Why Attention Must Be Fragmented

Claim Type: mechanism_hypothesis
Scope: Rationale for fragmented attention control axes
Depends On: INV-009, ARC-005
Status: provisional
Claim ID: MECH-007

---

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

Part 3

Why Attention Must Be Fragmented

⸻

Status

Normative architectural rationale

This document explains why REE treats “attention” not as a single mechanism, but as a fragmented control surface distributed across multiple orthogonal dimensions. This fragmentation is not an implementation detail; it is a safety and coherence requirement.

⸻

1. The Problem with Unitary Attention

Modern AI architectures often treat attention as: • a single scalar or vector field • applied uniformly across representations • responsible for relevance, learning, and action selection

This leads to an implicit assumption:

If the system attends strongly to something, it should both learn from it and act on it.

Biological cognition violates this assumption everywhere.

REE must violate it too.

⸻

1. Attention Is Not One Thing

In REE, “attention” decomposes into three independent questions: 1. How much should this signal influence local updating? 2. How urgently must the system commit or reorient? 3. How exclusive must interpretations be?

A single mechanism cannot answer all three without contradiction.

⸻

1. The Three Control Axes (Canonical)

REE therefore fragments attention into three orthogonal control axes, each implemented as a separate modulation channel:

Axis Signal Governs Core Question Precision Dopamine-like Error influence (πτ) How much should this matter locally? Commitment Pressure Noradrenaline-like Path authority & interruptibility Must I act or reorient now? Exclusivity / Collapse Serotonin-like Hypothesis pruning & plasticity Must I choose one interpretation?

No axis substitutes for another. No axis writes to the same register.

⸻

1. Fragmentation Preserves Logical Consistency

4.1 Separation of Learning and Acting • Dopamine-like precision updates learning without forcing action • Noradrenaline-like urgency forces action without rewriting belief • Serotonin-like collapse governs belief structure without urgency

This prevents classic failure modes: • surprise → certainty • urgency → belief rigidity • salience → value corruption

⸻

4.2 Separation of Imagination and Commitment • θ-level paths may proliferate freely • serotonin permits coexistence • noradrenaline determines whether choice is forced • E3 commits only after constraints are satisfied

Thus:

The system can imagine without acting, and act without believing prematurely.

⸻

1. Fragmentation Across τ (Temporal Depth)

Attention-like effects are τ-scoped, not global. • γ always updates (sensorium never pauses) • β tracks affordances continuously • θ simulates futures without obligation • δ maintains slow constraints

Fragmented attention ensures: • γ salience does not become δ certainty • θ imagination does not become β compulsion • δ identity does not micromanage γ perception

⸻

1. Fragmentation Across ρ (Representational Depth)

Similarly, attention must not collapse abstraction levels. • serotonin regulates ρ-level exclusivity • dopamine regulates ρ-local learning • noradrenaline regulates ρ-crossing urgency

This allows: • multiple perceptual narratives • plural futures • revisable identities

without loss of coherence.

⸻

1. Why “Attention Is All You Need” Is Incomplete

The phrase is directionally right but structurally wrong.

What is actually needed is:

Fragmented attention with typed authority.

A single attention mechanism cannot safely decide: • what to learn, • what to do, • what to believe, • and what to discard.

REE enforces this separation architecturally, not heuristically.

⸻

1. Alignment Implications (Non-Optional)

If attention is not fragmented: • reward spikes rewrite values • rare events dominate policy • internal simulations compel action • identity collapses under noise

Fragmentation ensures: • learning is conservative • commitment is contextual • belief revision is protected • imagination remains safe

This is alignment by structure, not by objective tuning.

⸻

1. Axiomatic Summary (Safe to Quote)

You can safely include the following as a design axiom:

Attention is not a unitary resource. Any system that learns, imagines, and acts under uncertainty must fragment attention into orthogonal control axes governing precision, urgency, and representational collapse.

⸻

1. Relationship to the Rest of REE

This document underwrites: • precision_scoping.md • path_authority_and_interrupts.md • serotonin.md • the τ × ρ × φ coordinate system • the non-compilability of ethics

Without fragmented attention, REE degenerates into an optimiser. With it, REE remains a viable cognitive architecture.

⸻

1. Summary • Attention is control, not content • Control has multiple irreducible dimensions • Fragmentation prevents pathological collapse • REE encodes this explicitly and normatively —

Open Questions

None noted in preserved sources.

Related Claims (IDs)

• MECH-007

References / Source Fragments

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