Architecture decision record

Context

Cases must progress reliably and be testable and auditable at every step. A language model is probabilistic and cannot be the component that decides what happens next in a critical operation.

Options considered

1. Deterministic workflow orchestrator that calls services · chosen The workflow engine owns state and sequence; the model is one service it calls.
2. Autonomous agent that plans its own next actions The model decides sequence and tool use at runtime.
3. Linear script or RPA only A fixed script with no model reasoning.

Decision

A deterministic orchestrator owns the workflow and the case state machine. The model is invoked only as a called service inside a single state, and never advances the case on its own.

Context

The operation needs the leverage of AI for reading documents and drafting assessments, without allowing ungrounded output to influence a decision.

Options considered

1. AI drafts inside one state, grounded by retrieval and citations · chosen Output must cite retrieved sources and pass policy checks.
2. AI available across several states The model contributes at triage, enrichment, and assessment.
3. No AI; deterministic rules only Rules and templates handle everything.

Decision

The model is confined to the Assessed state. It must ground every claim in a retrieved citation, and its draft must pass the deterministic policy-and-rules checks before it can reach the approval gate.

Context

Some actions are routine and safe to automate; others carry enough consequence that a human must remain accountable. A single policy for all actions would be either unsafe or unusably slow.

Options considered

1. Consequence-tiered approval gate · chosen Autonomy granted in proportion to the consequence of the action.
2. Human approves every action Maximum oversight, minimum throughput.
3. No human gate Maximum throughput, no accountability for consequential calls.

Decision

An explicit approval gate is a named state. Low-consequence actions run automatically within set bounds; medium-consequence work is decided by a case handler; high-consequence actions require a senior approver.

Context

The operation must be fully traceable and must enforce least privilege, while still acting in the organization's systems of record.

Options considered

1. Single data boundary with an append-only audit log and integration adapters · chosen All state inside one boundary; actions leave only through adapters.
2. Services write directly to systems of record Each service integrates independently.
3. Shared database with no explicit boundary Convenient but unbounded access.

Decision

Cases, documents, and an append-only audit log live inside one case-data boundary under least-privilege access. Actions reach systems of record only through integration adapters the boundary can observe.