I'm definitely not an expert, but I seem to recall that without knowing the mechanism, the ability to observe the outcomes in the reduced space can help too.

Maybe by coming up with a model that closely enough correlates to what the actual mechanism produces. No perfect match needed, a really good overlap should be enough to generate collisions.