A Government Report Was Reissued for Its Sources

A report was reissued twice, once over its own references.

The document is the Targeted Compliance Framework (TCF) Assurance Review Final Report, produced for the Australian Department of Employment and Workplace Relations (DEWR). The DEWR resource page that serves it carries one line above the download link: "This Report was updated on 3 February 2026 to address identified corrections. It replaces the report dated 26 September 2025 previously published on 3 October 2025." On the resource page and the departmental announcements fetched for this article on 26 September 2026, that sentence is the whole published explanation of the second reissue.

The lesson sits one level below the vendor and the tool. A report on the computer system which reduces and cancels social security payments was commissioned, drafted and published carrying citations and reference list entries which, in the report's own words, contained errors. Separately, the report's own Limitations page, describing the supplementary sources it footnotes and references, states: "We have not attempted to verify these sources independently unless otherwise noted within the Report." What the published words do not record is where in that sequence a reference check was meant to sit.

The fix is small, boring and testable. Treat every reference an assistant supplies as unverified until a named person opens the source, and make that opening a visible, dated step in the sign-off rather than an assumption buried inside the word review.

What was the review for, and what was it not?

The scope is on the record and it is narrow. The Secretary statement of 14 August 2025 describes it this way: "The review focused on the department's IT system, and whether rules coded into the system could be clearly traced back to the law and policies governing it."

The same statement records what the review did not do. It "did not consider systems in partner agencies, such as Services Australia, nor did it examine human decision-making under the framework". That matters for anyone tempted to read the report as a verdict on how people were treated. It is a verdict on whether code matched law.

The resource page records that the review was "undertaken by Deloitte from December 2024 to June 2025". The Secretary statement describes a review that "includes detailed assessments of business rules underpinning the operating system (with over 200 pages of detail)". The file served from the resource page when it was retrieved for this article on 26 September 2026 ran to 237 pages.

The Commonwealth Ombudsman report Automation in the Targeted Compliance Framework: when the law is changed but the system isn't (August 2025) records, on the DEWR Secretary's advice of 4 December 2024, that decisions to cancel income support had occurred automatically without consideration of a job seeker's circumstances, and that this affected 964 job seekers who "had their income support cancelled ... due to not meeting their mutual obligations". The report footnotes that figure to the Secretary's opening statement to Additional Senate Estimates of 26 February 2025. The same report records that "A further 45 job seekers received automated cancellations after the Secretary's decision to pause cancellations, when they should not have."

That is the subject matter of the document whose reference list had to be corrected: the machinery that stops payments to people who rely on them.

Two corrections, in the words of the publishers

The correction history runs across two artefacts and two rounds, and every description below belongs to the organisation that published it. The resource page records a Creation Date of 14 August 2025, the day the Secretary's statement in response to the review was published.

The Final Report carries its own update note on page two of the file served from the department download link. It states that "The Report has been updated to correct those citations and reference list entries which contained errors in the previously issued version" and, in the same sentence, to amend a case summary, being "the summary of the Amato proceeding which contained errors". That note describes the round dated 26 September 2025, which replaced a report dated 4 July 2025.

The companion artefact moved with it. The Statement of Assurance resource page records that "This Statement was updated on 26 September 2025 and replaces the Statement of Assurance dated 18 June 2025". The statement itself uses near-identical language, being updated "to correct those citations and reference list entries which contained errors in the previously issued version as well as revisions to improve clarity and readability".

Two documents. Same defect. Same day. Both notes also state that the updates "in no way impact or affect the substantive content, findings and recommendations" of the document.

Then the department put its own characterisation on the record. The Secretary statement on progress under the integrity assurance program, 3 October 2025 states: "Deloitte conducted this independent assurance review and has confirmed some footnotes and references were incorrect." The same statement notes that "There have been media reports indicating concerns about citation accuracies which were contained in these reports", and that a correct version had been released.

Four months after that release, the Final Report was reissued again, on 3 February 2026, "to address identified corrections".

A note on language, because it is part of the discipline. No departmental or firm page fetched for this article uses the words fabricated, invented or hallucinated. The words on the public record are "incorrect", "contained errors" and "identified corrections". Anything stronger is a characterisation someone else added, and repeating it as though the publisher said it would be the same class of error the corrections were made to fix. That applies to the web address of this article, which carries the word fabricated and did not get it from a publisher.

Where the AI disclosure actually sits

The report discloses a generative AI tool chain in its own methodology appendix. Appendix C, Detailed Methodology: Technical Review, at page 147 of the served file, describes a verification assessment of whether the state of the system code could be mapped back to business requirements, "which included the use of a generative artificial intelligence (AI) large language model (Azure OpenAI GPT 4o) based tool chain licensed by DEWR and hosted on DEWR's Azure tenancy." Section 6.3, Methodology, at page 48 of the same file, carries the same disclosure against the Traceability and Documentation Gaps item.

So there are two facts in one document: the references were corrected, the report was then reissued again to address identified corrections, and a generative AI tool chain was used in part of the technical review.

Nothing in the report joins them. No departmental page joins them. This article therefore does not join them either, and no reader should. Asserting that the tool produced the wrong references would be precisely the failure under discussion: a plausible claim carried forward with no source opened to support it.

Why do reference lists go unchecked?

The Ombudsman report supplies the general answer, in a finding about a different failure in the same program. Finding 4 is headed "Quality assurance activities did not identify the error", and the Ombudsman wrote of the department that "we could not be assured that its functional IT system tests check the compliance and consistency of the TCF computer system with the relevant laws."

An assurance activity that is never tested against the thing it claims to assure produces a tick and nothing else.

Apply that to citations. Where a review is specified, it is specified as reading for sense, structure, tone and risk. Opening the sources is rarely written into it as a step. Reference lists are the last thing added to a draft and the first thing skimmed in review, and they are skimmed for format rather than for existence.

Then add distributed responsibility. The drafting team assumes the reviewer will check, the reviewer assumes the quality process will check, and the quality process checks that a reference list is present and formatted. Nobody has lied, and the document still goes out unchecked.

A check that leaves no trace cannot be audited, and a check that cannot be audited will eventually stop happening without anyone noticing.

What does a reference check look like as an artefact?

It looks like a table, attached to the sign-off page, that a reviewer can be handed.

TheAICommand works to the Verified Draft Method: de-identify the inputs, ground the model in your own source material, keep a person at the decision point, verify against the primary source, and log what happened. The reference table is the log-what-happened half of that, made concrete.

The second row is the one that stops the sign-off, and it is not an exception to be waved through on a deadline. A table that is fully ticked and names nobody has not been checked.

Three rules keep the table honest.

• One named person per row, because a team cannot open a document.
• The date recorded is the date the source was opened, not the date the draft was reviewed.
• The sign-off names who checked, rather than stating that checking occurred.

A reference an assistant supplies is built to look correct: a plausible author, a plausible year, a plausible publisher, a title that reads exactly like the title such a document would have. The check that catches it is opening the thing, not reading it more closely.

The first prompt turns a finished draft into that table. Run it against the draft itself, with no search enabled and no external material in the context, so the model cannot quietly supply what it is being asked to expose. Redact the draft before you paste it, and use a tool your organisation has approved for material of that sensitivity.

You are helping me verify a draft document before it is signed out. Use only the text I paste below or the file I attach. Do not search, do not add sources, and do not fill any gap from memory.

If any personal name, internal file, matter or claim number, or any organisation other than the publisher of a cited source survives in the text I paste, list it at the end under Redaction missed, do not repeat it anywhere else in your output, and where a quoted claim contains one, quote the sentence with that item replaced by [REDACTED] and mark the row. Leave document numbers, compilation numbers, case citations, dates and URLs inside a cited reference exactly as written, because those are what the check opens.

Return one markdown table, one row per factual claim that carries a reference, with these columns in this order: Claim (quote the sentence), Reference as written, URL if the draft supplies one, Opened by, Date opened, Outcome. Leave the last three columns empty.

Then return a second table listing every factual claim in the draft that carries no reference at all. Do not suggest references for them. If a claim is ambiguous, put it in the second table and say in one line why.

Draft: [PASTE THE DRAFT OR ATTACH THE FILE]

What to check: that the row count matches the number of referenced claims you can find by hand in the first two pages, that no reference appears in the table which is absent from the draft, and that the second table is not empty when you know the draft contains assertions nobody sourced. If the model has invented a tidy reference for an unsourced claim, the control has failed at the first step and the output is unusable.

The second prompt is the one that changes reviewer behaviour, because it forces the test to be written before anyone opens anything.

Use only the table I paste below. Do not open anything, do not search, and do not assert whether any reference is real or unreal.

For each row, add a column called Test and write one sentence in this form: "For this claim to stand, [SOURCE_TITLE] must state [THE SPECIFIC THING], at or about [PAGE OR SECTION]."

Substitute the reference as written in the row for [SOURCE_TITLE]. Keep [THE SPECIFIC THING] and [PAGE OR SECTION] as your own words.

Where the reference as written is too vague to produce a test, write "Not testable as cited" and name what is missing: publisher, year, document number, page, or URL.

Do not rewrite the claims. Do not soften them. Do not add rows. Return the table only.

Table: [PASTE THE REFERENCE TABLE FROM THE PREVIOUS STEP]

What to check: that every test names a specific thing the source must say, rather than restating the claim in different words. A test reading "the source must support the claim" is not a test. A filled test reads: for this claim to stand, [SOURCE_TITLE] must state that the figure was [FIGURE] in the period, at or about page 4. Count the rows marked not testable as cited, because that count is your real exposure.

The line you will have to write if the check fails

The cost of a bad reference is easiest to understand as the sentence the document carries afterwards. Drafting that sentence before publication makes the risk legible to the people signing off.

Use only the facts I supply below. Do not invent dates, document numbers, findings or names.

Document: [DOCUMENT_TITLE], published by [ORGANISATION] on [DATE_PUBLISHED], version dated [DATE_VERSION].
Error found: [DESCRIBE THE ERROR IN ONE SENTENCE, WITHOUT NAMING ANY INDIVIDUAL].
Document set: [LIST EVERY OTHER ARTEFACT PUBLISHED WITH THIS DOCUMENT, TITLE AND VERSION DATE, ONE PER LINE].
Date the correction will issue: [DATE_CORRECTED].

Draft the version replacement line the reissued document will carry, in no more than two sentences, in this form: "This [DOCUMENT TYPE] was updated on [DATE_CORRECTED] to address identified corrections. It replaces the [DOCUMENT TYPE] dated [DATE_VERSION] previously published on [DATE_PUBLISHED]."

Then, from the document set I listed and nothing else, name each artefact that carries the same reference and will need its own replacement line. If I listed none, write No document set supplied.

Then draft, in no more than sixty words, the internal note recording which control failed and at which step. Name the control and the step, not a person.

What to check: that the replacement line contains only the three dates you supplied and no others, that every artefact it names is one you listed yourself, because an artefact you did not supply is an invention, and that the internal note names a step in your own process that a colleague would recognise. If the note blames a tool rather than a step, send it back, because a tool is not a control and cannot be fixed by a corrective action.

What the receiving side now expects

The department that received this work states that it has AI in Procurement Principles in place covering third party use of AI. The DEWR AI Transparency Statement, on a page the department records as last updated 27 February 2026, states that those principles "help to ensure" that where third party providers are using AI in delivering services to the department, "the usage is transparent to, and agreed to by the department, and that AI-aware quality assurance processes are in place". The same page states that "Human-in-the-loop mechanisms are embedded at critical stages to ensure accountability and accuracy in AI outputs" and that "AI is not used for automated decision making".

Read the middle clause again, because it is the institutional version of the table above. Human In The Loop (HITL) is not a claim about attitude. It is a claim about a step that exists, has an owner, and leaves evidence.

Above that sits the whole of government position. The Digital Transformation Agency publishes the Policy for the responsible use of AI in government, which states that "This version of the policy (v2.0) is effective 15 December 2025. The first version (v1.1) took effect on 1 September 2024." The policy states that it "applies to all non-corporate Commonwealth entities, with some exceptions", and sets out its mandatory requirements as separate linked pages. Version 2.0 was the version published on the policy page when it was checked for this article on 26 September 2026. Confirm the version line before citing it.

For a supplier delivering written work to DEWR, the department has published what it expects, so AI-aware quality assurance is a stated expectation rather than a private matter. A supplier who cannot describe that process in a sentence has an exposure it has not measured.

Where the framework itself stands

This is not closed history. The DEWR announcement Update on work to return the Targeted Compliance Framework to lawful administration, 4 August 2026 states that "Resuming these decisions on 26 October 2026 is contingent on successful completion of assurance activity and system testing ...". The same page records that other provisions are expected to resume in the first quarter of 2027, that compensation runs through the Scheme for Compensation for Detriment caused by Defective Administration, and that a further confirming statement was promised for October 2026. As at 26 September 2026 the department had published nothing after that update, so whether the provisions resumed, and whether the confirming statement has issued, sits on the departmental announcements page rather than in this article.

The announcement names sections 42AM and 42AG(1)(a) of the Social Security (Administration) Act 1999 (Cth) as the provisions to resume. Checked on 26 September 2026 against Compilation No. 163 on the Federal Register of Legislation, compilation date 5 December 2025, section 42AM is headed "Reconnection requirements for mutual obligation failures and work refusal failures" and section 42AG is headed "Compliance action for work refusal failures". Anyone citing those sections after 30 October 2026 opens the compilation header first and confirms the compilation has not moved.

Do this Monday

The artefact is a Reference Verification Table, appended to the sign-off page of one document, built exactly as set out above.

The owner is the person who signs the artefact out, not the drafting team and not the reviewer. The signature and the table belong to the same role, or the control has no teeth.

Steps 2 to 4 fit inside an hour once the draft is redacted, and nothing else changes about the document this week.

1. Take a document currently in draft that you are permitted to paste into your assistant, and redact it before you paste it.
2. Paste its reference list and its body into the first prompt and produce the table.
3. Run the second prompt over the five rows carrying the claims that would do the most damage if they were wrong, and record why those five, so the next reviewer knows what was not checked.
4. Open those five references yourself and fill in the opened-by and date-opened columns.

The hour worked if each of the five rows carries a written test, plus either a name and a date you can trace to the page you opened, or a blank nobody filled. It failed if all five rows are complete and none traces to a URL and a page, because the table then records an intention rather than a check, and the next hour goes into the same five rows.

The bottom line

A report about the system that cancels social security payments was corrected over its own citations, then reissued four months later to address identified corrections, and the published explanation for that second round runs to one sentence. The department and the firm described the problem in narrow, specific language, being incorrect footnotes and references, entries that contained errors, and identified corrections, and that language is the ceiling on what anyone else can responsibly claim. A generative AI tool chain is disclosed in the same report, in a methodology appendix, and no published source connects it to the citation errors, so nobody should. What transfers to every other desk is a control rather than a conclusion: a reference is unverified until a named person has opened the source on a recorded date. Build the table first, then argue about the tools.

This article is general information and education only. It is not legal, compliance, financial or professional advice.