Yep, it checks out for large matrices (e.g. with a 5000x5000 data matrix, check_finite is at least half of the execution time). Here is another thread where we showed that in the way you described: Math library input checks - #41 by seantalts
As of now, I haven’t seen operands_and_partials showing up in profiling anymore (after the bug fix). Does anyone have a new example?