1 · System architecture

How a governed AI action moves through the control plane and produces tamper-evident decision evidence. Every surface — UI, API, CLI and agent tools — reaches the same policy decision point, so no route bypasses evaluation. Conceptual component view.
flowchart LR ui([Controller UI]):::ext intg([API integration]):::ext cli([CLI]):::ext mcp([MCP agent tools]):::ext edge[Secure ingress
TLS · rate limiting]:::edge subgraph plane["AI Control Plane"] api["Control-plane API
single policy decision point"]:::core mandate["Delegation mandate check
fail-closed"]:::svc policy["Policy decision engine"]:::svc gw["Model gateway"]:::svc idp["Identity & access
agent credential issue · rotation · revocation"]:::svc sec["Secrets & signing
credential & API-key redaction"]:::svc end evid[("Decision-evidence store
append-only · hash-chained · signed")]:::db timeline[("Agent activity timeline")]:::db models[("AI model providers")]:::ext ui --> edge intg --> edge cli --> edge mcp --> edge edge --> api api -->|require valid mandate| mandate api -->|authorize| policy api -->|model call| gw --> models api -->|authenticate| idp api -->|signing keys| sec api -->|record decision| evid api -->|record agent activity| timeline classDef ext fill:#eef2f8,stroke:#9fb3d1,color:#0B1F3A; classDef edge fill:#e7eefb,stroke:#7fa1d8,color:#0B1F3A; classDef core fill:#e3f1ea,stroke:#5fae87,color:#0B1F3A; classDef svc fill:#eef2f8,stroke:#9fb3d1,color:#0B1F3A; classDef db fill:#f6ecf3,stroke:#c79ab8,color:#0B1F3A;

2 · Governed decision & evidence chain

Every governed AI action is policy-checked first, then recorded as tamper-evident, signed evidence with an audit trail — before the outcome is returned.
sequenceDiagram autonumber participant Caller as User / AI agent
(UI · API · CLI · MCP) participant ACP as AI Control Plane participant Mandate as Delegation mandate participant Policy as Policy engine participant Evidence as Decision evidence
(signed · hash-chained) participant Audit as Audit trail &
agent timeline Caller->>ACP: Requested AI action ACP->>Mandate: Valid mandate for this actor and scope? Mandate-->>ACP: Granted · missing · out-of-scope Note over ACP,Mandate: No valid mandate, or action outside its
constraints, is denied (fail-closed) ACP->>Policy: Evaluate tenant-scoped policy Policy-->>ACP: Allow · deny · approval-required ACP->>Evidence: Record decision evidence with the
accountable owner captured at decision time ACP->>Audit: Append audit event and agent activity
(credential and API-key material redacted) ACP-->>Caller: Outcome + evidence reference
Property: evidence is append-only, hash-chained, and signed, so any later alteration is detectable on review. Mandate enforcement is fail-closed — an action without a valid, in-scope mandate is denied rather than allowed by default — and each decision record carries the accountable owner as at the moment the decision was made, rather than reconstructed afterwards.

3 · Proof pack & regulatory-review boundary

How governed evidence rolls up into a repeatable, signed proof pack that a regulated firm and its reviewers can assess — and what stays the firm's own responsibility.
flowchart LR gov["Governed AI decisions
+ decision evidence"]:::core proof["Signed proof-pack manifest
repeatable release gate"]:::svc reg["FCA/PRA evidence-pack register"]:::svc review["Firm review
risk · compliance · audit · accountable owner"]:::svc dd([Buyer diligence decision]):::ext pub["Published run summary
status · freshness · attestation state"]:::svc gov --> proof --> reg --> review --> dd proof --> pub --> review reg -. boundary .-> b1["Technical & operational evidence only
not an approval or certification"]:::bound review -. boundary .-> b2["Firm-specific service mapping, impact
tolerance & ownership remain buyer-owned"]:::bound classDef core fill:#e3f1ea,stroke:#5fae87,color:#0B1F3A; classDef svc fill:#eef2f8,stroke:#9fb3d1,color:#0B1F3A; classDef ext fill:#e7eefb,stroke:#7fa1d8,color:#0B1F3A; classDef bound fill:#fbf4e7,stroke:#d8b877,color:#0B1F3A;