B · The AI build lifecycle

Chapter 8. Architecture Decision: S3

Settling sourcing, model, hosting, grounding, and authority boundary — and recording the decision so that everything downstream has something to compare against.


8.1 What S3 decides

S3 produces the Architecture Decision Record that G2 assesses and that the Control Plane later compares deployed reality against. Its scope is the architecture boundary, not the engineering interior (Chapter 1, §1.3).

Six decisions belong here.

Table 27. What S3 decides

Decision Recorded as Downstream dependency
Sourcing Build, fine-tune, retrieve, API, SaaS, or no AI Determines governance weight (Chapter 4, §4.5)
Model or service Named model family from the approved list, or an exception Policy-as-code checks deployed model against this
Hosting pattern Vendor API, private hosting, on-premises Sovereignty and continuity implications
Grounding Which governed data sources, if any, ground the system Lineage record at S4; lineage drift detection later
Authority boundary For agentic systems: tools, actions, limits Agent Card; machine-readable boundary at G3
Pattern conformance Which reference pattern, unmodified, modified, or none Determines whether the fast path at G2 is available

8.2 The sourcing decision framework

S0 raised the sourcing question provisionally. S3 settles it. Twelve considerations bear on the choice; not all matter for every case, but each should be consciously dismissed, not overlooked.

Table 28. The sourcing decision framework

Consideration Pushes toward build or fine-tune Pushes toward API or SaaS
Differentiation The capability is a durable competitive advantage The capability is table stakes
Domain specificity Domain language or format is genuinely distinctive General language suffices
Data sensitivity Data cannot leave the enterprise boundary Data is low-sensitivity or can be redacted
Data sovereignty Jurisdictional restrictions on processing location No location constraint
Intellectual property Outputs or training data create IP that must be retained No IP consequence
Explainability need High-consequence decisions requiring defensible reasoning Explainability not material
Model control Behavior must be stable and version-pinned Vendor-managed updates acceptable
Latency Strict latency requirements favor local inference Network latency tolerable
Cost profile High sustained volume amortizes fixed cost Variable, low, or unpredictable volume
Vendor dependency Concentration risk is unacceptable Dependency acceptable and diversifiable
Compliance evidence You must be able to produce evidence a vendor will not supply Vendor attestations suffice
Skills You have and can retain the capability You cannot staff it durably

Two observations from the framework’s design worth carrying into the conversation. First, the sourcing decision is the architecture decision for externally consumed models; the vendor and data-flow choice is where nearly all the governable content sits. Second, fine-tuning commits you to treating every subsequent training-data change as a major change with full revalidation. That recurring cost is frequently omitted from the business case.

Sourcing and the “no AI” option

The no-AI option remains live at S3, not just S0. Architecture work sometimes reveals that the actual requirement is a deterministic rule with good data behind it. Recording that outcome as a valid S3 decision, not as a failed initiative, makes it available as a real choice.

8.3 Reference patterns and the fast path

Ten pre-approved patterns exist (Chapter 35). Their governance function is specific: a Tier 1–2 system built from an unmodified pattern may use the largely automated G2 path. That is the framework’s main genuine burden reduction, and it only works if pattern conformance is assessed honestly.

Three conformance outcomes are possible, and the distinction matters.

Conformant. The system uses the pattern as published, with only the configuration the pattern anticipates. Fast path available at Tier 1–2.

Modified. The system follows the pattern’s structure but changes something material — an additional data path, a different trust boundary, an extra integration. Fast path not available. The modification is documented in the ADR and reviewed on its merits.

Novel. No pattern applies. Full review regardless of tier, and a candidate for promotion into the pattern library if it proves reusable (Chapter 35, §35.3).

[Practice recommendation] Watch for pattern-conformance inflation, where teams describe a materially modified architecture as conformant to obtain the fast path. The counter is not suspicion but specificity: patterns should be published with an explicit list of what may vary and what may not. A pattern that says only “use a gateway” cannot be assessed for conformance and will be claimed for everything.

8.4 Writing an ADR that survives contact with the Control Plane

An architecture decision record in IRGF carries a load that a conventional ADR does not: it is the approved-state reference the Control Plane compares deployed reality against. Fields that cannot be compared produce no signal.

Table 29. Writing an ADR that survives contact with the Control Plane

Field Requirement Why the Control Plane needs it
Decision statement What was decided, in one sentence Human readability
Alternatives considered What else was assessed, and why rejected Review quality; not machine-compared
Rationale Why this choice Review quality
Model or service Named family and version policy Compared against deployed model registry
Hosting pattern Named deployment topology Compared against cloud inventory
Grounding sources Named governed data products Compared against data catalog
Authority boundary For agents: tool list, action limits Compared against IAM grants
Risk tier at decision Tier and Risk Classification Record reference Sets alert severity
Pattern conformance Conformant, modified with detail, or novel Policy-as-code check
Accountable owner Named individual Alert routing

The distinction between a field written for a human reviewer and a field written for machine comparison is worth making explicit to authors. “We will use a modern managed model with appropriate controls” is acceptable prose and useless as approved state. “Model family: [approved list entry]; version policy: pinned minor, vendor patch auto-accepted; grounding: Data Product PD-114 only” can be compared tomorrow morning.

Superseding, not editing

An ADR is never edited to reflect a later change. A new ADR supersedes it and references the one it replaces. This preserves the decision history that makes a drift investigation tractable and prevents the common failure where the record silently tracks reality and therefore never shows drift at all.