Thinking / Action Selection

brain/think/ decides what runs next, every cycle. It is the module where control signals, the workspace, goals, and learned value all converge into a single pick.

The selection path

1. Deliberation gate (deliberation_gate.py, should_think()) — decides whether this cycle ignites into full deliberation or stays in low-power default mode (see Workspace and Ignition). Non-ignited cycles damp effortful functions so quiet time drifts toward cheap work.
2. Function selector (think_utils/select_function.py) — a contextual bandit picks the next cognitive function. Control-signal floats, demands, cost predictions, and the workspace winner (an additive prior, ORRIN_WORKSPACE_PRIOR) all bias the draw. The bandit honors an explicit learning rate (constant-step), with a modulated adaptive rate available. Candidate scoring (think_utils/selection/score_actions.py) also adds each action's learned reward EMA (once the action is mature, ≥8 observations) directly, so realized outcomes — not just priors — decide what wins (see Learning and Adaptation).
3. Action arbiter (action_arbiter.py) — the convergence layer's action half: reactive and analytical subsystems propose, the arbiter resolves, so nothing races on shared state.
4. Execution — the chosen function runs; sandbox_runner.py executes generated code safely; outcomes feed reward accounting and the effect ledger.

Deliberate reasoning (System 2)

• inner_loop.py — the deliberate draft→critique→revise reasoning loop, recruited on ignited and conflicted cycles (ORRIN_CONFLICT_RECRUIT) rather than fired on a schedule.
• inner_loop_symbolic.py — the symbolic-only variant, so deliberation works without an LLM.
• inner_loop_critique.py — the critique pass; depth_bandit.py (UCB1) learns how many rounds are worth running; meta_controller.py chooses shallow vs. deep chains for goal pursuit.
• scratchpad.py and simulate.py — working notes and look-ahead simulation of candidate actions.

Cost-aware selection

brain/cognition/cost_prediction.py is an expected-value-of-control layer: each function's predicted cost (time, tokens, energy) competes against its predicted payoff, so expensive functions must be worth it. Attention can still be hijacked by urgent signals (brain/cognition/attention.py).

Speech

speech_builder.py, speech_coherence.py, speech_comprehension.py — building and checking person-facing utterances. Actual delivery goes through the one door: Expression Membrane.

Debugging

Bandit statistics, per-function EMAs, and EVC predictions are surfaced in the UI's Cognition room; instrument those first when diagnosing mis-selection. The Reward Auditor peer flags a reward signal that has collapsed to noise.

Code pointers

• brain/think/ — everything above
• brain/think/bandit/ — bandit implementation
• brain/cognition/cost_prediction.py — EVC layer