Transaction logs may be reporting illegal actions

[eharpste]

When running a batch training learning curves sometimes show more opportunities than are possible in the problem set. This seems to be from times that agents gets a step A correct, then does some step B, then attempts to go back to attempt step A, which is marked as incorrect. In a human using CTAT this would never happen because if you try to take a previously correct steps the field will be locked and so it will never appear. Unsure how we want to handle this long term, because its kind of interesting, but in the short term it violates the assumptions of transaction roll ups for learning curve analysis.

[eharpste]

It is possible this is desirable behavior and the rollup process we were using from PyAFM is actually wrong.

[eharpste]

We are pretty sure this is an issue related to a particular tutor and not the framework itself. Closing for now unless it comes up again.