Have you looked at the transitions with a scope?
Is there any undershoot?

The non-filtered version would be more likely to catch short lived pulses just under the threshold.