Chapter 35. The Reference Pattern Library

Ten pre-approved patterns, what each governs, and how the library earns its keep.

35.1 What a pattern is for

A reference pattern is a pre-approved architecture that a Tier 1–2 system may adopt unmodified in order to use the largely automated G2 path. That is the library’s governance function, and it is the framework’s principal genuine reduction in burden.

A pattern that cannot be assessed for conformance provides no such reduction. A usable pattern publishes what may vary and what may not. “Use a gateway” is a preference; a pattern specifies the components, the trust boundaries, the required controls, and the configuration points that are open.

35.2 The ten patterns

Enterprise RAG

Problem. Ungoverned proliferation of retrieval assistants, each with its own grounding source and no lineage. Architecture. Shared vector store, governed knowledge base, retrieval gateway. Required controls. Data Lineage Record per source; S4 gate; maximum staleness defined per source. Primary risk. Grounding-source staleness or poisoning. Applicability. Tier 1–3; fast-path eligible at Tier 1–2. Watch for. Sources added after approval without re-running lineage. This is the most common drift for this pattern.

Secure LLM Gateway

Problem. Direct, ungoverned application-to-model calls with no logging or control point. Architecture. Centralized gateway providing authentication, logging, rate limiting, and redaction. Required controls. Gateway logs feed Control Plane observation; model allow-list enforced at the gateway. Primary risk. The gateway becomes a single point of failure and a performance constraint. Applicability. All tiers; foundational infrastructure rather than a per-system choice. Watch for. Bypass routes. A gateway with an undocumented direct path provides assurance it does not deliver.

AI Agent

Problem. Ad hoc tool-access grants with no recorded authority boundary. Architecture. Agent Card-governed identity, permission, and tool chain (Chapter 16). Required controls. Machine-readable boundary; kill-switch at Tier 3–4; preventive control for standing authority. Primary risk. Autonomy creep, confirmed in the framework’s stress-testing as a real failure mode. Applicability. Tier 2–4. Never self-service at Tier 1, because an agent at Tier 1 usually indicates mis-scoring.

Human-in-the-Loop

Problem. High-consequence decisions made without a genuine human check. Architecture. Mandatory approval step before a consequential action executes. Required controls. Documented oversight point; override-rate instrumentation. Primary risk. Automation bias hollowing out the check, confirmed in stress-testing. Applicability. Tier 3–4 decision-consequence systems. Watch for. An approve button that requires nothing the system did not already supply. Design the interaction so the human must contribute something.

AI Decision Support

Problem. Advisory AI mistaken for authoritative. Architecture. Recommendation-only output, no direct action capability. Required controls. D2 capped by architecture, not by policy alone. Primary risk. The formal-versus-effective autonomy gap, unresolved in the framework. Applicability. Tier 1–3.

Private or Sovereign AI

Problem. Sensitive data cannot leave the enterprise boundary. Architecture. Self-hosted or network-isolated model with no external egress. Required controls. Lineage record confirming no egress; data-location policy rule. Primary risk. Higher operational burden and capability lag against frontier models. Applicability. Tier 3–4 with high D5.

AI SaaS Consumption

Problem. Vendor AI features with no visibility, a confirmed structural gap. Architecture. Vendor feature behind a procurement and extended-ADR due-diligence gate. Required controls. Vendor evidence requirement; Tier-2 cap where evidence is unavailable. Primary risk. Opacity and evidence stalemate. Applicability. All tiers; the primary entry point for procured AI and the one most often ungoverned.

Multi-Model Routing

Problem. Single-vendor lock-in with no fallback. Architecture. Gateway routes across models by cost, capability, or risk. Required controls. Each candidate model separately ratified enterprise-wide. Primary risk. Inconsistent behavior across routed models complicates D1–D5 scoring, because the system’s behavior depends on which model served the request. Applicability. Tier 1–2 primarily.

Responsible AI Evaluation

Problem. Bias, fairness, and toxicity testing done inconsistently or not at all. Architecture. Standardized, reusable evaluation suite invoked at S6. Required controls. AI Assurance Summary referencing results. Primary risk. Tier 1–2 systems receive only sampling-based checks. Applicability. All tiers; intensity scales by tier.

AI Observability

Problem. No visibility into live AI behavior after deployment. Architecture. Telemetry feeding Control Plane observation, tagged to the drift taxonomy. Required controls. Eight-category drift classification. Primary risk. The underlying mechanism is not novel; value lies in the AI-specific categorization. Applicability. Mandatory Tier 2–4; recommended Tier 1.

35.3 Maintaining the library

A pattern library decays without maintenance, and a decayed library is worse than none because teams self-certify against patterns that no longer reflect good practice.

Promotion. A novel architecture reviewed at G2 and judged reusable is a candidate. Promotion requires ARB ratification, published variation points, and named required controls.

Review. Annually at minimum, and whenever the Learn function reports repeated drift against a pattern. A pattern generating drift across multiple initiatives is a pattern problem.

Deprecation. Retiring a pattern triggers re-review of every system still depending on it. Without that trigger, deprecation is a documentation change with no effect.

[Practice recommendation] Track pattern coverage — the share of new builds using an unmodified pattern — as a standing metric. It is the best available proxy for whether the library is reducing governance load, and it is the number that justifies continued investment in maintaining it.