During a GDC conference held by John Macdonald (Engineer largely in the data side of Valve's affairs), he stated that the training of VACnet was largely a "war of datasets", and he expected the quality to improve as time went on and Valve labelled and processed more internal data for their anticheat.

Considering the leak for "referring to your training manual" for the employee-access version of Overwatch match reporting for CS2 might be an indicator that they were labelling the data the whole time with contractors. Now seems to be the release moment :)