Government AI Assurance Network · a reference build
Completed is not verified.
Government AI assurance rarely fails inside a system. It fails between them — in the gap between “someone said it was fixed” and “someone checked”. This network connects the systems that sit either side of that gap, and you can watch it work.
A network is a group of separate applications that cooperate.What is a network?
5 applications and an Orchestrator: AI Register, Controls, Independent Audit, Governance, Management View.
Fictional scenario and synthetic data, on a DSLCore reference build. Not an official system, and the department in the story does not exist. Every replay is a real recorded run of the network; nothing on these pages is live.
The story: one issue, six steps
An AI assistant goes into service
A department runs a public-facing AI assistant. It is entered in the AI register, and management’s view starts overseeing it. Its position is Green: nothing is wrong.
See it happen →A control turns out not to work
The assistant is meant to have human review of its answers. The controls system finds that nobody is doing it. Governance opens an exception, and management’s view turns Amber — with the reason written next to the colour.
See it happen →Someone is told exactly what to do
A decision is recorded and becomes an obligation on a named role: start weekly human review. The obligation reaches the system that owns the fix, so nobody has to chase it by email.
See it happen →The owner says it is fixed
In most places the story ends here, with a status changed to “closed”. In this network, “fixed” is only a claim. It goes to independent audit to be tested.
See it happen →The independent check can say no
Audit looks at the evidence. In one of the recorded runs it fails: ten decisions sampled out of thousands. The work goes back, the failed check is itself recorded as an exception, and nothing closes.
See it happen →Only then does it close
The second check holds. The exception closes, the obligation is marked met, management’s view returns to Green — and the whole trail, from the first finding to the final check, stays on the record.
See it happen →
Who is who
Five systems, each doing one job, and an orchestrator that carries messages between them. These are existing DSLCore reference applications, connected rather than rebuilt.
AI Register
Knows which AI systems exist, their risks, incidents and monitoring.
Controls
Tracks the controls that should protect them, the weaknesses found, and the fixes.
Independent Audit
Tests the evidence behind a “fixed” claim, separately from the person who made it.
Governance
Turns serious issues into exceptions, decisions and obligations with owners and dates.
Management View
Shows one position per AI system — Green, Amber or Red — with the reasons, decisions and commitments behind it.
Orchestrator
Carries every message between the five, retries when a system is down, and keeps a record of what was sent and received.
Why connect them
One trail, not five
Every step from finding to closure is linked across the systems, so anyone can see who knew what, who decided, and who checked.
A claim is not a result
Marking work complete does not close the issue. Independent verification has to pass first.
Failure is handled, not hidden
A failed check reopens the work and is recorded as its own exception, instead of disappearing into a status field.
Management sees why
A Green, Amber or Red for each AI system, with the signals behind it, updated as things happen rather than at the next report.
Each system keeps its job
The owner of each kind of data keeps it and its rules. Systems cooperate by passing messages; none reaches into another’s database.
Nothing falls between the cracks
Messages are retried if a system is briefly down, and a commitment that passes its date is flagged overdue and escalated automatically.
In one simulated run of 30 scenarios, 249 messages were carried between the systems with none lost, and 2 of the 15 independent checks failed first and sent the work back — which is the system doing what it is for.
Watch it happen
Each recording is a real run of the network. Play it, pause it, step through it, and open any message to see who sent it and who received it.
Governed remediation
A serious control deficiency is found, decided on, assigned, fixed, independently verified and only then closed.
The first check fails
The owner says it is fixed. The independent check fails it. The work is redone and checked again before anything closes.
Management intervention
The position worsens, management decides and commits, the commitment is carried out and verified, and the position recovers.
A required control fails
A control is required for an AI system. Controls assesses it and finds it ineffective; the AI register is told and governance opens an exception.
An overdue commitment
Management commits to a fix. The date passes without it being met, and the commitment is marked overdue and escalated.
What this is, and is not
It is a working reference build showing how existing DSLCore applications can be connected so an AI issue is found, decided, fixed and independently checked before it closes. It is not an official government system, it is not presented as compliant with any named framework, and the numbers come from synthetic data on a single machine.