Precision Control

Claim Type: mechanism_hypothesis
Scope: Precision control analogues and cognitive regimes
Depends On: ARC-005, INV-008, ARC-004
Status: provisional
Claim ID: MECH-002

Source: docs/processed/legacy_tree/architecture/precision_control.md

Precision control (monoamine analogue)

REE uses depth-indexed precision gains (\alpha_k) to control how strongly errors at different depths shape belief and action.

Minimal knobs

For each depth (k):

  • alpha_k multiplies the prediction error term in training / inference.
  • Optionally, alpha_k modulates sampling temperature (higher precision → lower entropy).

Functional roles (analogue)

Precision control does not merely tune learning rates. It induces qualitatively distinct cognitive regimes by reshaping how and where temporal collapse, commitment, and exploration occur.

Functional regimes (analogue)

  • Dopamine-like (commitment / trajectory locking): Increases precision at action and policy depths. Promotes temporal collapse of a selected trajectory. Phenomenology: confidence, motivation, goal-directed flow. Pathology when excessive: over-commitment, mania, compulsivity.

  • Noradrenaline-like (unexpected uncertainty / interrupt): Transiently increases gain on surprising inputs and resets prior expectations. Breaks existing phase-locks and suspends temporal collapse. Phenomenology: alarm, vigilance, attentional narrowing. Pathology when excessive: panic, hypervigilance, fragmentation.

  • Acetylcholine-like (expected uncertainty / sensory precision): Increases precision at sensory depths ((\alpha_\gamma)) under expected noise. Forces perceptual updating against top-down prediction. Phenomenology: clarity, vividness, perceptual anchoring. Pathology when excessive: sensory overload, derealisation.

  • Serotonin-like (anti-collapse / horizon widening): Reduces premature commitment by limiting precision escalation. Stabilises exploratory hippocampal rollouts across longer horizons. Phenomenology: patience, emotional buffering, tolerance of ambiguity. Pathology when excessive: apathy, blunted affect, indecision.

Design constraint: expected uncertainty (ACh-like) must be separated from unexpected uncertainty (NE-like); they are distinct control channels, not a single precision scalar.

Dopamine Precision‑Weighting (MECH-043)

Claim Type: mechanism_hypothesis
Scope: Dopamine-like modulation of precision-weighting for unsigned prediction errors
Depends On: ARC-005, INV-008, MECH-003
Status: provisional
Claim ID: MECH-043

Dopamine‑like signals should modulate the precision weighting of unsigned prediction errors, shaping learning and commitment without becoming a scalar reward objective. Misallocation of this precision is a plausible mechanism for hallucination‑like failures.

⸻

Precision as a controller of temporal experience

Precision allocation determines whether temporally displaced predictions are:

  • Collapsed into a lived present (high action-depth precision),
  • Held in exploratory suspension (balanced precision),
  • Or fragmented into competing, incoherent trajectories (misallocated precision).

Thus, shifts in precision do not merely alter accuracy or learning speed. They alter:

  • Whether a unitary “now” is constructed,
  • How tightly perception is bound to action,
  • How emotion and value predictions bias trajectory selection,
  • And whether experience feels continuous, urgent, fragmented, or unreal.

Signed Harm/Benefit PE Precision (Implementation Sketch)

This section operationalizes MECH‑054 (see control_plane.md) in precision‑control terms. It treats harm‑related and benefit‑related prediction errors as distinct channels with separate precision weights.

Let:

  • (e_H) = harm‑related prediction error (from HARM / HOMEOSTASIS streams),
  • (e_B) = benefit‑related prediction error (from VALENCE / reward‑proxy streams),
  • (\pi_H, \pi_B) = precision weights for harm vs benefit channels.

Local updates (illustrative):

[ \pi_H(t+1) = (1-\alpha_H)\pi_H(t) + \alpha_H f(|e_H(t)|, context) ] [ \pi_B(t+1) = (1-\alpha_B)\pi_B(t) + \alpha_B f(|e_B(t)|, context) ]

Where (f) is bounded and saturating, and (\alpha_H, \alpha_B) may differ by mode.

Usage:

  • Learning: ( \Delta z \propto \pi_H e_H + \pi_B e_B ) (applied per channel, not collapsed into a scalar).
  • Commitment gating: if ( \pi_H |e_H| ) exceeds a threshold, feed the aversive gate (MECH‑053) to suppress or delay commitment.
  • Unsigned precision: dopamine‑like gain (MECH‑043) can scale the overall learning rate, while signed precision determines which errors dominate.

This keeps hedonic tone (μ/κ overlays), valence vectors (MECH‑035), and signed PE precision conceptually distinct. See MECH‑055 in control_plane.md for the separation rule.

Operational control state (LC/astrocyte aware)

To make control effects observable and calibratable, REE should expose explicit state variables:

  • (A_t): tonic arousal baseline (maps to control-plane baseline channel, K7-like).
  • (N_t): phasic volatility/imminence signal (maps to interrupt pressure, K8-like).
  • (R_t): action readiness bias (maps to motor gating readiness, K9-like).
  • (C_t): slow regulatory context integrator (astrocyte-like modulation field, tied to MECH-001).

Illustrative updates: [ K2t = k2_0 + w_A A_t + w_N N_t + w_C C_t ] [ K2{H,t} = K2t + \Delta_H,\qquad K2{B,t} = K2_t + \Delta_B ] [ K1_t = k1_0 + u_C C_t - u_O overload_t ] [ K10_t = k10_0 - v_N N_t + v_S safety_margin_t ]

Interpretation:

  • (N_t\uparrow) lowers the effective veto threshold ((K10_t\downarrow)) and increases interrupt propensity.
  • (C_t) modulates slower plasticity/precision drift, capturing regulatory inertia rather than instant switching.
  • (K2_H) and (K2_B) are channel-specific precision weights; they should not collapse back to a single valence scalar.

Telemetry requirement (MECH-042-aligned):

  • Log (A_t,N_t,R_t,C_t,K1t,K2_t,K2{H,t},K2_{B,t},K10_t) per episode window.
  • Log commitment breaks/interrupts with triggering channels and thresholds.
  • Distinguish tonic drift from phasic spikes in diagnostics.

Calibration note: Quantitative tuning rules are intentionally deferred until the signal‑map wiring (MECH‑004) and regulatory‑stack framing (MECH‑001) are clarified. The current separation rules are invariant‑compatible scaffolding, not a final calibration protocol.

Astrocyte-aware regulatory stack

Note: The above framing treats monoamines as direct control knobs. A more neuroscience-informed perspective reinterprets monoamines as broadcast meta-regulatory signals that bias a slower astrocytic regulatory substrate, which then modulates precision/gain/plasticity with spatial and temporal lags.

See docs/astrocyte_aware_regulatory_stack/ for:

  • Why monoamines are not direct knobs (astrocytes mediate their effects).
  • How to model the slow regulatory field (R(x,t)) that produces (\alpha_k(x,t)).
  • Implications for care budget, inertia, and sleep recalibration.

For REE-v0, the direct-knob model (above) is a valid simplification. Future implementations should account for the layered regulatory architecture documented in the astrocyte-aware module.

Open Questions

None noted in preserved sources.

  • MECH-002
  • MECH-043
  • MECH-054
  • MECH-035
  • MECH-055

References / Source Fragments

  • docs/processed/legacy_tree/architecture/precision_control.md
  • docs/thoughts/2026-02-11_habenula_signed_pe.md

REE is developed by Daniel Golden (Latent Fields). Apache 2.0.