For saving evaluation proofs

Runs `global_group_check` once per *optimizer step*. The trainer accumulates `group_id`s across microbatches; this callback fires after the o

Expects groups, and will calculate advantages for the groups

These guides just use c, c_p, and x to make a review prediction

For storing evaluation proofs for a series of rounds.

For keeping track of how we are doing on the target statements.

A lightweight HTTP task server for distributing a list of string tasks to many workers. - Start/stop programmatically from the master. -