Skip to main content

Blueprints

Reference architectures, full operating models for a capability, not a one-off essay. One G.A.I.N subject maps to one playbook series. Governance has two blueprints (operating and runtime) because the audience splits; it is still one G.A.I.N subject.

1. Evaluation: eight planes

G.A.I.N EvaluationPrinciples, ownership, and pillar mapping: published
Eval BlueprintSources, scorers, modes, gates, eight planes: draft

Playbooks: Evaluation. Planes: ① Input · ② Data · ③ Context · ④ Reasoning · ⑤ Tool · ⑥ Memory · ⑦ Action · ⑧ Outcome.

2. Agents: intent, orchestration, manifests

Intent at ingress, orchestration in the loop, versioned manifests (pinned tools, schemas only), four autonomy patterns on the route. Model endpoint selection is the LLM Blueprint. PEP is Governance. MCP is transport only.

G.A.I.N AgentsPrinciples, policy gate, orchestration: published
Agents BlueprintIntent, orchestration, manifests, Pattern 0-3: draft

Playbooks: Agents (Plane ① Intent router · Plane ② Orchestration including Memory · Manifests). Patterns: 0 Single inference · 1 Autonomous loop · 2 Deterministic · 3 Guided hybrid.

3. LLM: gateway model routing

G.A.I.N LLMGateway stack principles: published
LLM BlueprintCapability matrix, task routing, canary: draft

Playbooks: LLM. Handoff from agents: Inference handoff.

4. Governance: operating model and runtime

Two blueprints, one G.A.I.N subject. Operating is the estate control system. Runtime is PGAR on the request path. PGAR gates side effects. It does not own RAG corpora, tool manifests, or the memory store.

G.A.I.N GovernanceControl system: policy, inventory, runtime, evidence: published
GovernanceTwo views, one subject: draft
OperatingFour questions, one control system: draft
RuntimeSARAC contracts, five boundaries, release gates: draft

Playbooks: Governance (Operating · Runtime). Boundaries: ① Ingress · ② Agentic app · ③ LLM proposal · ④ PEP + PDP · ⑤ Downstream.

5. Observability: three layers

G.A.I.N ObservabilityAI capture path, retention tiers, audit: published
Observability BlueprintThree layers, graph, maturity, reference architecture: published

Playbooks: Observability.

6. RAG: governed retrieve

G.A.I.N RAGGrounded retrieve, validation, attribution: published
RAG BlueprintMode, scope, PEP pack, validate then synthesize: draft

Playbooks: RAG. Eval: Context plane.

7. MCP: transport

G.A.I.N MCPTool transport, registry, policy gate: published
MCP BlueprintGateway, server registry, adapter (not the catalog): draft

Playbooks: MCP (transport). Tool contracts: Agents: Manifests. Eval: Tool plane.