AGENT
AI that is bound to the evidence
The agent layer does not answer from a language model's memory. It plans, queries the ontology under the caller's permissions, calls registered engines, criticises its own result, and cites every claim. Anything it cannot support, it says it cannot support.
01
The loop
USER QUESTION
|
PLAN decompose into steps
|
RETRIEVE ontology query, permission-filtered
|
COMPUTE call a named engine, not a guess
|
CRITIQUE check result against evidence
|
ANSWER with citations
|
ACTION? -> human approval -> execute -> log02
Tools, not free text
Every capability the agent has is a registered tool with a schema, a permission requirement and an audit record.
TOOL REGISTRY query_ontology read objects and links run_model invoke a registered model simulate run a scenario search_documents retrieve source documents compute_metric evaluate a defined metric open_problem create a Problem object notify send to a person or unit dispatch request a field action [approval] write_back update an object [approval]
03
Citations are mandatory
ANSWER
claim 1 ............... [obj:<id> @ <date>]
claim 2 ............... [doc:<id> p.<n>]
claim 3 ............... [model:<name> v<n>]
not found in the ontology:
<the parts it could not support>An uncited sentence from the agent is treated as a defect, not as a style preference.
04
Where the human stands
read
Reading and analysis
Automatic, within the caller's permissions. Nothing is exposed that the person could not open directly.
recommend
Recommendation
Automatic, always accompanied by the alternatives considered and rejected.
act
Irreversible action
Requires a named human approver. Dispatch, write-back, notification to a citizen and enforcement all sit here.
No automated action on a person
The platform never takes an adverse action against an individual on model output alone. A named official decides and is recorded as having decided.
Disagreement is data
When a human overrides a recommendation, the override and its reason are stored and used to evaluate the model.