G · Artifacts

Chapter 24. The Artifact Set

Sixteen primary records, four derived views, and the non-duplication rule that keeps them from going stale.


24.1 The rationalization test

Every candidate artifact was tested against one question: what decision would become impossible or materially weaker without it? Artifacts failing the test were rejected outright or reclassified as derived views compiled from records that pass.

Three candidates were rejected entirely. An Enterprise Transformation Map duplicates a portfolio roadmap the framework does not own. An AI Capability Map is a filtered view of the business capability map. A Model Card is ML engineering documentation; only a reference to it is architecture-relevant.

Four were reclassified from records to views: Business Capability Map, AI System Card, Benefit Map, and Assurance Dashboard. Each is genuinely valuable and none should be independently authored.

24.2 The non-duplication rule

Every primary record has exactly one authoritative owner and one location. Every derived view is compiled by reference and never re-entered.

This is the rule that determines whether your governance data stays accurate. The AI System Card illustrates why. It is the single-pane summary practitioners most want: tier, owner, architecture, controls, authorization, open changes, open exceptions, assurance status. Every one of those fields already exists in another record.

Maintaining it by hand means every field has two homes, and the copy drifts from the source within weeks. Governance data that is confidently displayed and quietly wrong is worse than data that is visibly absent, because nobody checks it.

[Practice recommendation] If you cannot compile the AI System Card automatically, do not create it. A manually maintained version will be out of date at the moment it matters most.

24.3 The primary records

Table 64.

# Record Purpose Owner Created at
1 Capability Record Baseline and target maturity, gap, dependencies, AI-enablement flag Sponsor with Architect T1
2 AI Use-Case Canvas Business justification and provisional risk signal Business Sponsor S1
3 TG0 Initiative Qualification Record Evidence for the transformation qualification questions PMO, with mandatory architecture input T2
4 Architecture Decision Record Architecture decision and rationale, AI-extended Architect S3
5 Reference Pattern Library entry Pre-approved pattern enabling self-certification Architect On pattern proposal
6 Data Lineage and Sensitivity Record Grounding lineage and classification Data Owner S4
7 Risk Classification Record D1–D5 scores, computed tier, override flag Risk Lead S2
8 Control Matrix Control-to-risk-to-standard mapping Risk Lead S6
9 AI Assurance Summary Pass/fail evidence reference from technical evaluation System Owner S6
10 Agent Card Agent identity, tools, boundaries, kill-switch System Owner with Security S3
11 Deployment Authorization Record The G3 decision System Owner or AI Governance Body S7
12 AI Change Record Change classification and evidence System Owner S9
13 Architecture Exception Time-boxed exception Requester Any point
14 Benefit Record Realized value against baseline Sponsor T6
15 Drift/Alert Record Control Plane output System-generated On detection
16 Regulatory Overlay Reference Mapping of applicable regimes to affected systems Risk Lead On adoption, maintained

Four derived views compile from these: Business Capability Map, AI System Card, Benefit Map, Assurance Dashboard.

Three optional artifacts improve decision quality without being load-bearing: Technical Debt Register entry, Pattern Retrospective, and Stakeholder Map.

24.4 Worked templates

Risk Classification Record

Table 65.

Field Content
Purpose Record D1–D5 scoring and the resulting tier
Owner Risk and Compliance Lead
Trigger S2 initial; any Material or Major change; any risk-relevant drift alert; annual for Tier 3–4
Mandatory fields D1–D5 scores each with a one-line justification; D5 sub-scores for sensitivity and uncertainty; computed Impact and Control Deficit; resulting tier; safety-override flag; countersignature with date and any adjustments made
Approval Self-service for downward or lateral change at Tier 1–2; AI Governance Body for any upward change at Tier 3–4
Update Re-scored on every Major trigger; history preserved, never overwritten
Retention Life of system plus 3 years
Related Use-Case Canvas (input), Change Record (trigger), Control Matrix (downstream)

Example. Claims triage assistant. D1=3 (misrouting delays a claim and may disadvantage a claimant). D2=2 (proposes routing, adjuster approves each). D3=2 (routing reversible; elapsed delay is not). D4=3 (all claims; feeds claims system of record). D5=3 (sensitivity 3 restricted personal data; uncertainty 3 generative; composite = higher of the two). Impact = 3. Control Deficit = 2. Tier 2 by matrix. No override. Countersigned by Risk Lead, D3 raised from 1 to 2 — original score reasoned from mechanism reversibility.

Architecture Decision Record, AI-extended

Table 66.

Field Content
Purpose Record an architecture decision and rationale, with AI-specific criteria
Owner AI / Enterprise Architect
Trigger S3 for any new or materially changed AI System, Application, or Platform component
Mandatory fields Decision statement; alternatives considered; rationale; pattern conformance result or documented deviation; model or service named; hosting pattern; grounding sources named; authority boundary for agentic systems; risk tier at decision; accountable owner. For procured systems: vendor evaluation evidence, model update policy, data processed and location, contractual audit rights
Approval G2 — self-certified at Tier 1–2 against the pattern library; ARB at Tier 3–4
Update Superseded, never edited; supersession references the prior ADR
Retention Life of component plus 3 years
Related Pattern library (checked against), Risk Classification (referenced), Data Lineage (referenced)

Example. ADR-0142. Use Enterprise RAG pattern v2, unmodified. Model: [approved list entry], version policy pinned minor with vendor patch auto-accepted. Hosting: vendor API, EU region. Grounding: Data Product PD-114 (Policy Documents, Restricted) only. Alternatives: fine-tuned model rejected on recurring revalidation cost; no-AI rejected, rule-based routing tested at 61% accuracy against a 90% requirement. Pattern conformance: conformant. Risk tier at decision: Tier 2. Owner: A. Mensah.

Agent Card

Fields and a worked example appear in Chapter 16, §16.3.

Architecture Exception

Table 67.

Field Content
Purpose Record a time-boxed decision to proceed without meeting a requirement
Owner Requesting role
Mandatory fields The specific requirement excepted; justification; compensating control if any; hard expiry date; approving authority
Approval Per decision rights; always ≤ 90 days; no auto-renewal at Tier 4
Update Never silently renewed; expiry notifies the approving authority
Related Feeds portfolio health reporting as “initiatives operating under exception”

AI System Card, derived view

Table 68.

Field Content
Purpose Single-pane governance status, compiled not authored
Owner No independent owner; the System Owner is accountable for the underlying records
Fields, all by reference Current ADR; current tier; Control Matrix status; Agent Card if applicable; latest Deployment Authorization; open Change Records; open Exceptions; latest assurance status
Approval Not applicable; a compiled view has no approval step
Update Real-time or near-real-time from evidence-as-code capture

24.5 Tier-scaled documentation depth

Documentation depth scales with tier. This is where much of the framework’s proportionality actually lives.

Table 69. Tier-scaled documentation depth

Record Tier 1 Tier 4
ADR Architect self-certified, minimal narrative Full ARB-reviewed rationale
AI Assurance Summary Pass or fail reference Full underlying evaluation reference chain
Agent Card Kill-switch field not required Kill-switch conditions mandatory and separately approved
Risk Classification Countersigned, brief justifications Countersigned, full justification per dimension, annual re-score
Control Matrix Mapped controls listed Mapped controls with current effectiveness evidence

24.6 Where to keep records

The framework does not mandate tooling. Three properties matter more than product choice.

One authoritative location per record type. Records split across a wiki, a ticketing system, and a spreadsheet cannot be compiled into derived views, which means the views get maintained by hand, which returns you to §24.2.

Machine-readable where the Control Plane must read it. Model family, hosting pattern, grounding sources, authority boundaries, and tier need to be fields, not prose in a document.

Referenceable identifiers. Every record needs a stable identifier that other records cite. This sounds trivial and is the thing most often missing; without it, the traceability chain is a set of documents that mention each other by name.