De-escalation labels
From an Astra history, compare staying on Astra with handing the rest of the turn to Sol.
s
t
stay label: keep Astra to the end of the turn
switch-and-commit label: hand off to Sol, sampled to the end of the turn
Trains "Safe to use Sol?"
stay
Qπstay(st, Astra)
1 / 1
switch
Qπstay(st, Sol)
2 / 3
Compare success rates and cost-to-go, incl. switch cost
Escalation labels
From a Sol history, compare staying on Sol with handing the rest of the turn to Astra.
s
t
stay label: keep Sol to the end of the turn
switch-and-commit label: hand off to Astra, sampled to the end of the turn
Trains "Astra needed?"
stay
Qπstay(st, Sol)
0 / 1
switch
Qπstay(st, Astra)
2 / 3
Compare success rates and cost-to-go, incl. switch cost
Astra request
Sol request
turn succeeds
turn fails