. Diagnostics at least showed that one of the test points was way off in the comparison, as you've pointed out.
Restricting test values seems to be the safest bet. This solves the issue.