Control_Signals

Control Signals (Overview)

Control signals are Orrin's regulated internal state: a reward signal, an activation level, and a set of named pressures (demands) that bias what runs next. They are the affective layer, modelled on core-affect and reward-prediction-error theory (Russell & Barrett; Schultz), and they are stored as raw numbers in context["affect_state"] (with a core_signals sub-vector).

Two readers, on purpose

The two halves of cognition read the same state differently:

• Background machinery reads the raw floats. The bandit function-selector, the attention hijacker, and the cost-prediction layer use the numbers directly to bias selection.
• The reasoning layer never sees a number. brain/control_signals/signal_summary.py renders the signals into qualitative descriptions that name the felt quality ("a heaviness, like moving through something thick"), never the signal label or its value. Only that text reaches the inner-loop prompt, the self-descriptor, and the speech gate.

The intent: the reasoning layer reads its own state estimate, not a raw readout. Signals are adaptation-adjusted first, so a state the runtime has habituated to stops dominating.

Regulation

Control signals decay toward per-signal setpoints (not a flat midpoint) under a velocity budget, so state integrates rather than lurches (setpoints.py, regulation.py, homeostasis.py). A convergence layer (arbiter.py) means reactive and analytical subsystems submit proposals to a single writer instead of racing on the shared state file.

Where they come from and go

• Sources: prediction error, demand satisfaction/frustration, host-coupling deviation, peer signals, and workspace writeback.
• Effects: function selection, ignition likelihood, exploration vs. exploitation, attention capture, and how Orrin describes its own state.

Tuning

Most setpoints are learned or in-code; a few knobs are env-exposed (ORRIN_ALLOSTATIC_SETPOINT, ORRIN_SIGNAL_DECAY, ORRIN_INTEROCEPTIVE_AFFECT). See Tuning Control Signals and the Deep Dive.

Code pointers

• brain/control_signals/ — the whole subsystem
• brain/control_signals/signal_summary.py — the qualitative renderer (authored copy, load-bearing)
• brain/control_signals/arbiter.py — the single-writer convergence layer