How do you usually debug semantic similarity scores?

When you compute similarity between two words, do you just trust the number, or do you inspect which senses were matched? One reason I built WordNet Explorer was to make similarity explainable at the synset level.

Curious how others approach this when building NLP systems.