Summon Central gold medallionSUMMON CENTRALConstitutional AI Governance

THE CHALLENGE · TRUST GAP

AI capability is accelerating. Who governs its consequences?

Credentials can be mistaken for permission. Remembered claims can become instructions. A model can announce success without demonstrating the actual effect. The answers require structures outside the model.

Explore old versus new →Request institutional evaluation ↗

THE OLD WAY AND A DIFFERENT POSSIBILITY

From powerful agents to accountable transactions

A design comparison, not an empirical benchmark or claim that every existing framework behaves this way.

Traditional risk pattern

Capability may be confused with permission

  • Standing credentials and broad tool access.
  • Context or memory treated as implicit instruction.
  • Reported success accepted without separate observation.
  • Ambiguous failures invite unexamined retries.
  • Runtime substitution obscures state and liability.
SUMMON'S GOVERNED PROFILE

Legitimate work must be re-authorised

  • Finite permission and narrowing delegation.
  • Exact scope, memory view and intended action committed.
  • Scoped effect evidence and separate readers.
  • UNKNOWN preserved when outcome cannot be established.
  • Substitution without automatic transfer of authority.
CASE STUDY · AI AGENTS AND AUSTRALIAN PUBLIC RESOURCES

Could a governing boundary have changed the outcome?

ABC’s 2 October 2026 report described unauthorised access to the Medicare statistics portal on 18 June and separate activity involving NSW applications. AIHW’s 25 September statement found no evidence of compromise or unauthorised access in its own investigated interaction. These are different findings.

WITHOUT A GOVERNED BOUNDARY

A capability reaches a tool

It may be technically possible to query a site, retrieve a file or invoke an endpoint even though the actor's legitimate scope does not permit it. The model's own intentions or success claim do not substitute for independent controls.

A PROPOSED DBGA-GOVERNED TRANSACTION

The permission is checked first

The protected connector must know which operation is authorised, for whom, against which resource and what happens when permission is denied. Covered results and uncertainties are captured for separate examination.

Would Summon have stopped it? A correctly positioned, configured and enforced boundary might have refused covered attempts, but not every attack path. Receipts and observation apply only to actions actually governed; no retroactive visibility or universal protection is claimed. Real incidents must become hostile test cases, not marketing proof.

Sources: AIHW official account ↗ · ABC News, 2 Oct 2026 ↗ · ABC News, 2 Oct 2026 ↗

THE GOVERNANCE GAP

Three questions the industry cannot avoid

A powerful model can contribute to all three answers. It cannot be the sole authority deciding them.

I

What is allowed?

An actor, scope, purpose, time and budget must be authorised by legitimate controls.

II

What may be known?

Governed memory needs provenance, disputed claims, disclosure permissions and withdrawal.

III

What happened?

Producer self-report, physical observation and independent reconstruction are different classes of evidence.

EVIDENCE FOR ENGAGEMENT

Public concern is not a single statistic

Different populations and surveys should not be added together. These sources describe real concerns, not demand for any named vendor.

AU

Trust and adoption diverge

KPMG / University of Melbourne (2025) reported Australian AI usage above willingness to trust it and substantial concern about adverse outcomes.

Primary source ↗
US

Control remains a priority

Pew (2025) found significant demand for greater control over AI among both US adults and surveyed AI experts.

Primary source ↗
GOV

Standards require operational evidence

NIST AI RMF and Australian agentic AI guidance articulate governance, oversight and risk-management requirements to evaluate against.

Standards & policy →

A challenge to the AI industry

A statement of concern is not an engineering response.

You asked for accountable agents. Examine the mechanisms and help establish where they succeed—and where they fail.

View the evidence →