Governance › Decision Register
Decision Register
The register that keeps every other page durable. Principles are the standing rules. Decided items are settled calls with their reasoning, so nobody relitigates them from memory. Open and deferred items are the only place in this portal where transient content is allowed to live: when something changes, this page changes; the architecture pages do not.
Principles: the standing rules
- Values never code. The client plane carries configuration only. Custom logic is built in the Solution layer as an overlay and selected on the plane; a client-specific edit to shared code is a defect.
- Deterministic by default. Any task with a testable right answer is code; agents get judgment only, and only where a measured comparison shows judgment wins. An agent is the last resort, not the first draft.
- Two MCP boundaries, never merged. The north gateway serves hosts (platform as MCP server); the south tool servers reach data (platform as MCP client). Two security planes with different threat models.
- Routing policy is coded, transport is configured. The routing table is a versioned P1Z asset shipped through CI. The cloud carries only inference plumbing; no cloud-native intelligent router ever chooses a tier.
- Business-blind harness. The environment layer never learns what the business terms mean; the business-blindness linter in CI enforces it. The harness builds slots, the layers above fill them.
- Classification inheritance. Feedback and every derived artifact carry the classification of their source data. Restricted-class content is stripped before anything reaches the shared knowledge shelf.
- Deterministic rendering. A deliverable is a pure function of an immutable, versioned template and approved structured data. Corrections land as data updates and the artifact regenerates; output is never hand-edited.
- Source of truth stays in the source estate. The platform is never the system of record. Every platform store holds derived, disposable copies; the originals remain in the client's own estate.
- Two sealed loops. The operational loop tunes how the platform runs; the knowledge loop promotes what is true. Neither ever crosses into the other.
- Generated views, never hand-edited. Every published page and per-client file is regenerated from its source of truth; CI blocks hand edits. If a view and its source disagree, the source wins.
The determination rule, in full
The default is deterministic. Four tests; a single yes on any of them means the task is built as deterministic code, not as an agent:
- Can you write a unit test with an exact expected output?
- Would two competent professionals always produce the same result?
- Must it run identically every time (routing, gates, permissions, state transitions)?
- Is failure machine-detectable without judgment?
Deterministic code owns: all calculation (the domain rule library), validation, schema transforms, retrieval mechanics, orchestration and state machines, gate and permission mechanics, formatting, and data movement. Agents own only judgment under ambiguity: synthesis across conflicting sources, narrative drafting, interpretation, and exception triage; and only when a measured comparison shows the agent beats the simpler structure, and the output faces a human gate. Agents may call deterministic tools; workflows may invoke an agent for one judgment step; every responsibility keeps its label. An agent is the last resort, not the first draft.
Decided: settled calls, and why
| Decision | Rationale |
|---|---|
| Seven layers, Experience at the top | The stack reads the way a client experiences it. Finer slicings (personas, use cases, output, HITL as layers) collapse into these seven without losing a boundary. |
| Integration is its own layer, above Data | Different technology and different skillset; movement fronts storage, and every agent tool call and outbound API call enters at the middleware. |
| Intelligence sits above Integration, behind one router | Reasoning is rented, not owned. Nothing above the router names a vendor, so a model swap is a one-file change and model progress is absorbed, not rebuilt for. |
| Two MCP boundaries | Hosts connect down through the north gateway; sources are reached down through the south servers. Merging them would merge two security planes. |
| Cloud overlays live in Environment & Operations only | Nothing above L7 changes with the cloud choice; an overlay on any other tab would claim otherwise. |
| Lab tenant folded into the ADLC | Sandboxing is lifecycle, not architecture. It appears as ADLC step 2, with exit only through the promotion checklist. |
| HITL is plane configuration, not a layer | Review style varies per client. Blocking mechanics are harness machinery; placement and approvers are plane values. |
| The client plane is a vertical band, C1 to C7 | Client customization drawn with equal visual weight to the platform, so it is planned in discovery rather than improvised in delivery. |
Open: being worked, expected to change
| Item | Status |
|---|---|
| Per-engagement evaluation criteria | Set during discovery (sections C and H baselines); no platform-wide default declared yet. |
| A2A specialist-agent exposure | Any platform agent consumed by an external host over the A2A protocol must sit behind the north gateway, or be recorded here as an explicit exception. The orchestrator stays server-side and authentication stays at the gateway; a host reaching a specialist agent directly bypasses both, which is the Shadow Architecture failure mode. Decide before the first external host integration. |
| Managed harness runtimes, buy versus build | Anthropic Managed Agents (beta), AWS AgentCore Runtime, Azure AI Foundry Agent Service and Google Vertex AI Agent Engine each offer the L7 runtime slice as a service. Renting the harness from a model vendor moves the lock-in question from L4 to L7, and Managed Agents currently carries no Zero Data Retention or HIPAA coverage. Evaluate before the first enterprise deployment. |
| Register cadence | Each architecture review confirms or retires the entries on this page; a decision confirmed three times graduates into the environment spec. |
Deferred: deliberately parked, tracked with activation triggers
| Item | Scope | Activation trigger |
|---|---|---|
| Engagement EXIT procedures | data return, secure destruction, retained permitted aggregates | activate before first enterprise contract signature |
| Cross-layer compatibility matrix | knowledge vX ↔ solution vY release train | activate at the second concurrent client on different versions |
| Regulatory / compliance profile mapping | EU AI Act, client model-risk policies, audit evidence packs | activate when compliance review first appears in procurement |