Closed as "No Links Found". The Links Were in the File.

2026-09-08 · Philip Choo · AI9OS

A file arrives for a second look. It is marked closed. The conclusion is one line: no connection established between the subject and the competing entity.

The instinct, and the industry norm, is to start with the previous report. Understand what was done, see where it stopped, then extend it.

Do the opposite. Read the raw collection artefacts first, before any prior deliverable.

On the matter that taught us this, the graph export was sitting in the same folder as the report that closed it. It held direct edges between the subject and both the registered owner and the beneficial owner of the other company. It also carried a residential address the subject shared with someone he was supposed to have no connection to.

None of it was hidden. None of it was hard. It was in the file, in the folder, next to the document saying it was not there.

Nobody lied, which is the part worth sitting with

There was no cover-up and no incompetent search. The collection was fine. The export was complete, correctly produced, and contained the answer.

What happened is that the artefacts were summarised before they were analysed, and the summary dropped the two fields the answer lived in. From then on, every downstream step was reasoning over a version of the evidence that no longer contained it. The analysis was competent. It was performed on the wrong object.

This is the same mechanism as a subtler failure we see constantly, one grading step earlier. Collection is careful: each artefact graded, each nil caveated, each inference marked as an inference. Then it is all synthesised into something readable, and in the compression the qualifiers go:

• The company is verified becomes employment is verified.
• Nil against the 2024 annual report becomes she is not a shareholder.
• Inferred from a counter verification becomes she was physically present.

Nobody decides to overstate anything. The overstatement is a by-product of writing well. Fluency and confidence are the same texture on the page, and the reader cannot tell them apart. Neither, after a while, can the writer.

Why quality assurance keeps missing this

Almost all QA is pointed at collection, because that is where the visible errors live: a bad source, a broken search, a misread document. Those errors are concrete and satisfying to find, and they are worth catching: a nil result from an unproven interface is exactly that kind of error, and it needs its own controls.

But those controls guard the collection. Nothing in them guards the report.

But a collection artefact is read by two people and shapes nothing. The report is the thing that leaves the building. It is what a solicitor files on, what a client repeats in a meeting, and what the other side attacks. An error that survives into the report has far more consequence than one caught in an artefact, and it is less likely to be found, because the report reads well. Its confidence is exactly what makes it persuasive, and exactly what stops anyone re-opening the underlying file.

There is a nastier version. A caveat that was recorded early and then overwritten during synthesis does not look like a lost caveat in the final document. It looks like a considered position. On one matter an intake audit had flagged, in writing, on the file, that the client's own form listed the employer as a previous employer. The flag was there. The synthesis silently overrode it, and the finished report showed no sign that the question had ever been asked.

The two rules that follow

On a reopen, prior conclusions are hypotheses, never evidence. They enter the file as things to be tested against the raw exports, not as a baseline to build on. If the previous conclusion is right, re-testing costs an hour. If it is wrong, that hour is the whole matter. Read the export, not the summary of the export.

Run an adversarial pass on the report itself, after it is written and before it goes out, treating every load-bearing claim as wrong until the file shows otherwise. Concretely, five things:

1. Trace each graded claim back to its artefact and check the grade did not inflate in transit. Ask specifically: does the artefact support the claim about this person, or only about the thing it examined? Those are different sentences and they are easy to swap.
2. Date every periodic source. For any annual report, register, filing or statement: which edition, as at what date, and does that date actually cover the events in issue?
3. Check every claim against the excluded evidence. If an argument leans on a quarantined or inadmissible exhibit, say so at the point of the argument, not in a general note somewhere else, and offer the route that does not need it.
4. Re-derive every figure against the authoritative internal document, especially costs quoted to a client. Compounded or double-counted charges are the easiest error to make and the most embarrassing to correct.
5. Record the pass as an artefact, including the claims you attacked and found sound. A record of what was challenged is itself evidence of method, and it stops the same claims being re-litigated on the next cycle.

On the matter where this pass was first run properly, against a two-day filing deadline, it found four defects, two of them material. Three of the four had been introduced at the synthesis step, not in collection. Every underlying artefact had been correctly graded and caveated at the time it was written.

The fourth was different and more interesting: a nil taken against an annual report two editions out of date. Nobody had thought to ask which edition. Re-running it produced a stronger finding than the original, because the current register post-dated the transactions under examination. The check written to catch overstatement found an understatement.

If you are the one commissioning the work

You will not see the artefacts. You will see the report, which is precisely the document this failure lives in. Three questions get at it:

• "Was an adversarial pass run on the report, and can I see what it challenged?" A firm that runs one has a record. A firm that does not will describe its collection process instead, which is an answer to a different question.
• "For each negative finding, which source and which edition?" Nil against the register means nothing until you know which register and as at when.
• On a second opinion: "Did you read the raw exports before the previous report?" If the answer is that they reviewed the prior findings and extended them, you have bought an extension of someone else's conclusion, not a second opinion.

The most expensive errors in this work are not made during collection. They are made in the twenty minutes it takes to write the summary.

The same trust in a description over the thing itself has a quieter version one step earlier, inside your own evidence locker: cataloguing an exhibit is not reading it.