Deterministic complement to a RouteLLM Thompson-sampling bandit:
small/local = easy / low-"frequency",
large/cloud = hard / high-"frequency", with a tunable
PI-correction bandwidth ω_c as the crossover knob. Adopted from Li Yu / IEEE 911711,
generalized under SZL governance. ← a11oy
Routing weight vs query difficulty (active-flux crossover)
A lower PI bandwidth keeps the small/local model
dominant to higher difficulty; a higher bandwidth hands off to the
large/cloud model sooner. The crossover line marks the
difficulty at which routing flips. MODELED 1st-order complementary blend.
small/local weight
large/cloud weight
crossover
current query