159c783c2e
- docs/decisions/ADR-0008-allocation-physics.md Formalizes salience, attention, inhibition, and coherence-budget as the allocation physics of cognition. Replaces attention-as-weights with attention-as-field-curvature over the versor manifold. - docs/decisions/ADR-0009-compositional-physics.md Defines temporal binding, digest cycles, reasoning trajectories, and articulation planning as the compositional physics layer — how CORE assembles pressure into structured thought and output. - docs/decisions/ADR-0010-identity-physics.md Establishes IdentityManifold, DriveGradientMap, ExertionMeter, and CharacterProfile as structural identity primitives. Identity is a field over the geometry, not a prompt veneer. Grounded in John 1:1–2 and the Logos theology that anchors the architecture. - docs/architecture/MIND-PHYSICS-BLUEPRINT.md Integration blueprint showing how allocation → compositional → identity physics layers compose into the full cognitive cycle. - core/physics/ (11 Python interface stubs) SalienceOperator, AttentionOperator, InhibitionOperator, BindingFrame, DigestCycle, ReasoningTrajectory, ArticulationPlanner, DriveGradientMap, ExertionMeter, IdentityManifold, CharacterProfile — all typed, all frozen where stateless, all carrying explicit field contracts. Third Door: no off-the-shelf cognitive architecture borrowed. All operators defined from the geometry up.
109 lines
5.8 KiB
Markdown
109 lines
5.8 KiB
Markdown
# ADR-0008 — Allocation Physics
|
|
|
|
**Status:** Accepted
|
|
**Date:** 2026-05-12
|
|
**Deciders:** Joshua Shay (Architect)
|
|
**Supersedes:** None
|
|
**Related:** ADR-0001 (Field-State), ADR-0003 (Versor Injection Gate), ADR-0007 (Ingest Layer)
|
|
|
|
---
|
|
|
|
## Context
|
|
|
|
The versor field carries active pressure. Not all pressure is equally relevant to a given cognitive task at a given moment. The architecture needs a principled mechanism for allocating cognitive resources — determining which pressure regions are foregrounded, which are suppressed, and how coherence budget is spent.
|
|
|
|
The naive solution is attention-as-weights (the transformer pattern): a learned matrix projects queries against keys and returns a weighted sum over values. This is rejected on three grounds:
|
|
|
|
1. **It conflates geometry with bookkeeping.** Dot-product attention has no geometric meaning in the versor manifold. It operates on flattened token embeddings, not on structured field regions.
|
|
2. **It is opaque to correction.** When attention produces wrong salience, there is no dual-correction path — no conjugate operator that can restore coherence. The weights simply update at the next gradient step.
|
|
3. **It violates Semantic Rigor.** A learned weight matrix does not know *why* it attends to something. The salience of a claim should derive from its structural relationship to the active context, not from statistical co-occurrence in pretraining data.
|
|
|
|
This ADR defines the allocation physics layer: a set of operators that govern foregrounding, suppression, and budget within the versor field.
|
|
|
|
---
|
|
|
|
## Decision
|
|
|
|
### 1. Salience as Field Curvature
|
|
|
|
Salience is not a scalar score attached to a token. It is a **curvature property of the versor field** at a given region. A pressure region is salient when it causes measurable deflection in the trajectories of neighboring regions — when it bends the field around itself.
|
|
|
|
The `SalienceOperator` computes a local curvature estimate over a `FieldRegion` and returns a `SalienceMap`: a structured record mapping region identifiers to curvature magnitudes and directional vectors.
|
|
|
|
```
|
|
SalienceOperator: FieldRegion → SalienceMap
|
|
```
|
|
|
|
This is geometry-first allocation. Salience derives from structure, not from a learned score.
|
|
|
|
### 2. Attention as Controlled Field Traversal
|
|
|
|
Attention is the act of **directing cognitive traversal** along high-salience curvature gradients. The `AttentionOperator` takes a `SalienceMap` and a `CoherenceBudget` and returns an `AttentionPlan`: an ordered traversal schedule over field regions, constrained by the budget.
|
|
|
|
```
|
|
AttentionOperator: (SalienceMap, CoherenceBudget) → AttentionPlan
|
|
```
|
|
|
|
The plan is not a weight distribution. It is a **schedule** — a sequence of field regions to activate, with associated depth and duration, that can be inspected, overridden, and corrected.
|
|
|
|
### 3. Inhibition as Dual Correction
|
|
|
|
Every attention plan has a conjugate: an `InhibitionOperator` that suppresses field regions whose activation would reduce coherence. Inhibition is not the absence of attention — it is an active structural force that prevents interference between competing pressure regions.
|
|
|
|
```
|
|
InhibitionOperator: (AttentionPlan, FieldState) → InhibitionMask
|
|
```
|
|
|
|
The mask is applied before traversal begins. This encodes the **dual-correction axiom** directly into the allocation layer: every forward attention plan is paired with a corrective inhibition pass that restores field coherence.
|
|
|
|
### 4. Coherence Budget
|
|
|
|
Cognitive resources are finite. The `CoherenceBudget` is an explicit resource object that tracks:
|
|
|
|
- `total_capacity` — the maximum pressure activation units available in a cycle
|
|
- `committed` — units already allocated to active traversal
|
|
- `reserve` — units held back for inhibition and correction passes
|
|
- `spent` — units consumed in the current cycle
|
|
|
|
Budget is consumed by attention depth and region breadth. Inhibition draws from reserve, not from committed. When budget is exhausted, traversal terminates and the cycle closes.
|
|
|
|
---
|
|
|
|
## Consequences
|
|
|
|
### Positive
|
|
|
|
- Allocation is inspectable and correctable at every step.
|
|
- Salience derives from field geometry, not from learned weights — it generalizes across domains without retraining.
|
|
- The inhibition/attention duality ensures coherence is actively maintained, not assumed.
|
|
- CoherenceBudget makes resource consumption explicit and measurable.
|
|
|
|
### Negative
|
|
|
|
- Computing field curvature is more expensive than dot-product attention in naive implementations. The Rust hot-path (ADR-0003) must cover the curvature kernel.
|
|
- SalienceMap construction requires a populated FieldState — allocation physics cannot run on an empty field.
|
|
|
|
### Neutral
|
|
|
|
- This layer replaces transformer-style attention entirely. There is no compatibility shim with softmax attention weights. Any external model integration (D3 instruments per ADR-0007) operates above this layer, not within it.
|
|
|
|
---
|
|
|
|
## Alternatives Rejected
|
|
|
|
| Alternative | Reason Rejected |
|
|
|---|---|
|
|
| Transformer dot-product attention | Geometrically meaningless on versor manifold; opaque to correction |
|
|
| Sparse attention (Longformer, BigBird) | Structural improvement on wrong foundation; still no conjugate |
|
|
| Memory-augmented attention (Memorizing Transformer) | External retrieval bolted onto broken base; not field-native |
|
|
| Learned salience scoring (MLP over embeddings) | Violates Semantic Rigor; salience must derive from structure |
|
|
|
|
---
|
|
|
|
## Implementation Notes
|
|
|
|
- `core/physics/salience.py` — `SalienceOperator`, `SalienceMap`, `FieldRegion`
|
|
- `core/physics/attention.py` — `AttentionOperator`, `AttentionPlan`, `CoherenceBudget`
|
|
- `core/physics/inhibition.py` — `InhibitionOperator`, `InhibitionMask`
|
|
- Rust acceleration target: curvature kernel in `core_rs::physics::salience`
|
|
- `SalienceMap` is content-addressed (SHA-256 over region IDs + curvature values) for cache reuse across cycles
|