Diagram interlude
Execution is a separate authority surface.
HELM keeps proposal generation separate from the decision to act, so missing policy, approval, or proof can deny or escalate before dispatch.
The Separation of Intelligence and Authority
Many agent frameworks mix intelligence and authority. The model decides what to do and then calls the tool. That creates a direct path from guesswork to side effects.
HELM makes a clear split: Models Propose. HELM Governs Execution.
The Proposal
In HELM, models advise and draft. They read context and suggest work. The request becomes a structured spec that can be checked.
The model’s output is never executed directly. It is treated as untrusted input.
The Governance Boundary
Once a request is created, it crosses the HELM boundary. HELM checks it before any tool runs:
- Shape check: Is the request written in the expected form?
- Policy check: Does the action break company rules or role limits?
- Human review: Does this action need a person to approve it first?
Execution and Evidence
Only after the request passes the checks can a tool run. The model does not run the tool by itself.
Every step can be recorded: the request, the checks, any human approval, and the final result.
This separation lets a company use AI reasoning without giving the model final authority over tools, systems, or customer-facing work.