Statement Summaries Need Source IDs, Not a Neat Story

A polished narrative can hide who said what, where it appears and what remains contested. A source-level assertion ledger keeps each proposition attached to the original statement, so AI assists evidence navigation while the authorised decision-maker retains every factual and credibility judgement.*

A statement summary should not give a delegate a neat story. It should give them a faster route back to the evidence.

The safer artefact is a source-level assertion ledger. Each row records one assertion, its source anchor, the speaker's knowledge type, linked material, corroboration status, conflicts and gaps. It does not score truthfulness or recommend which account to accept. The delegate reads the originals and makes every finding.

That distinction matters under the current Safety, Rehabilitation and Compensation Act 1988, Compilation No. 82 in force on 1 July 2026. The statutory and delegation framework assigns the function to authorised decision-makers. AI can make a large statement set navigable, but it cannot become an invisible witness, fact-finder or substitute decision-maker.

Why is a neat statement summary the wrong artefact?

Statements do different work. A speaker may describe an observed event, repeat another person's words, explain an understanding, or identify an inaccessible document. Compress that material and the boundaries between direct observation, reported account and inference start to disappear.

Comcare's April 2026 guide to submitting employee statements says a statement may add information that did not fit in the claim form or respond to a request for further information. It asks for factual and objective information, relevant dates, a helpful chronology and available supporting records. It also tells employees that the statement forms part of the claim file and will be released to the employer under section 59.

That guidance does not turn every claim into a statement exercise. Section 54 addresses the written claim and, subject to its exceptions, the medical certificate. A further statement is supplementary evidence. Section 58 provides a formal power to request relevant information or documents from the claimant when its conditions are met. A model-generated gap list is not a section 58 notice, it cannot decide that the power should be used, and it does not move any of the statutory clocks.

A conventional summary creates a verification tax. If it says, "The allocation changed after the meeting", the reviewer must rediscover who asserted it, which meeting was meant, whether the date was explicit and whether another account differed.

An assertion ledger is not another claim chronology. A chronology organises events by time. The ledger organises propositions by source, including assertions with no settled date. Time is one field, not its organising thesis.

What should the assertion ledger contain?

Use one row per material assertion. "The meeting occurred, [WITNESSA] attended, and duties changed afterwards" is three rows, not one. Atomic rows let a reviewer confirm one proposition without accepting the others.

A useful minimum schema is:

1. Assertion ID: a stable identifier such as A-014.
2. Source anchor: document ID, page and paragraph, for example S-01 p4 [12].
3. Speaker: a placeholder such as [CLAIMANTNAME] or [WITNESSA].
4. Assertion: a neutral, close paraphrase with no added conclusion.
5. Knowledge type: direct observation, reported account, inference, opinion or unclear.
6. Date status: exact, approximate, relative, disputed or not stated.
7. Linked source: the ID and anchor of a document or statement said to support or test the assertion.
8. Corroboration status: not checked, supported by identified source, partly supported, in tension with identified source, or no matching source located.
9. Conflict or gap: a precise question for human review, never a credibility conclusion.

Those statuses are descriptive. "No matching source located" does not mean false. "Supported" does not mean the account is accepted. "In tension" marks propositions that are not yet reconciled. The human decides what the evidence establishes and its weight.

Before using AI, replace direct identifiers with merge fields and assess indirect identifiers. The OAIC explains that removing names and addresses alone may not achieve de-identification; context and access affect re-identification risk, which is why de-identification is a contextual judgement rather than a find-and-replace. In a bank, insurer or superannuation fund, treat a rare role, small work location, distinctive product event or dated system incident as a possible indirect identifier. Use an approved environment and preserve the original outside the model workflow. The de-identification toolkit covers the discipline in full.

This prompt creates the first-pass ledger. A human must compare every row with the original statement before it enters the claim record or informs any further enquiry.

You are assisting with evidence navigation in a de-identified SRC Act claim.

Input: statements labelled with document IDs, page numbers and paragraph numbers.

Create one row per discrete assertion with these columns:
Assertion ID | source anchor | speaker | neutral assertion | knowledge type | date status | linked source named by speaker | corroboration status | conflict or gap.

Rules:
- Use only the supplied text.
- Preserve placeholders such as [CLAIMANT_NAME] and [DATE_1].
- Do not merge assertions from different speakers.
- Do not infer missing dates, motives, causation or legal conclusions.
- Do not score credibility, reliability, consistency or claim prospects.
- Use "not checked" until a specifically identified source has been compared.
- Quote no more than needed to distinguish the assertion.
- If a page or paragraph anchor is missing, write SOURCE ANCHOR REQUIRED.

Return the ledger and a separate list of rows requiring human source verification.

Fictional worked example: A de-identified file contains statements S-01 and W-01, meeting record M-02 and allocation record R-03. Four ledger rows show the method:

• A-001, S-01 p2 [6]: [CLAIMANTNAME] attended [MEETING1] on [DATE1]. Direct observation, supported by M-02 p1. Check that both sources identify the same meeting.
• A-002, S-01 p3 [9]: [MANAGERROLE] said duties would change the following week. Direct account, not checked. Look for a contemporaneous note.
• A-003, W-01 p2 [5]: [WITNESSA] understood the change would start after [EVENT1]. Inference, in tension with A-002. Check whether both accounts concern the same change.
• A-004, R-03 [ROWID]: the allocation changed on [DATE2]. Documentary record, partly supporting A-002. Check whether it shows approval, communication or only system entry.

The ledger exposes a question without resolving it. It also prevents R-03 from being presented as proof of what was said at [MEETING1]. That separation is the point.

How does a human test the ledger before using it?

Comcare's best-practice decision-making guidance says decisions need legal authority and relevant facts, with full and accurate records linking findings, evidence and the applicable provisions. It also asks whether sufficient factual and medical information is present and other relevant material is reasonably obtainable.

For Comcare, section 69(a) makes accurate and quick determinations a statutory function. Section 72 refers to equity, good conscience and the substantial merits, and says Comcare is not bound by the rules of evidence. Section 108E separately requires a licensee authorised to manage claims to determine them accurately and quickly and observe its licence conditions. None lets software choose facts.

Test the ledger in two directions. First, move from every row back to its stated source anchor. Second, move from every material paragraph in the original statement forward to a ledger row or an explicit "not material to this task" notation. The first test catches invention. The second catches omission.

This prompt tests conflicts and gaps after the first human source check. A human must decide whether each comparison is material, whether more information should be sought, and how any competing evidence is treated.

Compare the verified assertion ledger with the supplied de-identified source index.

Produce three lists only:
1. Assertions with a specifically identified supporting source.
2. Assertions in tension with another specifically identified source.
3. Material assertions for which no matching source has been located.

For every item, cite both assertion ID and source anchor. Describe the relationship neutrally. Do not decide credibility, prefer one account, infer motive, fill a gap, apply an SRC Act test or recommend an outcome. If the sources cannot be compared, state why.

Then apply five human controls:

• Open every anchor. Reject any row that does not resolve to the cited passage.
• Compare the neutral paraphrase with the speaker's actual level of certainty. "May have" cannot become "did".
• Check that absence has not been converted into contradiction. A witness who does not mention an event has not necessarily denied it.
• Separate source conflict from legal significance. A difference may be immaterial to the provision being considered.
• Record the reviewer's corrections, date and purpose. Do not overwrite the source statement or silently repair the AI output.

Tone analysis, sentiment labels and credibility percentages have no place here. Do not let stress, language differences, disability, cultural communication styles or ordinary variation in recollection become machine proxies for credibility. The workflow shows the authorised decision-maker where to read, not whom to believe. It belongs to the same family as the NWE evidence pack: AI prepares evidence for a decision it never makes.

Where the Privacy Act applies, APP 10 requires reasonable steps to ensure personal information used or disclosed is, having regard to the purpose of the use or disclosure, accurate, up to date, complete and relevant. The OAIC's commercial AI guidance identifies heightened accuracy risk in AI systems and calls for human oversight and testing. Source anchors, omission tests and recorded corrections turn those principles into operational controls.

Do this Monday

1. Select one closed, de-identified training file with two statements and at least two linked documents. Do not start with a live determination.
2. Assign immutable source IDs and page or paragraph anchors before using an approved AI tool or equivalent.
3. Run the extraction prompt, then require a second practitioner to complete the backward and forward source tests.
4. Record false additions, omitted assertions, broken anchors and changes in certainty. Treat these as test failures, not drafting preferences.
5. Approve a narrow use case only if the tool can reliably produce traceable rows. Keep credibility findings, legal analysis, information requests and determinations outside the automated step.

Bottom line

A neat summary is easy to read and hard to audit. An assertion ledger is deliberately less elegant because it preserves the boundary between source, proposition and human judgement. Use AI to reduce search time, not to manufacture a coherent account. The authorised decision-maker must read the originals, resolve material conflicts and own every factual finding.

This article is general information and education only. It is not legal advice, and it is not advice about any individual claim. Decisions under the Safety, Rehabilitation and Compensation Act 1988 are made by human decision-makers on the individual merits of each claim, and claimants have reconsideration and review rights in respect of determinations. Seek advice specific to your scheme and circumstances.

References

1. Federal Register of Legislation, Safety, Rehabilitation and Compensation Act 1988, Compilation No. 82, 1 July 2026: https://www.legislation.gov.au/C2004A03668/latest/text
2. Comcare, Guide to submitting employee statements in support of a workers' compensation claim, April 2026: https://www.comcare.gov.au/sites/default/files/docs/submitting-injured-worker-statements-guide.pdf
3. Comcare, Scheme guidance: Best-practice decision making under the SRC Act, updated April 2025: https://www.comcare.gov.au/scheme-legislation/src-act/guidance/best-practice-decision-making
4. Office of the Australian Information Commissioner, De-identification and the Privacy Act: https://www.oaic.gov.au/privacy/privacy-guidance-for-organisations-and-government-agencies/handling-personal-information/de-identification-and-the-privacy-act
5. Office of the Australian Information Commissioner, Guidance on privacy and the use of commercially available AI products, published 21 October 2024, last updated 17 January 2025: https://www.oaic.gov.au/privacy/privacy-guidance-for-organisations-and-government-agencies/guidance-on-privacy-and-the-use-of-commercially-available-ai-products
6. Federal Register of Legislation, Privacy Act 1988, Compilation No. 104, 4 June 2026: https://www.legislation.gov.au/C2004A03712/latest/text