Every cycle, Orrin chooses one cognitive function to run. That choice is a contextual bandit — each arm is a function — biased by control signals, demands, the workspace winner, and predicted cost. It's the point where affect, goals, and learned value converge into a single pick.

flowchart LR
    CS[Control signals] --> SC[Score each function]
    DM[Demands] --> SC
    WW[Workspace winner<br/>additive prior] --> SC
    CP[Cost prediction / EVC] --> SC
    HS[Learned per-function value] --> SC
    SC --> AR[Action arbiter]
    AR --> PICK[Chosen function]

• Bandit selector (select_function.py) — scores functions from learned value plus the biases above, then draws. It honors an explicit learning rate (constant-step), with a modulated adaptive rate available.
• Workspace prior (ORRIN_WORKSPACE_PRIOR) — the broadcast winner is an additive prior, so "what's in mind" biases "what gets done."
• Cost prediction / EVC (cost_prediction.py) — an expected-value-of-control layer: a function's predicted cost (time, tokens, energy) competes with its predicted payoff, so expensive functions must be worth it.
• Action arbiter (action_arbiter.py) — resolves reactive vs. analytical proposals so nothing races on shared state.

Anti-repeat penalties, a meta-rut breaker, and outcome-devaluation keep a mediocre function from monopolizing cycles; the Reward Auditor peer flags a reward signal that has collapsed to noise.

Immediate reward updates the bandit each cycle; delayed credit arrives later from the evaluator daemons. Reward itself is grounded in the effect ledger. See Learning and Adaptation and Thinking / Action Selection for the deep dive.

• brain/think/think_utils/select_function.py, brain/think/bandit/
• brain/think/action_arbiter.py, brain/cognition/cost_prediction.py