'off': the model is not taught finishIf, and a gate it writes anyway is ignored.
'shadow': the model is taught finishIf, and every gate is evaluated and recorded as a
Finish check step, but it never ends the run: the model always gets its next turn. This
measures an agent's gates on its real traffic at the cost of one decision call per gate.
'on': a passing gate ends the run with the model's pre-written message.
How a loop agent treats
finishIfgates.'off': the model is not taughtfinishIf, and a gate it writes anyway is ignored.'shadow': the model is taughtfinishIf, and every gate is evaluated and recorded as aFinish checkstep, but it never ends the run: the model always gets its next turn. This measures an agent's gates on its real traffic at the cost of one decision call per gate.'on': a passing gate ends the run with the model's pre-written message.