SieF Architecture

Four systems, two of which use AI

The architecture separates deterministic authority from generative assistance, in the structure itself rather than in policy.

Pillar	AI	Responsibility
Brain	No	Holds the domain rules, the job catalogue, routing, permissions and approval gates. Every job starts here.
Engine	No	Executes approved workflows and tools, produces records and files, and logs what happened.
Judge	Yes	Gives bounded advice to Brain where a case needs interpretation. Never authorises execution.
Fleet	Yes	Runs the model workloads Brain delegates, inside the resource policy the OS enforces.

The ordering matters: Brain decides, Engine acts, Judge advises, Fleet computes. Advice is not a decision, and computation is not authority.

The four execution routes

Every piece of work follows one of four paths, and which one it followed is part of its record:

1. Brain to Engine. The rules fully decide. Engine executes and records.
2. Brain to Judge to Brain to Engine. The case needs bounded interpretation. Judge advises, Brain decides, Engine executes. The advice is kept with the record.
3. Brain to Fleet to Engine. A marked gap needs generative work. Fleet fills only that gap, and the result is checked before it counts.
4. Brain to a person to Engine. The decision belongs to a human. Work stops at the gate until an authorised person acts.

What a framework supplies

The operating system implements the universal core. A framework extends it through four packages and never patches it:

• Brain extension: domain rules, jobs, schemas, gates, approvals, fixtures.
• Engine extension: workflows, domain tools, integrations, records, failure and recovery behaviour.
• Judge extension: advisory scopes, evaluation criteria, escalation rules.
• Fleet workload manifest: the model capabilities required, the quality bars and the quotas, declared rather than requested at runtime.

A framework that needs the core to behave differently is asking for a change to the core, which is reviewed as one.

Capability promotion

A new framework may lean on bounded AI more at the start. Each resolved case can become a candidate rule, but a successful inference does not become a trusted rule by itself. Promotion requires regression tests, domain review, security and permission testing, a named approver, a version and rollback.

On a customer site, people approve new rules in the gate, customer-added rules may only tighten behaviour and need two approvers, and no model is trained there.