NDIA Accountability Project

This project is in DRAFT

Submission 3177 — Three Primitives Research Lab — NDIS Future Generations Bill

‹ PrevPage 1 of 5 · Source p. 1Next ›

Submission to the Senate Community Affairs Legislation Committee

Inquiry into the National Disability Insurance Scheme Amendment (Securing the

NDIS for Future Generations) Bill 2026

Three Primitives Research Lab · Melbourne, Australia · ABN 68 932 608 853

Stacy Gildenston, Lead Architect · Pyrate Ruby Passell, Lead Developer

https://3primitives.io ·

v1.2 · July 2026

  1. Summary of position

  2. Three things make a decision about a human being legitimate: a declared purpose, a named human who holds authority, and operational constraints. These are not policy preferences. They are the structural requirements of any governed action, proven in a published, DOI-registered formal record (FR01, Canonical Logic Sequence, v2.2, December 2025). Remove any one and you do not have governance. You have a process performing governance without containing it.

  3. The I-CAN framework this Bill enables removes the second one. The computed budget is treated as the determination. Public reporting indicates that agency staff cannot meaningfully depart from the computed amount, and that the Administrative Review Tribunal will no longer have authority to alter plans or funding. No human holds authority at the point where consequences attach to a person. A published impossibility proof, FR02, AI Cannot Govern AI (v2.3, April 2026), proves this is not a risk to be managed. It is a structural defect: no computation can generate the permission to act, because permission is not a function of any system’s internal state. It depends on context outside the machine, the legal framework, the consent, the named human who answers for the act. The published record names this condition Ghost Authority, and it is not a metaphor. It is a proven structural defect with a published definition and a test.

  4. This is the architecture of Robodebt, and the Bill legislates it as reform. Robodebt had ministerial authority behind its framework, human designers behind its formula, and human staff in its offices. None of that placed a human decision-maker at the point where consequences attached to an individual. The Commonwealth has since agreed to the largest class action settlement in Australian legal history. The formal record proving this failure mode is structural was published before the rollout this Bill enables. Parliament legislates with that record in front of it.

  5. We do not oppose assessment tools, consistency, or automation as such. A computed output used as an input to a human decision is legitimate and often valuable. A computed output treated as the decision is Ghost Authority: authority exercised that no one holds. The Bill, as we understand it, does not merely permit that substitution. It removes the mechanisms, planner discretion and full merits review, by which the substitution could be corrected.

  1. We make five recommendations (Section 6). Each is a structural condition, not a safeguard bolted on afterwards: a named human decision-maker with genuine power to depart from the computed figure, recorded per decision; automated outputs classified in law as inputs, never decisions; a gate that fails closed where no human authority is declared; an immutable audit record identifying the decision-maker for every determination; and full merits review preserved.

  2. About the submitters

  3. Three Primitives Research Lab is a two-person independent research lab in Melbourne, Australia. It publishes a corpus of thirteen formal records (FR01 to FR13) on the structural requirements of governed automated action, under CC BY 4.0 with DOI registration. The corpus proves what any governed action requires: a declaration of purpose, a named authority holder, and operational constraints, and it proves what happens when a computation is placed where the named authority must stand. The main public source is https://3primitives.io.

  4. Stacy Gildenston has thirty years of experience in virtual governance (1995 to 2025) and is a World Summit on the Information Society Award winner (2003). Pyrate Ruby Passell is a CERN Open Quantum Institute Friend and an ITU Citiverse Mentor, and is first author of the corpus result establishing that action-mapping is non-unique and therefore requires explicit declared authority. The lab has no clients, funders, or commercial relationships in the disability sector.

  5. The structural defect the Bill legislates

  6. The core result needs no symbols. Decisions entail responsibility. Responsibility requires agency. Protocols, whether a spreadsheet formula, an actuarial model, or a frontier AI system, lack agency and cannot bear responsibility. When a protocol is treated as the source of a decision, agency is displaced from every human actor, and authority is exercised that no one holds. The published record names this condition Ghost Authority and gives it a test:

At the point where consequences attach, if no human explicitly declares ‘I decide this,’ the decision lacks legitimate authority.

FR05, Ghost Authority Lemma, Corollary (the Ownership Test)

  1. Underneath the plain language sits a formal impossibility result, FR02: there exists no total computable function capable of autonomously resolving authorization without an exogenous declaration. Compliance with a specification is computable. Permission to act is not, because it is not a property of the system at all. The two are different objects. Every architecture that computes a figure and treats the figure as the determination has substituted the computable object for the one that must be held by a person. The defect is present at design time, and no improvement to the accuracy of the tool removes it, because the tool was never the problem. Its job description was.

  2. One defense should be answered before it is raised: a human designed the framework, a Minister made the instrument, therefore humans decided. No. A standing rule is a constraint on action, not an authorization of this determination, at this time, affecting this person. The two are legally and structurally distinct. Robodebt had ministerial authority behind its framework, human designers behind its formula, and human staff in its offices. None of that placed a human decision-maker at the

point where consequences attached to an individual, and that absence is what the Prygodicz settlement and a Royal Commission spent volumes trying to locate.

  1. Application to the Bill and the New Framework Planning arrangements

  2. Applying the Ownership Test to the arrangements this Bill supports, as publicly described:

  3. The budget model engine. I-CAN assessment data will feed an algorithmic system that generates funding amounts. Reporting in March 2026 indicated that agency staff directly involved in implementation formally escalated concerns about this engine to critical risk status. Our submission does not depend on whether the engine is accurate. Grant that it computes perfectly, every time: the defect remains, because the computed figure is being treated as the determination of a person’s support, and no computation can hold that authority.

  4. Approval as costume. The government’s assurance is that trained staff, not computers, approve plans. The Ownership Test asks a sharper question: can the staff member who approves a plan meaningfully depart from the computed budget? If public reporting is correct that they cannot, and that the remedy for an unsuitable result is only another computer-generated assessment, then the approval is not a decision. It is a costume. A signature on an output the signer cannot alter attributes authority without conferring it, and it will not survive the first serious examination of who decided.

  5. Review without authority. The same reporting indicates the Administrative Review Tribunal will no longer have authority to alter plans or funding. Merits review is the mechanism by which a human authority can be located after the fact. Removing it does not merely reduce a participant’s remedies. It completes the architecture: no human decides at the point of determination, and no human can be reached afterwards. The Committee’s own interim process has already recognised part of this problem by securing greater transparency requirements around automated decision-making. Transparency is necessary and insufficient. A transparent computation deciding a person’s support is still a computation deciding a person’s support.

  6. Thresholds by instrument. The Bill’s reliance on ministerial instruments to set assessment thresholds and reduce funding across categories of supports compounds the defect. It moves the standing rules further from the individual determination while the individual determination is simultaneously handed to a computation. The distance between the last genuine human decision and the person affected by it grows at both ends.

  7. Foreseeability: the record predates the rollout

  8. Robodebt examined this architecture at national scale: a computation treated as the source of hundreds of thousands of decisions about human beings, with human approval present in form and absent in substance. The Commonwealth has since agreed to the largest class action settlement in Australian legal history, and a Royal Commission made 57 recommendations. The question that inquiry struggled to answer, who decided, is the question the Ownership Test is built to ask in advance.

  9. The formal record proving this failure mode is structural, not emergent, was published with DOI registration before the deployments it describes: the canonical logic sequence in December 2025,

the impossibility proof and Ghost Authority Lemma in current form in April 2026, the completed corpus in May 2026. The Bill was introduced on 14 May 2026. When harm follows a computed determination under this framework, the defense of unforeseeable error will have to answer a published theorem that predates the rollout and a parliamentary record showing the Parliament was told. This submission places that record before the Committee so that the dates are not in dispute later.

  1. A structural alternative exists and is specified in the same public corpus. FR01 defines what must cross any gate before a governed action executes: a declaration naming the bounded purpose, the live human authority, and the constraints, issued prior to execution and immutable for that event, with every gated action leaving an audit record. No declaration, no action: the gate fails closed. The lab has implemented and tested this architecture, documented at https://3primitives.io/uss. Whether government systems adopt this implementation or another, the structural requirements it satisfies are the minimum for any lawful determination, and they are publicly available. ‘Nothing better was possible’ is not available as a defense.

  2. Recommendations

  3. Recommendation 1. The Bill be amended to require that every determination of a participant’s plan, budget, eligibility, or supports be made by a named, identified human decision-maker with genuine authority to depart from any computed or assessed figure, and that this authority be exercisable in substance, not merely recorded in form.

  4. Recommendation 2. The Bill classify all outputs of assessment tools, budget models, and any automated or algorithmic system as inputs to a decision, and prohibit any provision, instrument, or administrative arrangement that treats such an output as a determination or confines the human role to accepting the output or requesting its regeneration.

  5. Recommendation 3. Where no human decision-maker has declared authority for a determination, the determination must not take effect. The system fails closed to a human decision, never open to a computed one.

  6. Recommendation 4. Every determination carry an immutable audit record identifying the human decision-maker by role and name, the computed inputs considered, and any departure from them, available to the participant and to review bodies.

  7. Recommendation 5. Full merits review by the Administrative Review Tribunal, including the authority to alter plans and funding, be preserved without limitation. Review that cannot alter the outcome is not review.

  8. These amendments do not slow automation of what is legitimately computable. They restore the one element no computation can supply: a human who answers for the act. That element is not a policy preference. It is a structural requirement of governance, proven in the published record, demonstrated at national cost in Robodebt, and absent from this Bill as introduced.

  1. Attachments

  2. We consent to this submission being published.

  3. Attachment A: Negligence by Design, white paper, v1.1, July 2026 (Three Primitives Research Lab).

  4. Attachment B: The Second Line of Defence, white paper, v1.1, 2026 (Three Primitives Research Lab).

  5. Formal Records cited: FR01, Canonical Logic Sequence (v2.2) · FR02, AI Cannot Govern AI (v2.3) · FR05, Ghost Authority Lemma (v2.3). All records: https://3primitives.io/formal_records

Stacy Gildenston

Lead Architect, Three Primitives Research Lab

WSIS Award winner (2003) · 30 years virtual governance (1995 to 2025)

Pyrate Ruby Passell

Lead Developer, Three Primitives Research Lab

CERN OQI Friend · ITU Citiverse Mentor

Melbourne, Australia · https://3primitives.io ·