Unclassified

05 · Projections

The sole composition layer — shaping canonical Spine data into role-appropriate views without duplicating or leaking canonical truth

The sole composition layer.

Projections shape canonical Spine data into role-appropriate views. They are the only mechanism through which the Spine is exposed to users and AI. Nothing reads from the Spine directly — everything goes through Projections.

Why Projections Exist

The Spine contains canonical truth. But different roles need different views of that truth:

A Sales Manager needs pipeline data, not engineering metrics
A Finance Analyst needs revenue data, not support tickets
A Twin needs context for its user's role, not every fact in the Spine
A Domain Agent needs domain-specific data, not cross-domain noise

Projections solve this by shaping the same canonical truth into role-appropriate views.

The Projection Model

Canonical Spine
      ↓
Projection Engine
      ↓
Role / Domain / Authorization Filter
      ↓
Context Assembly
      ↓
Operational / Sandbox Context
      ↓
Twin / Agents / Workbench

Key Invariant

AI receives authorized projections, not unrestricted canonical organizational truth.

This is one of the strongest architectural invariants:

                    SPINE
                     │
          ┌──────────┴──────────┐
          │                     │
    CANONICAL LAYER       OPERATIONAL LAYER
          │                     │
     Organization             Sandbox
     Truth                    Context
     Memory                   Working State
     Governance               Agent Data
          │                     │
       HUMAN                 AI / AGENTS

Projection Types

Type	Description
Role Projection	Shaped by the user's role (Sales, Finance, etc.)
Domain Projection	Shaped by the business domain
Work Projection	Shaped by the current work item
Time Projection	Shaped by temporal relevance
Authorization Projection	Filtered by permissions

Projection Contract

interface Projection {
  id: string;
  tenantId: string;
  source: 'spine';                    // always from the Spine
  target: ProjectionTarget;           // who/what receives this
  scope: ProjectionScope;             // what data is included
  filters: FilterDef[];               // data filters
  transforms: TransformDef[];         // data transformations
  cacheability: CachePolicy;          // how long this is valid
  governanceToken?: string;           // governance approval
}

Projection Rules

Rule	Description
Canonical read-only	Projections read from the Spine, never write
No duplication	Projections do not copy canonical data
Role-scoped	Each projection is scoped to a role
Authorization-gated	Projections respect RBAC
Audit-logged	Every projection is logged
Cacheable	Projections can be cached for performance

Projections and AI Security

The Projection layer is the security boundary between canonical truth and AI:

Canonical Spine (organization truth, memory, governance)
      ↓
Projection (role + authorization filter)
      ↓
Context Assembly (relevance scoring, token budgeting)
      ↓
Operational / Sandbox Context (AI-safe data)
      ↓
Twin / Agents (operate on safe data only)

This ensures:

Twins never see unauthorized data
Agents operate within their scope
Context is always role-appropriate
Canonical truth is never leaked

Projections Connect to

Layer	Relationship
Spine Fabric	Reads canonical data from the Spine
Human Workbench	Provides role-specific views
Twin Workbench	Provides sandbox context for AI
Coworker	Provides context for ambient awareness
Domain Projections	Domain-specific shaping