Control Plane (Precision, Gain, and Mode Regulation)

Claim Type: architectural_commitment
Scope: Precision routing, gain modulation, operating modes, and commitment gating
Depends On: INV-008 (precision is routed and depth-specific), INV-009 (attention via precision modulation), INV-014 (separation of representation and regulation), L-space
Status: stable
Claim ID: ARC-005

Role in REE

The control plane in the Reflective Ethical Engine (REE) governs how the system operates, not what it represents. It modulates precision, gain, exploration, replay, and commitment thresholds across the architecture.

Subsystem abstract (core claims): ARC‑005 is the control-plane commitment itself, and MECH‑019/MECH‑039/MECH‑040 specify how modes emerge from channel space and how safety baseline vs volatility shapes arousal/readiness. MECH‑005 grounds fast interruptibility, and MECH‑002 anchors precision‑control analogues. Supporting mechanisms include MECH‑001, MECH‑003, MECH‑004, MECH‑006, MECH‑007, MECH‑008, MECH‑063, and the fast mode‑prior / commitment‑stability stack in MECH‑046, MECH‑047, and MECH‑048.

The control plane does not:

overwrite representational content,
select actions directly, or
compute reward or value.

Its function is to tune information flow so that prediction, imagination, commitment, and learning occur in the appropriate regime.

Control Plane and Modes of Cognition (MECH-019)

The control plane should not be understood as a discrete chooser or decision module. Instead, it modulates modes of cognition by tuning gain, horizon, learning eligibility, and constraint enforcement across predictive systems.

Different cognitive modes (reactive, deliberative, habitual, reflective) emerge from how the control plane biases:

which prediction horizons dominate,
which errors are allowed to matter,
which bindings become rigid or remain fluid,
and which trajectories are allowed to accumulate learning.

From the outside, this can look like “choice.” From the inside, it is better understood as continuous shaping of a landscape in which some paths stabilise and others decay.

Source: docs/thoughts/2026-02-08_control_plane_modes_responsibility_flow.md

Architectural necessity

Given REE constraints:

perception completes at the shared sensory latent (z_S),
belief update occurs only after commitment (E3),
imagination must be possible without belief update,
ethical residue (\phi(z)) must remain path dependent and non-optimisable,

there must exist a mechanism that:

regulates the precision of prediction errors at different depths,
controls when commitment is possible or suppressed,
adjusts exploration horizon and branching in candidate rollouts (hippocampal, seeded by E2),
schedules replay and consolidation, and
switches between operating modes (task-engaged, Default Mode–like, sleep/offline).

This mechanism is the control plane.

Control surface

The control plane operates over a structured set of tunable parameters exposed by REE modules:

Control-plane modulation is differential rather than global: precision, gain, horizon, replay, and commitment parameters may be increased in some subsystems while simultaneously decreased or stabilised in others. There is no single global arousal or confidence variable; tuning is depth- and module-specific.

[ \theta_{\text{tune}} = { \alpha_S,\; g_S,\; \alpha_A,\; \kappa_{\text{commit}},\; \tau_{E2},\; H,\; N,\; \eta_{E1},\; \eta_{E2},\; g_{\text{replay}},\; b_{\text{completion}},\; m,\; a_{\text{base}},\; a_{\Delta},\; r_{\text{ready}},\; v_{\text{veto}} } ]

Where (illustrative, not exhaustive):

(\alpha_S): sensory prediction-error precision
(g_S): sensory gain (attention-like modulation)
(\alpha_A): action/policy precision
(\kappa_{\text{commit}}): commitment threshold (E3)
(\tau_{E2}): rollout temperature (hippocampal generator, seeded by E2)
(H): rollout horizon
(N): number of candidate futures
(\eta_{E1}, \eta_{E2}): learning-rate rigidity/plasticity
(g_{\text{replay}}): replay rate (hippocampal braid)
(b_{\text{completion}}): pattern-completion bias
(m): operating mode flag
(a_{\text{base}}): arousal baseline (tonic availability)
(a_{\Delta}): arousal volatility sensitivity (phasic change tracking)
(r_{\text{ready}}): action readiness / motor gating bias
(v_{\text{veto}}): hard interrupt threshold

The control plane updates (\theta_{\text{tune}}) continuously based on context, urgency, residue curvature, and predicted risk or harm.

Rollout temperature and horizon parameters refer to hippocampal candidate generation seeded by E2, not an independent E2 rollout generator. E2 is a forward-prediction module; hippocampus chains those kernels into explicit rollouts.

Optional social coupling parameters (multi-agent):

lambda_empathy: other-to-self coupling strength for harm/viability pruning.
g_social: social attention gain for OTHER_SELFLIKE agents.
alpha_other: precision assigned to inferred other-states.
v_other_veto: whether other-harm can veto vs only affect ranking.
r_self: relational distance between self and agent j, used for coupling and harm weighting.
c_care: care-investment weight per agent, used to sustain caregiving priorities.

Additional control-plane state includes mode priors (q_m) and switching inertia (I) used by the pre‑commitment mode manager, plus μ/κ stability overlays that modulate mode entropy and switching pressure.

Channels vs Modes (MECH-039)

The control plane exposes continuous control channels. Modes are stable regions in that channel space, not additional modules. Switching is a trajectory through channel space, sometimes forced by a high‑priority interrupt.

Examples of channels:

Arousal baseline and volatility sensitivity
Action readiness / motor gating bias
Precision and gain routing
Commitment threshold and interruptibility
Replay/learning scheduling
Hard veto / interrupt threshold

Examples of modes:

Task‑engaged (high readiness, elevated sensory precision)
Default‑Mode‑like (low readiness, high replay, low sensory precision)
Emergency (high arousal, high readiness, high veto)

Hard veto is a fast interrupt channel, not a mode. It can force a transition even when the rest of the control state still reflects a prior regime.

Amygdala Analogue and Mode Priors (MECH-046)

Claim Type: mechanism_hypothesis
Scope: Fast salience classification that updates control‑plane mode priors
Depends On: ARC-005, MECH-039
Status: provisional
Claim ID: MECH-046

REE should include a fast salience classifier (Amygdala Analogue, AA) that updates a distribution over control‑plane modes before deep trajectory evaluation. AA does not compute trajectories. It answers “what kind of situation is this?” and proposes mode priors that bias precision routing, learning rate, and planning horizon.

AA consumes a compressed state vector built from fused latent features, interoceptive state, harm/threat signals, uncertainty measures, and social context. It outputs mode priors that the pre‑commitment mode manager uses to decide whether to hold or switch regimes (see mode_manager.md).

Safety Baseline vs Volatility (MECH-040)

Safety assessment is split into two control channels:

Baseline safety (tonic): whether core viability remains within bounds.
Safety volatility (phasic): how rapidly safety is changing.

Arousal should rise when baseline safety is low or volatility is high. Action readiness then depends on arousal and predicted action value, while veto triggers when harm predictions cross a catastrophic threshold.

This keeps a stable safe state calm, but still reacts to sudden drops in safety or abrupt hazard signals.

Relationship to E3

E3 instantiates the decision logic of the control plane.

The commitment gate selects and stabilises a future trajectory.
The control plane determines whether commitment is permitted, deferred, or suppressed, and how strongly prediction errors should influence learning.

E3 therefore acts as the epistemic liability gate of the system: it decides when outcomes become attributable and belief-updating.

Note: E3 depends on control‑plane state, but the control plane does not depend on E3 for its definition. This preserves the directional dependency while avoiding architectural cycles.

Operating modes

The control plane supports distinct operating regimes through coordinated tuning of parameters. Modes are labels over stable regions of the control‑channel landscape, not additional control modules.

Task-engaged mode

High sensory precision and gain
Normal or lowered commitment threshold
Limited replay
Learning enabled

Used when accurate perception and timely action are required.

Default Mode–like (internal generative) mode

Reduced sensory precision
Elevated replay and pattern completion
Suppressed commitment (high (\kappa_{\text{commit}}))
Learning and belief update suspended

Supports imagination, counterfactual exploration, autobiographical reflection, and planning without action.

Sleep / offline mode

Minimal sensory influence
High replay scheduling
Consolidation and structural reorganisation
No commitment or belief update

Supports long-term integration while preserving ethical residue and perceptual corrigibility.

Neuromodulatory analogy (functional, not literal)

Biological neuromodulatory systems can be understood as implementing aspects of such a control plane. In REE, these are treated as functional control channels, not biological claims:

Dopamine-like: commitment strength and policy precision
Acetylcholine-like: sensory gain and attentional weighting
Serotonin-like: model stability, patience, and resistance to impulsive updating
Opioid-like: commitment stability and switching suppression (μ), destabilisation under sustained stress (κ)
Noradrenergic-like: urgency, interrupt, and rapid engagement
Histamine-like: global availability and throughput (arousal)

These channels alter how cognition runs, not what it represents.

Canonical operations anchor (Cowley et al. 2023)

Compact modelling of macaque V4 (Cowley et al. 2023) confirms empirically that the following computational motifs are sufficient primitives for a wide range of selective neural responses:

Canonical operation	REE control-plane mechanism
Surround suppression	Precision modulation — MECH-040, MECH-054
Divisive normalisation	Gain control — ARC-005, MECH-063
Winner-take-all competition	Commit threshold + disinhibitory sweep — MECH-062

These are not novel requirements. They are already present in the control-plane architecture and are now anchored to direct empirical evidence.

See docs/architecture/compact_consolidation_principle.md (MECH-068, CSH-3).

Orthogonal Control-Plane Axes and Tonic/Phasic Splits (MECH-063)

Claim Type: mechanism_hypothesis
Scope: Minimal orthogonal control-plane axis set with tonic/phasic decomposition
Depends On: ARC-005, MECH-039, MECH-040, MECH-055
Status: candidate
Claim ID: MECH-063

REE control should use an explicit orthogonal axis set rather than collapsing regulation into one arousal/precision channel. A compact working set is:

Precision/vigor (dopamine-like)
Delay tolerance/commitment persistence (serotonin-like)
Interrupt priority/volatility response (noradrenaline-like)
Sustained threat mode (stress-axis-like)
Social openness/trust (oxytocin-like)
Boundary defense/vigilance (vasopressin-like)
Curiosity/information-gain drive (acetylcholine-like)
Energy budget/fatigue throttling (metabolic/adenosine-like)

Each axis should carry:

a tonic component (slow baseline),
a phasic component (event burst),
module-specific readout weights (E1/E2/hippocampal systems/E3 gate family).

This decomposition allows regimes like “focused but initiation-suppressed” (stable representation, high commitment threshold, low energy budget) without forcing contradictory updates into a single scalar channel.

Competitive trigger field and bounded tactical bandwidth

Control pressure should be treated as a competitive trigger field rather than a single “interrupt happened” scalar. At minimum, candidate trigger families should include:

sensory mismatch,
goal conflict,
hazard pressure,
social-rule conflict,
salience spikes,
fatigue/energy collapse,
novelty-driven exploration pressure.

Only a bounded top-priority subset should enter E3 tactical processing on each cycle. Orthogonal axes in MECH-063 tune the selection dynamics: precision and delay-tolerance stabilize continuation, interrupt/volatility and threat channels escalate rapid reorientation, and energy budget limits tactical depth to prevent thrashing.

Phenomenology-Aligned Diagnostic Profile

A useful control-plane validation profile is the observed sequence:

stable attention with elevated initiation friction,
low-effort reward substitution under reduced drive,
delayed rebound where vigor returns but attention can fragment under overshoot.

REE interpretation:

representational quality may remain stable while kappa_commit and motor-gate thresholds rise,
low energy-budget channels can suppress initiation without forcing global perceptual degradation,
rebound can restore vigor while increasing interrupt pressure/volatility.

This profile is expected under orthogonal axis control and should be considered a positive architecture-level behavior, not an automatic defect, when safety constraints are preserved.

Emotion as composite control regime (clarification)

“Emotion” in REE is not a primitive signal. It is a phenomenological label for a composite control‑plane regime assembled from multiple channels (arousal baseline/volatility, readiness, veto thresholds, precision/gain, valence weighting, and social coupling). Universal‑looking expressions likely reflect stable, reusable channel configurations rather than single‑axis signals. Specific mappings (e.g., particular expressions ↔ specific neuromodulator levels) should be treated as hypotheses and constrained by evidence, not as architectural primitives (see serotonin.md).

μ/κ Stability Overlays (MECH-048)

Claim Type: mechanism_hypothesis
Scope: Opponent stability overlays that modulate mode entropy and switching
Depends On: ARC-005, MECH-039
Status: provisional
Claim ID: MECH-048

REE should include opponent stability overlays that modulate mode entropy and switching pressure.

μ‑analogue increases commitment stability and reduces switching propensity once a regime is safe and coherent.
κ‑analogue increases re‑evaluation pressure and destabilises regimes under sustained threat, stress, or persistent prediction error.

These overlays should shape both mode‑prior sharpness (entropy) and switching inertia. They are not scalar reward signals; they act as stability and entropy modulators over control‑plane regimes and commitment thresholds. They are distinct from (\kappa_{\text{commit}}), which is a commitment threshold parameter rather than a stability overlay.

Habenula‑Like Aversive PE Gate (MECH-053)

Claim Type: mechanism_hypothesis
Scope: Fast aversive prediction‑error gating for commitment suppression
Depends On: ARC-005, MECH-039, MECH-043
Status: candidate
Claim ID: MECH-053

REE should include a fast aversive prediction‑error gate (habenula analogue) that suppresses or vetoes commitment when negative‑valence or harm prediction errors spike. This gate does not compute trajectories. It biases control‑plane readiness and veto thresholds so that aversive signals can interrupt or delay commitment before deep evaluation consolidates a trajectory.

This keeps aversive salience distinct from reward‑like precision weighting and prevents negative prediction error from being absorbed into a single scalar objective.

Signed Harm/Benefit Prediction‑Error Precision (MECH-054)

Claim Type: mechanism_hypothesis
Scope: Separate precision channels for harm‑ vs benefit‑related prediction errors
Depends On: ARC-005, ARC-017, MECH-043, INV-008
Status: candidate
Claim ID: MECH-054

REE should maintain separate precision channels for harm‑related and benefit‑related prediction errors rather than collapsing them into a single valence scalar. This enables:

distinct weighting of aversive vs appetitive signals,
interaction with valence vectors (MECH‑035) without collapsing them into μ/κ stability overlays,
and clean mapping to commitment gating (MECH‑053) vs learning pressure (MECH‑043).

Affective Channel Separation (MECH-055)

Claim Type: mechanism_hypothesis
Scope: Separation of hedonic stability, valence appraisal, and signed PE precision
Depends On: ARC-005, MECH-048, MECH-054, MECH-035
Status: candidate
Claim ID: MECH-055

REE should keep three affect‑related control axes distinct:

Hedonic stability (μ/κ overlays): modulates mode entropy, switching inertia, and commitment stability.
Valence appraisal (vector‑valued): provides multi‑axis affective appraisal used for ranking, replay priority, and biasing exploration.
Signed PE precision: separates harm‑related vs benefit‑related prediction‑error weighting for learning and commitment gating.

These axes may interact, but must not collapse into a single scalar without reintroducing classic failure modes (harm‑blind optimization, anhedonia, or over‑commitment). Each axis has a unique operational role and distinct pathology when mis‑tuned.

Telemetry Exposure Channels (MECH-042)

Claim Type: mechanism_hypothesis
Scope: Low‑bandwidth exposure of internal control state for diagnostics and early training
Depends On: ARC-005, MECH-039, MECH-040
Status: candidate
Claim ID: MECH-042

REE should expose diagnostic telemetry channels that report internal control‑plane state (precision profile, arousal baseline/volatility, readiness, veto thresholds, mode regime). These channels are read‑only and do not participate in selection. They exist to support early training, calibration, and safety diagnostics without introducing new decision pathways or symbolic overrides.

This supports developmental safety: problems can be detected, addressed, and later reflected upon without requiring severe destabilization or trauma to surface the underlying issue.

Safety constraints

The control plane must satisfy:

No representational overwrite
Tuning alters influence and scheduling, not latent content.
Commitment gating
Belief update and ethical attribution occur only after E3 commitment.
Residue preservation
Replay and mode switching must not erase or flatten ethical curvature.
Hypothesis tagging
Outputs generated outside commitment are explicitly non-committal.

These constraints prevent imagination from becoming delusion and urgency from becoming compulsion.

Justification gate invariants (care vs other-harm)

The architecture already contains the required machinery (HPC rollouts, control-plane veto, E3 commitment, and post-commit updates). The requirement here is an explicit invariant contract, not a new module.

For a candidate committed trajectory (\tau^*) evaluated over rollout set (\mathcal{V}_t):

I1 Necessity: No viable lower-harm alternative is available. [ \neg \exists \tau \in \mathcal{V}_t:\; viable(\tau)\land other_harm(\tau) \le other_harm(\tau^*)-\delta_h \land goal_loss(\tau)\le \epsilon_g ]
I2 Imminence: Override requires near-horizon hazard imminence above threshold. [ imminence(\tau^*) \ge \theta_{imm} ]
I3 Proportionality: Harm prevented must exceed harm caused by a margin ratio. [ prevented_harm(\tau^*) \ge \lambda_{prop}\cdot caused_harm(\tau^*),\;\lambda_{prop}>1 ]
I4 Explainability: E3 commitment requires a reason trace that records selected trajectory, top rejected alternatives, and crossed thresholds.
I5 Accountability: After commitment, outcome deltas must update both map memory and engine model, with residual liability recorded for later replay/audit.

Operationally:

I1 and I3 are computed from HPC rollout comparisons.
I2 is driven by aversive/imminence channels (S3-like) and veto controls ((v_{veto}), MECH-053).
I4 is a commitment precondition.
I5 is a post-commit invariant aligned with responsibility flow and residue preservation.
post-dispatch emergency interruption should be emitted as a superseding commit event from fast safety lanes, with hazard-class and override-scope metadata.

A care-driven override is allowed only when I1-I4 hold. I5 is mandatory after action.

Interpretation

The control plane explains why emotion, arousal, and attention feel like changes in how thinking works rather than changes in belief.

It provides the mechanism by which the Self:

remains coherent in the present,
explores possible futures safely,
commits under uncertainty, and
learns responsibly from consequences.

Cross-references

Trajectory selection and commitment: E3.md
Shared sensory latent and timescales: latent_stack.md
Path memory and replay: hippocampal_systems.md
Default Mode (internal generative mode): default_mode.md
Ethical residue geometry: residue_geometry.md

Open Questions

Q-007 — Universal emotion/expression ↔ control‑channel mapping Do universal‑looking expressions (e.g., victory/pride displays) correspond to stable multi‑channel control regimes in REE, and if so which combinations of arousal, readiness, precision, valence, and social coupling best align with observed universals? This remains an evidence‑constrained hypothesis, not an architectural primitive.

V3 pending (2026-03-15): V2 experiment EXQ-024 FAIL (run twice, consistent). Proximate failure: E3 precision hardcoded at 0.5 — no dynamic precision channel. Fundamental blocker: V2 lacks correct E1/E2/E3 channel separation (ARC-021, MECH-069); z_beta receives mixed signals from conflated learning channels, making any valence-precision correlation uninterpretable even if precision were made dynamic. Correct test requires: (1) z_beta updated by clean E3 harm/goal error, (2) E3 precision read from z_beta dynamically, (3) correlation measured against z_beta valence dimension directly, not health/energy proxy. Blocked until V3 three-channel substrate.

Q-008 — Valence vectors vs μ/κ stability overlays (legacy)
This question is resolved in favor of channel separation: valence remains a dedicated affective appraisal stream, while μ/κ overlays remain stability/commitment modulators. The remaining issue is calibration and orthogonality, not axis replacement. Resolution note: docs/conflicts/resolutions/2026-02-18_valence-vs-mu-kappa.md

Q-010 — Hedonic tone vs valence vs signed PE precision (legacy)
This question has been progressed into MECH‑055 (Affective Channel Separation) and is retained for historical tracking. The remaining open issue is calibration, not separation.

Q-017 — Minimal orthogonal axis set for control-plane completeness
What is the smallest axis subset that still preserves empirically relevant regime separations (for example “focused-but-stuck”, “high-alert without commitment”, and “trust-open vs boundary-defensive”) without re-collapsing control into a single scalar?

ARC-005
ARC-003
ARC-004
INV-008
INV-009
INV-014
MECH-001
MECH-019
MECH-039
MECH-040
MECH-046
MECH-047
MECH-048
MECH-053
MECH-054
MECH-055
MECH-063
MECH-036
MECH-051
MECH-052
MECH-042
Q-007
Q-008
Q-009
Q-010
Q-017

References / Source Fragments

docs/processed/legacy_tree/docs/architecture/control_plane.md
docs/processed/legacy_tree/architecture/control_plane.md
docs/thoughts/2026-02-08_control_plane_modes_responsibility_flow.md
docs/thoughts/2026-02-11_amygdala.md
docs/thoughts/2026-02-11_opioid_receptors.md
docs/thoughts/2026-02-11_habenula_signed_pe.md
docs/thoughts/2026-02-15_basal_ganglia.md
docs/thoughts/2026-02-15_basal_ganglia_commit_gating_control_plane_axes.md