base-on solved 6/8 with 2 aborted — the model ran out its eight-turn budget
without reaching a disposition, which scores exactly 0.000 and is recorded as
such rather than dropped. base-off solved 7/8. Both arms sit below the scripted
policies on the shipped weights, and above them on nothing, which is the honest
shape of a model that has not been trained on this task.
72 fixtures across both environments replay at delta 0.
karti-ai