RegistrationResult interpretation for point cloud matching

I’d like to use a (global) registration method to estimate whether two point clouds are of the same object/scene (after rotation/and or translation). That is, I don’t particularly care about the transformation matrix, I just need a scalar (say in the 0-1 range) that gauges the chance the point clouds are the same (and ultimately a binary yes-no decision).

Can the RegistrationResult returned by the various registration methods be used to that end? Are there any heuristics regarding the value of fitness, inlier_rmse, correspondence_set or any combination thereof to basically tell whether the point clouds are “most likely the same”, “most likely different” or “ambiguous”?