AI Application / Document Review

The AI Revised It Ten Times and It Is Still Wrong. Now What?: Cleaning the Data Is Only the First Layer

Same material, same model, same question: the tenth run usually differs from the second only in wording. What you need to change is the layer, not the number of tries.

What this article covers

A participant at the talk asked a very good question. He had made a 150-page training deck at work, first using one model to check the content, structure and chapter order repeatedly, revising it more than ten times until he thought it was close. Then he had another model review it once, and it turned up a standards error in the content that the first model had missed even with deep thinking switched on. He later traced it to the difference between two international standards.

"Both models I used were new, I used deep thinking, and I did it over ten times, and there were still differences. So this should prove you cannot trust just one company's AI, or rather, each company's capabilities and emphases may differ."Participant question at the talk, 2026-08-02

He did it more than ten times, both on the latest models, both with deep thinking on, and there was still a gap. Judging by this case, the problem looks more like the design of material, instructions and mechanism than something a stronger model would fix. Though one case cannot completely rule out model capability either.

My reading is this: when long documents go wrong, it is usually because three layers of design were never done, and most people only did the first, sometimes not even that. This article opens up all three, and gives you one passage of prompting that works better than "run it again".

Who this is for
  • People who need an AI to check long documents: training materials, reports, standards, contracts, tenders
  • People who have already switched models and turned on deep thinking and still do not feel confident
  • People who want to turn a one-off check into a workflow they can reuse next time
What you can take away
  • The order of the three checking layers and the criteria for each
  • How two slicing methods cover each other, and why one is not enough
  • A copy-ready prompt for discussing with the AI why it got something wrong
Sections five and seven are not limited to long documentsAnyone who goes back and forth with an AI on revisions can use them.

1. Why running it ten more times does not help

An uncomfortable fact first: with the same material, the same model and the same question, the tenth run usually does not differ from the second in any meaningful way. The wording will differ each time, and occasionally it will catch one or two extra things, but a blind spot at a given layer does not disappear because you ran it more times. There are three reasons, and they sit at different layers.

Layer oneMaterial

What the AI reads is not what you see. This is what data cleaning solves.

Layer twoSlicing

Read too much at once and it skips. This is what chunking solves.

Layer threePerspective

The same model carries the same set of preferences. What it considered right, it still considers right on review. This is what switching vendors solves.

Running it again just circles inside the same layer. Ten runs later, the material is still that hard-to-read material, the slicing is still that swallow-it-whole slicing, and the perspective is still the same perspective.

The order mattersFirst make the material something it can read, then decide how to slice it, and only then find a second perspective. Skip the first two layers and go straight to a second vendor, and all you get is two check reports built on the same misreading.

2. Layer one, material: confirm what it reads matches what you see

There are already two articles on the site covering this layer thoroughly, so here are the links:

So rather than repeat that, here are three points that bear directly on reviewing.

Serious standards material gets cleaned first

Training materials, regulations, standards, contracts: get one version number or one standard code wrong and the whole conclusion goes with it, and it goes wrong in a way that looks right. For this kind of content I always suggest converting to a plain-text Markdown file first, pulling the chapter structure and a content index out together, and then handing it to the AI. What comes out is a searchable copy for the AI to read, and the original has to be kept: tables, text inside images and footnotes are the things most likely to be lost in conversion, which is why the sampling check below goes back to the original to compare.

"Of the cases I have seen where documents produce a lot of hallucination, 90% of people see enormous improvement after cleaning the data."Live at the talk, 2026-08-02

Do one sampling check after cleaning

Do not start reviewing the moment cleaning finishes. Pick three to five spots at random and compare them against the original: did a table skip a row, did a footnote get swallowed, did the text inside an image vanish entirely. Five minutes here saves ten wasted runs later.

Make "where it came from" part of the rule

This is what the participant's case suggested to me: since what was wrong was the standard version, then require it to cite sources in the checking rule.

Wherever you mention a standard, a specification, a statute or a version number, add after the sentence: which document it came from, which chapter or section, which version. Anywhere you cannot find a source in the original material, do not write it as a statement of fact. Change it to "this point has no clear basis in the material" and list it for me.

What this does is turn things it says smoothly into things it must produce evidence for. Claims without a source surface by themselves.

But this does not stop the error in the case at the top. That participant's deck had two international standards mixed up. Both standards genuinely exist and both are citable, so when it cited the wrong one it could still show you a source, and it looked entirely reasonable. To stop that kind of error, the rule needs one more layer: specify which standard and which version this document is subject to, and then require it to compare every citation against that line.

The standard this document is subject to is: {your standard name and version, e.g. XXX Standard 2023 edition}. Every time you cite a standard, compare it against the line above first: 1. Matches → write it as normal, with the chapter cited afterwards 2. Does not match (cites another standard or another version) → mark it [VERSION MISMATCH] and list it 3. The material does not state which one applies → mark it [VERSION TO CONFIRM]. Do not pick one yourself At the end, give me one list: which places are [VERSION MISMATCH] and which are [VERSION TO CONFIRM].

The key to this layer is the third case. Most standards errors are not invented out of thin air. They are the wrong choice between two things that both hold, and when it chooses, it does not tell you it made a choice. What you have to do is turn "choosing" into an action it must report.

3. Layer two, slicing, and not just one way

The problem at layer two is that when it reads too much at once, it starts skimming.

Throw a hundred-plus pages in whole and the beginning reads clearly, but by the later parts it starts skipping. So chunk it. The real key is the next sentence: do not chunk it the same way twice.

"Next time ask it to try checking in chunks, to avoid blowing up the context. Chunk it, and do not use the same way of chunking, use two or three different ways, and that basically makes things much better."Live at the talk, 2026-08-02
Slice one: by chapter

Ten chapters means ten passes, one analysis per chapter, then one big consolidation of the ten results at the end.

Blind spot: cross-chapter contradictions are invisible, for example chapter two and chapter eight describing the same thing differently.

Slice two: by theme

First have it group the ten chapters into five themes, then run one pass per theme.

Blind spot: ordering problems inside a chapter are invisible, for example a build-up from simple to complex that has been broken.

The two slicing methods cover different areas, and only overlapping them gives you coverage. This is what "use two or more different ways of chunking" actually means. It does not mean running the same slicing twice.

Round one: this document has {chapter count} chapters. Analyse one chapter at a time, report problems for each chapter separately, then do one cross-chapter consolidation. Round two: first regroup these chapters into {theme count} themes, tell me how you grouped them, then analyse once per theme. If the two rounds conflict, list the conflicts. Do not pick one yourself.
That last line matters mostIf you do not tell it to list conflicts, it will pick one for you, and you will never know it chose.

4. Layer three, perspective: have both vendors run once, then cross them

Layer three is where you actually switch models.

"Have it check chapter by chapter, then cross-multiply with Claude, cross-multiply with GPT, then cross the Claude and GPT results with each other, and that basically makes problems very unlikely."Live at the talk, 2026-08-02

The method: run each slicing method through both vendors once, then cross the two sets of results. That leaves you with four sets of results, and what survives the crossing as disagreement is usually where a human genuinely needs to judge.

That participant's experience is the proof of this layer: both models on the latest version, both with deep thinking, more than ten repetitions, and still a gap.

My own addition: different models have different weightings, so they have different blind spots. This is not a question of which is stronger. It is that they care about different things.

What if you only have one vendorPlenty of environments have this constraint: the company only bought one, security policy only allows a specific service. Layer three is out of reach there, but there is still a substitute: with the same model, open two clean new conversations, run one with a "slice by chapter" role setup and one with a "slice by theme, only flag contradictions" role setup, then hand both sets of results to a third conversation to compare. To be honest, this is weaker than crossing vendors, because both sides still share the same preferences. It catches omissions but not a bias the model holds consistently. Treat it as reinforcement, not an equivalent substitute.
Would three vendors be saferI did the arithmetic and gave up. If every vendor has to review every other vendor, A reviews B then C, B reviews A then C, C reviews A and B, and then you run the whole thing again, the workload multiplies several times over and the time is not there either. In my experience two vendors crossing each other handle the large majority of problems. That is a judgement from experience, not a statistic.

5. More effective than one more run: discuss why it got it wrong

This is the step I think most people miss.

After finding an error, most people's reaction is "check it again". But the error has already happened, and running it again is just another roll of the dice. What works better is discussing the thing itself with it directly.

What I read out in class was this passage:

Why did you produce incorrect information about the standard? How do we make sure the information is correct? Suppose I am willing to spend more time and more money, to have you re-check a second time, or verify from a different angle on a second pass, and cross-check with another AI on a third. I also have two model setups I can combine. Would you work with me to design a review mechanism, a debugging mechanism, or an information-tracking mechanism, to make sure everything is correct?

This passage does two things.

  1. It moves the problem up a layer: from "this one was wrong" to "we need to design a process that does not go wrong again". What you get back is not one correction. It is a checking design you can keep.
  2. It draws out limits it knows about but will not mention unless asked: which section it was not actually confident about, which comparison it cannot do, which content needs external data to verify.

6. Three sources of error: work out which one this is first

After the discussion you will need to decide which kind of problem this is, because all three are handled completely differently.

"Some things are because we did not word the instruction well, some things are because we did not set the mechanism up properly, and some things may be beyond the current ceiling of the model's ability."Live at the talk, 2026-08-02
Source 01Instruction not worded well

Sign: it is doing something different from what you wanted, but doing it earnestly.
Fix: write the criteria, scope and output format clearly.

Source 02Mechanism not set up

Sign: it got this one right, but misses the same thing next time.
Fix: write the checkpoints, the order and the stop condition into a rule.

Source 03Beyond the model's ceiling

Sign: switching model, slicing and phrasing all produce the same error.
Fix: have a human review this part, or use external data that can be verified.

That participant's 150-page deck does not look like it exceeded the model's reading range, so I judge it as falling into the first two, a problem of instruction and mechanism. One thing to add: page count does not convert directly into how much a model can take in. What actually matters is the volume of text and the formatting after cleaning, so this is a judgement about this case, not a general rule.

Getting this wrong wastes workSwitch to a stronger model for a mechanism problem and even the most expensive one still misses it. Keep rewriting the instruction for a capability problem and you can rewrite until dawn without getting anywhere.

7. How many rounds before you stop

Easier to miss than "how to review" is this: how many revisions before you stop.

If you only tell it "revise according to my standard" with no upper limit, it will keep revising until it thinks the result is good. That usually goes one of two ways: the bill keeps climbing, or the result drifts further off.

My work is mainly document analysis and reasoning, and my rule of thumb is two to three rounds. One round here means:

  • Round one: A reviews B, and B reviews A
  • Round two: A revises once following B's suggestions, B revises once following A's, and then they review each other again
"What is clean after two or three rounds is clean, and what has a problem does not get better either... if two rounds have not solved it, do not sit there, because ten rounds or a hundred rounds are not going to solve it."Live at the talk, 2026-08-02

There is a simple way to judge that it is time to stop: when the same problem sticks in the same place two or three times in a row, stop and report. Do not keep running.

After stopping, what you do is go back and look at what the rule left out. Usually it is one of these three:

  1. No definition of the criterion: you never said what "correct" means.
  2. No rule for exceptions: you never said what to do when a situation is not covered.
  3. No requirement for sources: you never asked it to show why something is right.

8. Last step: turn this into a process it runs on its own

At this point you have a complete checking design: cleaning and sampling, two slicing methods, two vendors crossing, discussing the mechanism, judging the three sources, stop conditions.

All of that is still "you walking it through each time". The last step is handing it over.

Judging the moment is simple: when you find this standard has run once or twice with no problems, you can tell it:

"You know my standard, you should let yourself run it. I want you to be able to review and run to this standard yourself next time, when I am not beside you. That is what loop engineering means."Live at the talk, 2026-08-02

If it can, that is loop engineering. When it cannot, it is usually because it does not know what to do when something is wrong. So go back and add "how to notice it went wrong, how to fix it, when to stop" to the rule.

What that participant said at the end of the class stuck with me: once it is confirmed OK, turn it into a skill package. That order is right. Get the process running smoothly first, then save the smooth process, rather than trying to build a perfect skill package from the start.

How you can start

Today Sample-check your cleaning results

Take the document the AI most often misreads, pick three spots at random and compare them against the original. You will find out quickly whether the problem is in the material or the analysis.

This week Run both slicing methods once

Run the same document by chapter and by theme, and list the conflicts between them. Do not have it choose for you. The conflict list is itself the value.

After those two Write sources and stop conditions into the rule

Then open a new conversation, say only the sentence you would normally say, and test whether it starts up on its own.

None of the three steps needs programming. What you need is to state clearly what you check every single time, and that is something only you can do.

Three related articles

AI WorkflowCross-Vendor ReviewData OrganisationKnowledge ManagementFoundationalPublished 2026-08-05