Submission 494
Mr Kelvin Chau
Submitter group: Individuals.
Position labels are not shown for individual submitters.
An AI model identified the submitter group and wrote the summary, key points and recommendation summaries. Counts do not measure importance or merit.
Summary
Read the original submission for the full text.
Argues sovereign AI requires institutional capability and continuously revisable rules. Proposes capability tests, workforce renewal, orchestration skills, architectural accountability, portable Australian-data models and lifecycle controls through existing machinery, without a new body, levy or appropriation. The author created the open-licensed frameworks offered for adoption without dependency; acknowledges layered adaptation is unresolved and a shared Australian evaluation suite does not yet exist.
Key points
Quotes are the submitter’s own words. Text taken from documents, including scanned pages read by text-recognition software, can differ from the PDF. Check the original. Listed in the same order as on the submission’s page in the interactive guide.
Argues that the institutions that govern AI run on a slower clock than the technology, and that the gap is widening.
the institutions that must govern this technology run on a slower clock than the technology, and the gap is widening.
Argues that sovereignty is a matter of structure and capability rather than owning a large model.
The answer to someone else's giant model is not your own giant model. It is structure.
Proposes a vendor-swap test as the measure of whether sovereign capability is real.
a final exam, the swap test: a governance layer demonstrated on a live workflow across two different vendors' models, and, for the sovereignty the test is named for, against at least one model Australia can host and run under its own jurisdiction.
Argues for separation of duties: the component that acts should never be the component that judges, and the judging component must be independent on a stated axis, such as a different base model, vendor or a person.
the component that acts is never the component that judges, and nothing approves itself; and the judging component must be independent on a stated axis, a different base model, a different vendor or a person, because three similar AI operators do not create real checking if they reason through the same frame
Argues that a review trigger must require the responsible officer to re-decide whether an instrument remains fit for purpose, not merely gather information.
A review trigger that can only gather information is decoration; the trigger must require the responsible officer to re-decide whether the instrument remains fit for purpose, with the power to vary, suspend or retire it, and to record the decision.
Argues layered adaptations should be pursued through portability and data-handling procurement conditions, while acknowledging their composition is unreliable and unstandardised.
Composing layered adaptations is unreliable and unstandardised today, and this architecture should be pursued as architecture and demand, procured with portability and data-handling conditions, not asserted as solved technique.
Recommendations
Recommendations are shown beside the submission’s own words. Listed in the same order as on its page in the interactive guide.
Adopt sovereign AI capability measures in National AI Plan reporting, incorporating five capability dimensions and a swap test with an independently set pass criterion; report in aggregate on terms the Government settles.
adopt an evidence-based measure of sovereign AI capability within National AI Plan reporting, incorporating the five capability dimensions and a swap test whose pass criterion is set independently of the entity tested, reported in aggregate, in terms the Government settles.
A program corrected obvious text-recognition errors in this quote for display.
For the use dimension, require a live redesigned workflow with task boundaries written down, including identification of tasks that train future judgement.
USE choosing problems, decomposing wok, redesigning workflows; one live workflow, redesigned, with task identifying which tasks train future judgment boundaries in writing
For the direct dimension, require a decision class expressed as common decision objects owned by Australia.
DIRECT rules and decisions in your own gammar, readable by lawyer and machine __ one decision class expressed as common decision objects the country owns
For the verify dimension, require a baseline evaluation suite on every release and results readable by a stranger.
VERIFY evaluation sutes, regression discipline, evidence packets a baseline evaluation suite on every release, results a stranger can read
For the govern dimension, demonstrate a governance asset through a full lifecycle including retirement, with nothing approving itself.
GOVERN the full lifecycle with teeth: environment change sensed, anajsed, one governance asset through a full decided, announced, communiated, enforced, reviewed, retired; nothing lifecycle, including a retirement approves itself
For the capture dimension, demonstrate an approval stating its outcome in advance and verified against actual outcomes.
CAPTURE _ value landed in the country: outomes stated first, verified against actuals one approval that stated its outcome in advance and was verified against actuals
Demonstrate governance instructions, evaluations, lifecycle controls and evidence schema on a consequential live workflow across two vendors and at least one model Australia can host under its own jurisdiction.
The final exam: the swap test. Demonstrate the governance layer, instructions, evaluations, lifecycle controls and evidence schema, working on a live workflow of stated consequence across two diffeent vendors' models, and against at least one model Australia can host and run under its own juisdiction.
For the swap test, require held-out evaluation material not authored by the tested entity, independently set tolerance, and publication of the suite, tolerance and results.
Pass ctiterion: a baseline evaluation suite that includes a held-out component the entiy did not author returns results within a tolerance set by a function independent of the entty; the suite, the tolerance and the results are published.
Calculate national grades as shares of Commonwealth entities or their AI-supported decision classes meeting each pass mark; choose relied-upon workflows, decision classes and governance assets; treat pass marks as floors.
Who sits the exam. The pass marks are institution-scale, and each is a floor not a ceiling: the wokflow, decision class and go\ernance asset chosen must be ones the ently relies on, stated as such. A national gade is the aggregate share of Commonveatth entities, or of their AI-supported decision classes, that meet @ch pass mark, so the card measures diffusion, not exhibits.
A program corrected obvious text-recognition errors in this quote for display.
Have Commonwealth-administered skills products state review arrangements and retirement paths for competencies affected by AI, and have Commonwealth AI workflow guidance require identification of tasks that develop human judgment.
Commonwealth-administered skills products state review arrangements and retirement paths for competencies affected by AI; that Commonwealth AI workflow guidance require the identification of tasks that develop human judgment
A program corrected obvious text-recognition errors in this quote for display.
Draw the committee's observations on skills frameworks to the attention of accrediting bodies and the ministerial machinery for qualification frameworks.
that the committee's observations be drawn to the attention of accrediting bodies and the ministerial machinery for qualification frameworks.
Put governed multi-model orchestration, disciplined instruction, information condition and curated adaptation into national AI skills programmes, adoption supports for small, medium and family businesses, and procurement guidance.
Put orchestration, governed groups of models, disciplined instruction, (a), (d) skills frameworks, SME information condition and curated adaptation, into national AI skills programs supports, procurement guidance and adoption supports for small, medium and family businesses
A program corrected obvious text-recognition errors in this quote for display.
For consequential government AI, separate acting and judging components, prohibit self-approval and specify independence through a different base model, vendor or a person.
Separation of duties [3]: the component that acts is never the component that judges, and nothing approves itself; and the judging component must be independent on a stated axis, a different base model, a different vendor or a person, because three similar AI operators do not create real checking if they reason through the same frame [6].
For consequential government AI, default to non-execution and queue a human question when authority is unclear.
The designed halt: when authority is unclear the default is non-execution and a human question is queued.
Record consequential government AI decisions as made, reconstructable from actual inputs, logic and outputs rather than later explanations; distinguish the decision-maker from the recorder.
The trail, and here the distinction matters for transparency as much as accountability: decisions are recorded as they are made, in forms a stranger can reconstruct from the actual inputs, logic and outputs. Reconstructability is not explanation. An explanation generated afterwards is a new summary that may not faithfully represent the processing it describes; the record must be walkable, not narrated, and the actor who decides must be distinguishable from the actor who records the decision.
Reserve decisions about what is being made, where thresholds sit and whether results are accepted to humans.
The reserved decisions: three questions that never go to the machine, what are we making, where do the thresholds sit, and is the result accepted.
State accountability controls in tenders and test at acceptance; for managed services whose architecture cannot be inspected, require evidence of separation deliverable to the agency, beyond supplier attestation.
Each is an old control re-aimed at software components, and each can be written into a tender and tested at acceptance. For systems procured as managed services, where an agency cannot inspect the architecture, the requirement is met by a stated requirement in the tender and a stated test at acceptance, with evidence of separation deliverable to the agency; a supplier's attestation is not the test.
Specify portability and data-handling conditions in Commonwealth procurement of narrow Australian-data models for regions and sectors, treating layered-adaptation composition as unresolved.
The committee recommends that Commonwealth procurement of narrow, domain-specific models on Australian data specify portability and data-handling conditions, so that the instrument reaches regions and sectors, treating the composition of layered adaptations as an open problem.
Require every AI-specific instrument arising from this inquiry to state at creation a fit-for-purpose re-decision trigger, responsible-officer powers to vary, suspend or retire, a recorded decision, retirement path, lifecycle state and change trail.
The committee recommends that every AI-specific instrument arising from this inquiry's report state, at its creation, a review trigger that requires the responsible officer to re-decide whether the instrument remains fit for purpose, with the power to vary, suspend or retire it, and to record that decision; a retirement path; a lifecycle state; and a change trail;
A program corrected obvious text-recognition errors in this quote for display.
Carry lifecycle requirements in Commonwealth AI policy and assurance arrangements; have the responsible portfolio keep a register of covered instruments and report annually against outcomes.
that the requirement be carried in the Commonwealth's AI policy and assurance arrangements; that the responsible portfolio keep a register of such instruments and report annually against outcomes;
A program corrected obvious text-recognition errors in this quote for display.
Treat material changes to instructions, thresholds and rule sets of deployed Commonwealth AI systems as rule changes, applying lifecycle discipline proportionately.
and that material changes to the instructions, thresholds and rule sets of deployed Commonwealth AI systems be treated as rule changes at a proportionate weight.
A program corrected obvious text-recognition errors in this quote for display.
At rule creation, specify review triggers anchored to observed change, authority to declare a rule broken, proposed/active/deprecated/retired states, and trails recording changes, approvals and evidence.
The repair is made at the point of rule creation or not at all, and it has four parts, the elements of a rule-change engine: 1. Triggers, events that compel a review, anchored to observed change in the world rather than to a calendar. 2. Authority, clarity on who may declare a rule broken. 3. States, so arule is proposed, active, deprecated or retired and never merely there. 4. Trails, so anyone can see when and why a rule changed, who approved it and on what evidence.
Use whether removing a rule makes productive activity harder as the retirement test.
And the test for retirement is stated once: does removing the rule make productive activity harder? If not, itis governance weight without governance value.
Focus research and development on layered-adaptation composition and a shared baseline evaluation suite.
Research and development belongs at the two open problems this submission names, the composition of layered adaptations (§3) and the shared baseline evaluation suite (Appendix B).
Treat mentoring as infrastructure to help people reach the 'Why' level and bring others up to it.
The scarce resource is not AI-literate workers but people who work at the Why level and can bring others up to it, which is why mentoring belongs in the infrastructure and not in the category of personal virtue.
Keep some tasks partially human to train future decision-makers and preserve institutional capability when machines fail.
Some tasks should stay partially human not because AI cannot do them, but because they are where the next generation of decision-makers learns the discipline that Why-level work requires. Oversight is a training ground: it is how an institution keeps the ability to operate when the machine fails, preserves its memory, and brings up the people who will still be able to think clearly tomorrow.
For AI-supported systems inside critical infrastructure (energy, water, health, finance), make redundancy across model sources with outputs compared a procurement requirement.
For the AI-supported systems inside critical infrastructure, energy, water, health, finance, the same discipline is a procurement requirement, redundancy across model sources with outputs compared
Require regulators to observe their own machinery when identifying and responding to emerging AI risks.
Clause (m) asks whether regulators, the AI Safety Institute and the intelligence and security community can identify and respond to emerging risks. The first requirement is self-observation: a regulator that cannot observe its own machinery cannot know whether it is responding, and a rule system without self-observation is blind to its own dysfunction.
Define monitoring and reconsideration as ongoing administration rather than one-off reviews, detecting AI capability advances as environmental change.
The second is continuity: monitoring and reconsideration defined as ongoing administrative activities, not one-off reviews, so that AI capability advance is treated as a category of environment change the system is built to detect.
Where an existing institutional home is inexact, designate another existing home rather than create a new body.
No new body is proposed. Where the committee judges an existing home inexact, the recommendation is that government designate one, not that one be built.
Read the original
In some quotes a program corrected obvious text-recognition errors for display.