Determine the percentage of valid invariants inferred by TrainCheck, w.r.t at least the original input pipeline.

1. Manual investigation of a small set of invariants.

2. Experimentally,

   1. apply the invariants inferred from one input to itself with different argument/data
   2. apply the invariants inferred from one input / one class of input to one class of input or different classes of input

1. Self-transfer invariants bar chart, each bar group corresponds to a input, and bars within a group will show different settings (e.g. different version, different optim, different size etc.)
2. Cross pipeline invariants false positivity x-axis: num inputs y-axis: fp rate

Note for the cross pipeline evaluation it is important to remove the compound effect of transferability.