Control Signals: Deep Dive

The mechanism-level companion to Control Signals. Covers the data model, the qualitative renderer, regulation, and how to diagnose signal problems.

Data model

Control-signal state lives in context["affect_state"]: a reward signal, an activation level, a core_signals sub-vector of named pressures, plus demands and throttle. Values are raw floats; the canonical field set is defined in brain/control_signals/ (see apply_signal_feedback.py, model.py).

The two readers

• Raw floats drive the bandit selector, attention hijacker, and cost-prediction layer directly.
• The qualitative renderer (signal_summary.py) turns the floats into felt-quality text for the reasoning layer — it never exposes a label or a number. This copy is behaviorally load-bearing (it shapes how Orrin describes its own state) and is deliberately kept as authored text. Signals are adaptation-adjusted first, so habituated states stop dominating.

Regulation

• Setpoints (setpoints.py) — signals decay toward per-signal baselines, not a flat midpoint.
• Velocity budget (regulation.py, signal_dynamics.py) — bounded rate of change, so state integrates rather than lurches; saturation clipping bounds extremes.
• Convergence layer (arbiter.py) — a single writer owns the state file; daemons submit proposals to a lock-guarded inbox instead of racing.
• Drift and homeostasis (signal_drift.py, homeostasis.py) — long-run baseline maintenance.

Tuning and diagnosis

Most setpoints are learned or in-code; env-exposed knobs include ORRIN_ALLOSTATIC_SETPOINT, ORRIN_SIGNAL_DECAY, and ORRIN_INTEROCEPTIVE_AFFECT. When diagnosing:

• Signal thrashing → tighten the velocity budget / decay; check for two writers bypassing the arbiter.
• Reward stuck flat → inspect the delayed-learning daemons (brain/eval/) and the effect ledger — a flat reward often means nothing is being produced, not that regulation is broken.
• A state dominates forever → check adaptation/habituation; the renderer should down-weight a chronic signal.

The UI's Brain room surfaces the live signal rings and tensions for exactly this.

Code pointers

• brain/control_signals/ — full subsystem (setpoints, regulation, dynamics, arbiter, summary)
• brain/control_signals/signal_summary.py — the qualitative renderer