It Looked Like Serendipity. Then We Checked the File. — The Investigative Loop
It Looked Like Serendipity.
Then We Checked the File.
Memory, source pages, arithmetic and correction: how the Rent Assistance question became a better investigation.
There’s a moment when a number jumps out and you think: there it is. That’s the thing. Then you have to check whether it actually is.
That is what happened with the Rent Assistance email. I remembered the accommodation, the delays, the money I did not recall receiving. The file contained $430.80 in arrears and a later $2,840.61 overpayment entry. It looked like a collision between lived reality and the database.
So we put the actual LEX 93986 release under the magnifying glass. Not the previous blog post. Not another AI paragraph repeating the same claim. The file.
The person supplies the question
An agency record describes a person from the inside of a system: entries, dates, codes and adjustments. I approach it from the other direction. Where was I staying? What did I tell them? What money do I remember arriving? What was happening outside the office?
My current account places the start at BP Keith on 12 December 2025. The agency’s March note records 19 December. Those are two sources to compare, not permission to silently replace one with the other. Memory can identify the missing week; contemporary records must help establish it.
The same applies to Cages Road, Aratula and Collaroy. My account gives the investigation a chronology to test. Calling the whole chronology verified before testing it would make the starting point look like the conclusion.
The file changes the question
The first hypothesis was that $430.80 might be the whole backpayment for a much longer period. The source check found something we had to carry forward: on 23 February the OMC recorded Rent Assistance as paying. On 3 March, the OMC recorded that I had updated within time and decided to pay arrears back to 19 December.
That does not prove every payment arrived. It does mean we cannot responsibly say that $430.80 was necessarily the only arrears ever released. The sharper question is now: where is the implementation record for the March backdating decision?
If the evidence changes the question, change the question.
That is progress, even when the first theory was more dramatic.
The arithmetic gets its own check
The later $2,840.61 figure looked mysterious when reduced to an average daily amount. Split the period at September indexation and the mystery changes.
| Check | Result | Limit |
|---|---|---|
| $430.80 ÷ $215.40 | Exactly two fortnights, equivalent to 28 days at that rate. | Does not identify which dates were paid. |
| 168 days at $219.40/fortnight, plus 13 days at $223.80/fortnight | $2,840.61 after rounding. | Reproduces a possible calculation; does not prove payment or recoverability. |
| 12 December to 25 February | 75 elapsed days; 76 if both endpoints count. | The correct entitlement period still needs a decision and evidence. |
| 19 December to 25 February | 68 elapsed days; 69 if both endpoints count. | A date count is not a complete entitlement calculation. |
The calculation supports a precise reconstruction using published rates. It does not turn a case-note entry into a bank reconciliation. Those are different jobs.
The loop we can repeat
- Start with the experience. Write the discrepancy as a question, with the recollection attributed.
- Locate the original record. Keep the release reference, stamped page, entry date and surrounding text.
- Read around the apparent breakthrough. The later March decision matters as much as the earlier $430.80 entry.
- Calculate explicitly. State rates, periods, endpoint conventions and rounding.
- Try an alternative explanation. Could there have been an additional payment, an old file-review assumption, or a different payment period?
- Publish the correction with the finding. Do not make readers reconstruct what changed.
- Ask for the next decisive record. Here, that is the component ledger and the bank-payment trail.
Where the AI helps—and where it needs checking
AI helps search, structure, compare, draft and code the reading interface. A calculation can be executed and reproduced separately from the prose. A page can be rendered to check that extracted text belongs to the right entry.
That does not make a language model a deterministic compiler or give it zero-drift citations. Repeating the same corpus across models does not create independent documentary evidence. A second model can inherit the same mistake. The checks become stronger when they use different evidence or methods: original page, exact calculation, published rate table, actual transaction.
This audit used model-assisted reading, PDF text extraction, a rendered source page, Python arithmetic and official rate sources. It was not a separately delegated multi-agent investigation. The mention of Apple Notes in the process discussion describes the writer’s own record-keeping; those notes were not inspected in this audit.
Keep the record and the interpretation apart
There are at least four things on the table: what I remember, what an officer recorded, what a calculation reproduces, and what a bank statement shows. They may support one another. They may conflict. They should remain identifiable.
The August social-work entry is a good example. It describes a file review of accommodation information. It should be compared with October’s landlord account. It should not be rewritten as proof that someone visited and verified my accommodation in August.
The formal addendum is part of the process too. It records the questions I put to the agency. Sending it does not make every premise in it proven. If the file check reveals a qualification, that qualification belongs in the public account and any later follow-up.
What was actually achieved
We did not establish a missing dollar total. We identified a specific recorded backdating decision and a specific payment trail still needed to test whether it was carried out. We also reproduced the later figure rather than leaving it wrapped in a mysterious average rate.
That is useful work. The next question can be answered. It has a date, a page and a requested record.
The file says I updated within time.
The file says arrears were to go back to December.
Show me what was actually paid.