Models draft and suggest. People decide.

The problem
Every serious funding decision in this product requires a human. That is not caution about model quality — it is that accountability for a funding decision cannot be delegated to something that cannot be accountable.
What the system contributes
It reads the sources, extracts what they require, and keeps every derived figure attached to the record it came from — so a number can be checked rather than trusted.
What a person decides
Every consequential act here is a named human decision, bound to the exact version it was made against. The system prepares it; it never takes it.

The boundary, as the system enforces it

A model can never perform an act that commits the organisation: approving, certifying, submitting, awarding, attesting or waiving. This is enforced in the domain layer, not by policy.
A model-produced claim carries the epistemic state "Predicted" and can never be recorded as "Known". Observation is what makes something Known.
Model output is never a basis for a consequential act. It can inform the person who performs one.
Anything a model generates for a document is assembled from facts a person already approved.

Ask AwardBase
Planned
Ask questions about your funding position in ordinary language, answered from your own records with the citations attached.
Agent Observatory
Planned
A record of what every agent did, what it was authorised to do, and what it produced.