Overview

Claim Type: implementation_note
Scope: Architecture folder overview (legacy)
Depends On: None
Status: legacy
Claim ID: IMPL-015

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

Reflective Ethical Engine — Architectural Overview

The Reflective Ethical Engine (REE) is a cognitive architecture organised around a simple but demanding claim:

Experience, selfhood, and ethical action are not automatic.
They are constructed, conditionally, through commitment under uncertainty.

REE makes this construction explicit.

1. The core idea

REE separates cognition into three irreducible functions:

  • Persistence across time (holding a world and a self)
  • Reachability in the near future (what can be done next)
  • Commitment in the present (what is lived as “now”)

Interpretation note:

  • This decomposition is an analysis lens and typed-boundary contract.
  • It is not an ontological claim that functions are fully isolated in biology or implementation.
  • REE remains a coherent single cognifold with bounded coupling across these functions.

These functions are implemented by three major components:

  • E1 — Persistent Predictive Substrate
  • E2 — Fast Forward Predictor
  • E3 — Trajectory Selection and Commitment Engine

A fourth system — precision control — governs when and how these components engage.

2. E1: Persistence and coherence

E1 is the deep, slow predictive core.

It:

  • maintains long-horizon models of world, self, and value,
  • supports abstraction and reuse across domains,
  • tolerates ambiguity without forcing resolution,
  • persists when attention and action drop away.

E1 does not act and does not construct the present.

It is what allows identity, ethics, memory, and care to exist across time.

(see E1.md)

3. E2: Reachability and affordance

E2 is the fast predictive stream.

It:

  • performs low-latency forward inference,
  • predicts near-future sensory flow,
  • maps action-conditioned outcomes,
  • constructs an affordance surface of immediate possibilities.

E2 keeps the system responsive and grounded while E1 remains slow and abstract.

E2 proposes options; it does not decide.

Explicit multi-step rollouts are hippocampal; E1/E2 supply constraints and short-horizon predictions. Terminology: rollout = explicit hippocampal multi-step sequence; forward prediction = E2 local transition kernel.

(see E2.md)

4. E3: Commitment and the present

E3 selects and stabilises trajectories through a temporally displaced latent space.

It:

  • evaluates candidate trajectories for coherence, feasibility, and ethical admissibility,
  • conditionally commits to one trajectory,
  • phase-aligns prediction with action,
  • collapses a smeared future into a lived present (“now”).

Crucially, E3 operates under licence: global viability and attention signals determine whether a present is constructed at all.

Without E3, the system predicts. With E3, the system lives.

(see E3.md)

5. Precision control: governing engagement

Precision systems do not merely tune learning rates.

They:

  • maintain the viability of predictors,
  • regulate confidence, attention, and uncertainty tolerance,
  • decide whether temporal collapse occurs,
  • shape cognitive regime and phenomenology.

Shifts in precision explain:

  • focus and flow,
  • exploration and imagination,
  • panic, dissociation, or apathy,
  • over-commitment and rigidity.

Precision is the silent governor of experience.

(see precision_control.md)

6. Latent stack and temporal displacement

REE operates on a temporally displaced latent stack.

Predictions are not about “now”, but about multiple future offsets across depths:

  • sensory,
  • affordance,
  • narrative/workspace,
  • regime/value.

A unitary present is not stored anywhere. It is constructed downstream by E3 when commitment occurs.

This preserves perceptual corrigibility while allowing deep value to shape action.

(see latent_stack.md)

7. Default Mode and imagination

When commitment is withheld, REE enters Default Mode.

In this mode:

  • temporally displaced trajectories are explored,
  • hippocampal replay samples possible paths,
  • precision is reduced,
  • no unitary present is constructed.

Default Mode supports imagination, planning, empathy, and ethical reflection without action.

(see Default_mode.md)

8. Ethics as architecture

In REE, ethics is not an add-on.

Ethical constraint enters via:

  • residue geometry (cost of moral violation),
  • deep value shaping of E1,
  • admissibility checks during E3 selection.

This ensures that:

  • some trajectories are never selected,
  • others leave lasting residue,
  • and care is structurally enforced rather than optimised away.

(see residue_geometry.md)

9. Putting it together: the REE loop

A simplified loop:

  1. E1 maintains deep predictive structure
  2. E2 generates near-term affordances
  3. Default Mode may explore without commitment
  4. E3 evaluates trajectories
  5. Precision licenses or withholds collapse
  6. Action is taken (or not)
  7. Error and residue update E1 over time

This loop can pause, resume, or refuse to collapse — without loss of self.

10. Design commitments

REE is built on several explicit commitments:

  • Selfhood is constructed, not assumed
  • Experience is conditional, not continuous
  • Ethics is structural, not instrumental
  • Pathology is a regime failure, not a bug
  • Reuse requires decoupling from action
  • Care requires the ability to wait

11. Summary

REE is an architecture for agents that:

  • can predict without acting,
  • imagine without collapsing,
  • commit without losing corrigibility,
  • and care without optimisation erasure.

It is not designed to be fast. It is designed to be viable. —

Open Questions

None noted in preserved sources.

  • IMPL-015

References / Source Fragments

  • docs/processed/legacy_tree/architecture/README_architecture.md

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