Infrastructure 01 of 06
SieF Architecture
The common operating architecture every SieF framework extends: Brain, Engine, Judge and Fleet, with the authority between them fixed.
In development
What it enables
- A single way to describe how work is decided, executed, advised on and computed, so two frameworks built by two companies behave predictably on the same machine.
- A boundary that can be audited: deterministic work is separable from generative work, in the architecture rather than in a policy document.
- Domain extension without forking. A framework adds its own rules, tools, advisory scope and model workload, and changes nothing in the shared core.
Who it serves
- SieF Ventures and SieF Developers, who build against the four extension contracts.
- Enterprise customers, who need to know which decisions a machine may take and which it may not.
- Reviewers, who need the same structure in every submission to judge it fairly.
Scope and boundaries
- Brain and Engine run no AI. Brain holds the rules and makes the decision; Engine executes approved work and records what happened.
- Judge is AI, and advises Brain. It never authorises execution and its advice is not a decision.
- Fleet is AI, and runs only the model work Brain has explicitly delegated, inside the resource policy the OS enforces.
- A framework never patches the shared core to get domain behaviour. If it needs the core to change, that is a change to the core, reviewed as one.
In practice
- An invoice arrives. Brain routes it by rule. Where the rules fully decide, Engine completes it and writes the record.
- A case falls outside the rules. Brain asks Judge for a bounded reading, then decides. The advice is kept with the record.
- A document needs interpreting. Brain delegates that piece to Fleet, which runs a permitted local model and returns the result for checking.
- A decision needs a person. Brain stops at a gate and waits. Nothing proceeds on the machine’s own authority.
Where we are
In development. The extension contracts are being written against the first real implementations, so that what publishers build against is proven rather than theoretical.
Questions
- Is this a multi-agent system?
- No. A set of agents without reliable domain procedure is not this architecture. Rules come first, and bounded AI is used only where the architecture authorises it.
- Can a framework teach itself from what it sees on a customer site?
- It can propose. A resolved case may become a candidate rule or tool, but promotion needs tests, review, a named approver, a version and customer authority. Model weights are not trained on a customer site.