Tuning_Control_Signals

Tuning Control Signals

Control-signal behavior is mostly learned and in-code (setpoints adapt to experience), so tuning is a light touch, not a config dial. Reach for it when regulation misbehaves, not to chase a target number.

When to tune

• Signals thrash (rapid oscillation that destabilizes selection).
• A single state dominates and never habituates.
• Reward looks stuck and you've confirmed it's regulation, not a lack of production.

What you can actually change

Knob	Effect
ORRIN_ALLOSTATIC_SETPOINT	Baseline the signals regulate toward
ORRIN_SIGNAL_DECAY	How fast signals relax toward setpoint
ORRIN_INTEROCEPTIVE_AFFECT	Whether interoceptive (host/body) state feeds affect
ORRIN_IGNITION_GATE	Whether ignition is gated at all (0 = always deliberate)

Most setpoint and velocity behavior lives in brain/control_signals/setpoints.py and regulation.py — change those in code, with a test, rather than adding env knobs.

Process

1. Change one thing, small.
2. Run a scenario benchmark (ORRIN_BENCHMARK=1) or a short staging run and watch the signal rings / reward curve in the Brain room.
3. Compare before/after; keep only if the pathology improves without new instability.

Rollback

State is snapshotted on reset, and config changes are small and reversible. If a change destabilizes a run, revert the single knob and re-baseline. See Control Signals: Deep Dive for the underlying mechanics and Troubleshooting for symptom→cause tables.