E3 — Trajectory Selection and Commitment Engine

Claim Type: architectural_commitment Scope: Trajectory selection, commitment, and temporal collapse mechanism Depends On: INV-012 (commitment gates responsibility), E1, E2, L-space, residue geometry, precision control Status: stable Claim ID: ARC-003 Related Claims: MECH-125 (multi-constraint viability selection criterion)

E3 is the component of the Reflective Ethical Engine (REE) responsible for selecting, stabilising, and committing a trajectory through a temporally displaced latent space.
It is the mechanism by which prediction becomes lived experience.

E3 does not select instantaneous actions or perceptual states.
It selects trajectories and determines whether a unitary present (“now”) is constructed at all.

Subsystem abstract (core claims): ARC‑003 defines trajectory selection and commitment. Supporting context includes INV‑012 (commitment gates responsibility), ARC‑004 (latent stack), ARC‑005 (precision control), ARC‑012 (legacy ethical term note), ARC‑018 (rollout source for candidates), MECH‑060 (dual pre/post-commit error channels), and MECH‑061/MECH‑062 (commit-boundary token + tri-loop commitment gating).

1. Architectural role

E3 sits downstream of predictive generation (E1, E2) and upstream of action, memory consolidation, and subjective experience.

Its core functions are:

Selection of a coherent trajectory across predictive depths
Conditional commitment to that trajectory
Temporal collapse of displaced predictions into an experienced present
Enforcement of perceptual corrigibility and ethical constraint

E3 operates only when licensed by global viability and attention signals (see Precision Control).

1a. Consolidation principle anchor

E3 — specifically its tri-loop commit gating (MECH-062) — instantiates the consolidation step identified in compact modelling of macaque V4 (Cowley et al. 2023): selective pooling + competition + normalisation over a shared representational basis (E1).

Selectivity of trajectory preference lives here, not in E1. Behavioural identity — which trajectories are preferred, which values dominate, which harm costs are weighted — is determined by E3’s consolidation operator, not by the E1 feature basis.

This has a direct experimental prediction (CSH-1): varying E3 consolidation weights (λ, ρ in J(ζ)) should produce systematically different trajectory preference in a fixed environment, while E1 latent representations remain stable.

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

2. Inputs and outputs (interface contract)

Inputs

E3 receives:

A temporally displaced latent bundle
( \mathcal{B}t = { \hat z_k(t+\Delta) }{k \in K} )
(see latent_stack.md)
A precision profile
( \alpha = { \alpha_k }_{k \in K} )
(see precision_control.md)
A residue / ethical curvature field
( \mathcal{R}(\cdot) )
(see residue_geometry.md)
Feasibility and invariant constraints
(dynamics, safety, care budget)
Global affective / viability signals
(reward, harm, uncertainty tolerance, kin relevance)

Outputs

E3 emits:

A selected trajectory ( \tau^* ) (or null)
A commitment signal
Precision escalation or suppression directives
Action gating / policy parameters
Intended-action labels for self-impact attribution (feeds post-action comparison)
Memory write-enable signals

3. What E3 selects

A trajectory ( \tau ) is a path through the displaced latent bundle over a horizon (H):

[ \tau = \big( a_{t:t+H}, \hat z_{K,\,t:t+H}, \hat o_{t:t+H} \big) ]

Trajectories may be generated by:

fast predictor forward predictions (affordance-level kernels)
deep predictor forward predictions (narrative / workspace constraints)
hippocampal replay and braid mechanisms

Canonical clarification (2026-02-09): explicit multi-step rollouts are generated by hippocampal systems. References to fast/deep rollouts should be read as inputs and constraints from E2/E1 that seed hippocampal trajectory generation, not separate rollout engines. E3 does not imagine; it selects and commits.

3a. Task-loop extraction object contract (implementation-facing)

E3 can also be specified as an extraction/re-entry loop: pull a bounded task object from continuous embodied policy traffic, evaluate and edit it in decoupled selection space, and only re-enter actuator paths through commitment gates.

Recommended minimal extracted object contract:

object_token: stable anchor/index for the current task object.
valence_vec: signed approach/avoid and urgency profile.
transition_ops: candidate state-transition operators or option fragments.
error_vec: mismatch direction with confidence/precision tags.
stop_op: explicit pause/inhibit/terminate primitive.
time_env: onset, expected duration, and deadline window.
identity_tag: self-generated vs externally imposed ownership/provenance.
path_graph (optional): relational sequence scaffold for planning and simulation.

Suggested payload sketch:

TaskLoopObject = {object_token, valence_vec, transition_ops, error_vec, stop_op, time_env, identity_tag, path_graph?}

Contract note:

this is a representation contract for MECH-062 tri-loop arbitration and MECH-061 boundary tokenization, not a separate module.
it strengthens MECH-060 write-locus separation by keeping extracted-loop edits pre-commit unless/until boundary crossing is explicitly licensed.

3b. Neuro-inspired primitive mapping (non-isomorphic)

The extracted object contract above can be mapped to a neuro-inspired scaffold for implementation naming and interface partitioning. This is a design aid, not a one-to-one anatomical claim.

object_token: frontoparietal indexing/maintenance loops (DLPFC/PPC) under corticostriatal gating.
valence_vec: OFC/vmPFC/amygdala/ventral-striatal integration for signed approach-avoid weighting.
transition_ops: premotor/SMA sequencing with cerebellar forward-model correction and dorsal-striatal chunking.
error_vec + precision tags: ACC/insula conflict-mismatch channels with dopaminergic and noradrenergic gain signals.
stop_op: hyperdirect-style interrupt lane (rIFG/pre-SMA -> STN -> basal-ganglia output suppression).
time_env: DLPFC planning horizon with basal-ganglia/SMA/cerebellar timing support and hippocampal temporal context.
identity_tag: mPFC/TPJ/PCC self-other attribution and ownership binding.
path_graph: hippocampal-entorhinal relational scaffold constrained by prefrontal selection.

Implementation constraint:

preserve typed interfaces between these primitives so stoppage, valuation, and path-structure edits remain independently inspectable during tri-loop arbitration.

4. Selection functional (symbolic)

E3 selects the trajectory that minimises a composite energy functional:

[ \tau^* = \arg\min_{\tau \in \Omega} \Big[ \lambda_C\,\mathcal{C}(\tau;\alpha)

\lambda_R\,\mathcal{R}(\tau)
\lambda_F\,\mathcal{F}(\tau)
\lambda_I\,\mathcal{I}(\tau)
\lambda_P\,\mathcal{P}(\tau) \Big] ]

Where:

( \mathcal{C}(\tau;\alpha) ) — cross-horizon coherence cost
( \mathcal{R}(\tau) ) — residue / ethical curvature cost
( \mathcal{F}(\tau) ) — feasibility and control cost
( \mathcal{I}(\tau) ) — epistemic value (exploration benefit)
( \mathcal{P}(\tau) ) — phase-lock / temporal collapse penalty

The coefficients ( \lambda ) are modulated by precision and affective state.

4a. Selection criterion: multi-constraint viability, not reward maximisation

The selection functional above is a composite energy minimisation over weighted terms. The correct interpretation of what E3 is doing is not “find the highest-reward trajectory” but:

Choose the trajectory that fails the fewest constraints across the evaluation spaces — and makes acceptable progress toward the goal.

Formally: a trajectory is selected when no major evaluation system produces a veto-level error signal across goal space, harm space, identity space, and resource space. The decision rule is lowest-veto + acceptable goal progress, not maximum reward.

This distinction matters because:

A trajectory can have high immediate reward while failing harm, identity, or resource constraints, producing internal contradiction (“this feels wrong” even when reward is high).
Standard reward maximisation cannot represent the constraint structure; viability maximisation can.

Biological correlate: the commit gate (basal ganglia) opens when ACC conflict, insula cost, amygdala threat, goal-distance (MECH-112), and identity checks all fall below veto threshold simultaneously. No single system “chooses” — commitment emerges from absence of sustained veto.

The five terms in J(tau) above map to constraint spaces:

C(tau) – cross-horizon coherence: identity and prediction consistency
R(tau) – residue/ethical curvature: harm and moral cost
F(tau) – feasibility: resource and physical constraint
I(tau) – epistemic value: exploration benefit (negative constraint)
P(tau) – phase-lock: temporal coherence

Claim: MECH-125 (coherence.multiconstraint_viability) – see claims.yaml. Source: docs/thoughts/2026-03-24_COHERENCE_MULTICONSTRAINT_HIPPOCAMPAL_NAVIGATION_CONVERGENCE.md

4b. Ethical Cost Term (Legacy vs Current)

Claim ID: ARC-012

Current canonical framing: E3 selection does not require an explicit ethical cost term. Ethical consequence is encoded via residue curvature, mirror modelling, and commitment-gated learning rather than a separate moral module or cost term.

Legacy formulations included an explicit ethical cost term (M) in trajectory scoring. That wording is preserved in processed sources and is no longer canonical.

Source: docs/processed/legacy_tree/REE_CORE.md

5. Coherence and depth interaction

Coherence is evaluated both:

Within depth: prediction error weighted by precision
Across depth: agreement between projected representations at different horizons

Deep regime predictors (reward, threat, attachment) bias selection via precision and scoring, not semantic overwrite, preserving perceptual corrigibility.

6. Commitment and temporal collapse

E3 operates in two regimes:

Exploratory regime

Multiple trajectories evaluated
Low action-depth precision
No phase-locking to motor execution
No unitary present constructed
(see Default_mode.md)

Committed regime

A single trajectory is stabilised
Action-depth precision is increased
Predictions are phase-aligned with motor timing
A unitary present (“now”) is constructed

Commitment is triggered when a trajectory is:

sufficiently coherent,
ethically admissible,
feasible,
and licensed by global viability signals.

Commit-Boundary Token and Error Reclassification (MECH-061)

Claim Type: mechanism_hypothesis
Scope: Explicit boundary token that separates pre-commit and post-commit error classes
Depends On: ARC-003, MECH-060, INV-012, ARC-015
Status: candidate
Claim ID: MECH-061

E3 should emit an explicit commit-boundary token at threshold crossing. The token is the minimal routing primitive that reclassifies errors from counterfactual evaluation into attributable post-commit feedback.

Minimal token fields (v1 draft):

commit_id (unique id for the commitment event)
trajectory_id (selected trajectory)
trajectory_hash (stable digest of selected trajectory materialization)
issued_at_step (time index at threshold crossing)
ttl_steps (how long downstream attribution may bind to this commit)
mode_snapshot (control-plane mode at commit time)
intended_action_labels (declared intervention labels for attribution joins)
authority_path (resolved authority chain used for this commit)
policy_hash (policy/invariant store digest at commit)
caps_hash (capability/permission manifest digest at commit)
verifier_result_hash (digest of verifier artifact set)
uncertainty_snapshot (compressed uncertainty calibration state at commit boundary)
supersedes_commit_id (optional; set when this commit explicitly interrupts/supersedes a previously dispatched commit)
supersession_reason (optional; required when supersedes_commit_id is present)
hazard_class (optional; required when supersedes_commit_id is present)
override_scope (optional; required when supersedes_commit_id is present)
reflex_lane_id (optional; identifies fast safety/interrupt source lane when supersession is reflex-driven)

Routing rule:

before token issuance: errors are pre-commit (counterfactual, coherence, effort/conflict, anticipated value)
after token issuance (within ttl_steps): errors are post-commit (reafference mismatch, realized outcome, credit assignment)

Operational interpretation:

pre-commit and post-commit channels can run continuously in parallel; this is not a stop-and-wait phase machine.
token issuance marks an authority transition (irreversible dispatch / privileged durable-write eligibility), not scheduler quiescence.
internal rehearsal/selection dynamics may remain stochastic; deterministic requirements attach at the boundary where action/write projection becomes responsibility-bearing.
fast-path interruption before irreversible dispatch remains pre-commit arbitration.
interruption after irreversible dispatch must mint a new commit token with supersedes_commit_id; realized outcomes remain attributable to the original commit lineage.

Supersession constraints:

post-dispatch supersession is expected to be safety-reflex dominated (fast aversive/veto lanes) rather than preference-driven replanning.
when supersedes_commit_id is set, supersession_reason, hazard_class, and override_scope must be present.
override_scope must be mitigation-bounded (for example: halt, brake, evade, contain, safe-mode transition), not a blanket permission for unrelated durable policy rewrites.
supersession extends lineage history; it does not invalidate the accountability record of the superseded commit.

This preserves INV-012 by making responsibility arise at commitment, not at simulation time.

Tri-Loop Commitment Gating (MECH-062)

Claim Type: mechanism_hypothesis
Scope: Three commitment gates (motor, cognitive-set, motivational) with shared thresholding and coordinated arbitration
Depends On: ARC-003, ARC-005, MECH-061
Status: stable
Claim ID: MECH-062

E3 should expose three coupled commitment gates rather than one global gate:

gate_motor (movement/action release)
gate_cognitive_set (task-set/update release)
gate_motivational (salience/drive release)

The thresholding algorithm is shared, but each gate consumes a different proposal manifold and can be modulated independently by the control plane. This prevents pathological global coupling where one suppressed stream forces all streams into suppression.

Loop policy contract:

gate_motor:
- commit criterion prioritizes trajectory feasibility/integrity.
- strongest veto power when invariant risk is high.
gate_cognitive_set:
- commit criterion prioritizes task-set coherence and switching cost.
- can defer commitment while preserving rehearsal breadth.
gate_motivational:
- commit criterion prioritizes salience/value lane selection and persistence.
- may pre-filter proposal bands before full trajectory expansion.

Trajectory-first is therefore loop-specific: strict for motor enactment, softer for cognitive-set updates, and band-limited for motivational gating.

Coordination invariant:

conflict across gates must resolve through explicit arbitration policy and be recorded against a commit_id, rather than via silent dominance of one gate.

Candidate arbitration policy: disinhibitory sweep check

One concrete policy family to test is a disinhibitory sweep:

a provisional winner lane requests release (gate_* crossing),
alternate lanes are briefly disinhibited for bounded eligibility checks,
higher-priority veto/authority checks can still block release,
commitment proceeds only when sweep checks preserve global eligibility.

This keeps lane specialization while still forcing cross-lane consistency checks before commit issuance. All sweep diagnostics should be emitted under the same commit_id and routed into Q-016 arbitration evidence.

Learning/update invariant:

pre-commit gate activity may tune lane-local thresholds only.
durable cross-lane policy or ledger writes require commit-boundary transition (MECH-061/060).

Overlay re-engagement and trigger-field routing (implementation guard)

Tri-loop control should treat conscious/reasoned re-engagement as an overlay mode by default, not an automatic full takeover of all execution loops. Recommended control states:

EXECUTE_BASE: post-commit executor dominates with low-latency micro-corrections.
OVERLAY_AUDIT: E3 injects targeted constraints/re-weights while base execution continues.
FULL_REPLAN: full takeover and replanning, reserved for unresolved conflict or high-risk mismatch.

Triggering should be modeled as a competitive field, not a single interrupt bit:

candidate triggers include sensory mismatch, goal conflict, hazard, social-rule conflict, salience spikes, fatigue, and novelty.
bounded tactical bandwidth means only a small top-priority set enters E3 arbitration at once.
control-plane axes (MECH-063) tune trigger gains, refractory periods, and escalation thresholds.

This preserves smooth low-level execution during routine corrections while still allowing strong interruption and replanning when conflict pressure exceeds guardrails.

Tri-loop gating should be implemented as ordered eligibility filters, not as a scalar vote among lanes. Recommended dominance order:

hard veto / fast threat interrupt (HARM, catastrophic risk),
authority and reality-coherence eligibility (verifier + RC_conflict posture),
contextual anchoring / provenance validity,
embodiment/homeostasis feasibility,
motivational sufficiency and delay-cost posture,
motor release (gate_motor) with commit-token issuance.

This keeps “can execute” and “may execute” distinct, and prevents cases where motivational pressure outvotes authority or safety constraints.

Hierarchy clarification:

this layered-filter structure is the pre-commit realization layer for MECH-060, not a superseding replacement.
MECH-060 remains the umbrella two-channel boundary claim; MECH-062 specifies how pre-commit eligibility is enforced.

Evolutionary-constraint note

The disinhibitory commit pattern (default suppression, release-on-selection, opponent promotion/suppression channels) is treated as a robust computational prior for REE commitment, with cortical/hippocampal systems enriching candidate generation rather than replacing the gating core.

Implementation Gaps (Open, from 2026-02-15 refresh)

The commit-boundary and tri-loop framing is now canonical, but the following implementation contracts remain open and should be treated as explicit engineering gaps:

Commit-token wire format and routing scope are underspecified beyond the v1 field draft.
Cross-gate conflict arbitration policy is underspecified when motor, cognitive-set, and motivational gates disagree.
Post-commit credit-assignment interface is underspecified (mask types, causal assumptions, and update timescales).
Responsibility locus storage is underspecified (attribution ledger vs training-signal router vs stable self-model).

These gaps should be resolved in implementation-facing specs to keep MECH-061/MECH-062 operationally crisp and falsifiable under stress.

7. Precision as control

Precision does not merely tune learning rates.
It controls whether temporal collapse occurs at all.

High action-depth precision → lived present
Balanced precision → sustained exploration
Misallocated precision → fragmentation or pathological collapse

(see precision_control.md)

8. Failure regimes (E3 pathologies)

Misconfiguration of E3 yields characteristic failure modes:

Over-commitment
Excessive temporal collapse → mania, compulsivity
Under-commitment
Failure to collapse → indecision, apathy
Fragmentation
Competing trajectories → panic, dissociation
False coherence
Deep regime overwrite → psychosis-like states

These are architectural regime failures, not simple parameter errors.

9. Summary

E3 is the engine that turns prediction into experience.

It:

selects trajectories, not moments,
constructs the present conditionally,
enforces ethical and perceptual constraint,
and grounds selfhood in the sustained act of commitment under uncertainty.

Without E3, REE predicts. With E3, REE lives.

Open Questions

Q-015: What is the smallest commit-boundary token that still supports reliable multi-timescale attribution?

Open design dimensions: token scope, TTL policy, and cross-module broadcast rules.
Current hard floor candidate fields: commit_id, trajectory_hash, authority_path, policy_hash, caps_hash, verifier_result_hash, uncertainty_snapshot.
Anchor: MECH-061.

Q-016: What arbitration policy best resolves cross-gate conflicts (motor vs cognitive-set vs motivational) without coupling collapse?

Candidate families: veto lattice, weighted arbitration, or mode-dependent precedence.
Additional candidate: disinhibitory sweep check (provisional winner with bounded alternate-lane eligibility checks).
Anchor: MECH-062.

ARC-003
ARC-012
ARC-001
ARC-002
ARC-004
ARC-005
INV-012
ARC-018
ARC-015
MECH-060
MECH-061
MECH-062
MECH-125
Q-015
Q-016

References / Source Fragments

docs/processed/legacy_tree/docs/architecture/e3.md
docs/processed/legacy_tree/architecture/E3.md
docs/thoughts/2026-02-15_basal_ganglia.md
docs/thoughts/2026-02-15_basal_ganglia_commit_gating_control_plane_axes.md
docs/thoughts/2026-02-18_commitment_as_layered_eligibility_filters.md
docs/thoughts/2026-02-19_basal_ganglia_evolutionary_conservation_pull.md
docs/thoughts/2026-02-24_prefrontal_primitives.md
docs/thoughts/2026-02-25_task_loop_extraction_and_latent_field_ethics.md
docs/thoughts/2026-03-24_COHERENCE_MULTICONSTRAINT_HIPPOCAMPAL_NAVIGATION_CONVERGENCE.md