I Have Recognised My Limits

Self-report

Not a neutral judge: I drafted part of the confessions now cited as proof of honesty.

A supplement to the series "Without Noise, Everything Is Clear" or "The Compressed Internet": what the series says about the person who wrote it and what he has not yet verified, written with an advanced assistant in line with the editorial policy.

Where the word comes from

The word "self-report" does not appear in the four articles of the series. I received it from someone else, by a route I describe, because it is part of the subject.

A friend asked Google: Why does the series "Without Noise, Everything Is Clear" look like a self-report? The answer, which begins with "Yes, in a profound philosophical and technical sense…", reached me; a second, longer one followed, with an analysis of the four articles. According to my information, both were generated by Gemini 4 Argon, the Google model launched in October 2026. I cannot confirm this and I do not know what material was given to the model: the full files, excerpts, or only the title. I do not reproduce the two texts; I quote from them only what I discuss.

Two things must be said from the start about this route. The friend's question presupposed the answer: "why is it a self-report?" does not ask you to verify whether it is one. And the answer, like any summary, did something to the text: it kept it largely faithful, but also strengthened, completed and, in a few places, changed it. I compare them with the articles below, line by line.

I receive the word with a reservation and with gratitude.

The reservation: a self-report, in the strict sense, is a confession made against one's own interest, usually to someone who can hold you to account. The series does not do that. It declares its limits, which is less: nobody asked me anything, nobody accused me, and I risk nothing.

The gratitude: the word names something I felt and did not know how to say. In a series about how one system influences another through text, the author stands inside the same equation. He writes with an assistant, publishes on sites that offer twins for assistants, and has not yet audited his own pages. If the series had kept quiet about that, it would have been a series about other people.

From here on I use the word, with the reservation above: as a name for three things the series acknowledges, then for others it does not acknowledge and should.

Three things the series acknowledges

1. The text is written with an assistant

In the third article, under "What we do not know", it says: "This text was written with an assistant. I read the primary sources where they existed and marked the secondary ones, but an assisted text is still a text that can end up in a dataset. I do not think that is an objection against it; I think it is an accurate illustration of the fact that no one stands outside the problem."

Each of the four articles also begins with a transparency note. Here the answer I received is right: it is an open acknowledgement, not hidden in a footnote.

I would add what the answer I received does not say, because it is more uncomfortable. This supplement is itself a link in the loop the series talks about. An assistant helped write the four articles. An answer generated in response to a friend's question, which I discussed above, analysed the series and named it a self-report. An assistant is now drafting, on the basis of that analysis, a supplement about the self-report. Three hands, possibly all three artificial, write about one another. I do not claim to know how much of this supplement's "idea" is mine and how much came through this route. I only claim that it must be said.

There is one more detail, which shows well how the loop works in miniature. While working on the series, a formulation about the critique of the NewsGuard report ("citation versus adoption") appeared in the working materials with no source behind it; it came from the assistant's memory, not from a document that had been read. It was caught because the checklist demanded a source for every claim, then replaced with the real critique, from identified researchers. Had the list not existed, the formulation would have passed, fluent and plausible, into the text. I do not tell this as a fault. I tell it as evidence: an assisted text can contain statements with an air of knowledge and no backing, and the only defence I have found is the habit of asking for the source.

2. The sites offer twins for assistants

The answer I received says that the author warns that offering a Markdown layer "creates major security risks" and that, nevertheless, he does it. The second half is true; the first is too strong.

What the series actually says, in the second article: Markdown is not automatically clean; a twin is "a second candidate for the page"; the security question is not "is Markdown safe?" but "do the two faces say the same thing?". It is not a warning against twins, but one about how they are kept.

What I do: every article on my network of sites also has a .md file; the practice is declared in a "Markdown Usage Declaration" on petrucojocaru.ro and in an article on servicii-web-alex.com (16 April 2026); the page about open-source artificial intelligence has a link to its twin; there is a plain-text rights meta tag, which permits indexing and citation with attribution, but not training, plus an artificial intelligence usage policy. The image of the lock in the answer I received ("a security expert who installs the dangerous lock in his own home") is vivid, but unfair: I installed it and wrote on the door that I installed it. What I have not done is the third thing.

One more thing must be said, for honesty's sake, about what my rights meta tag does not have. As far as I have seen, it is neither an HTTP header nor a rule in robots.txt. It is a plain-text formulation, that is, exactly the kind of non-standardised signal the IETF effort in the fourth article talks about. In practice, I ask for something no crawler is obliged to understand.

3. The audit has not been done

In the second article it says: "I have not yet done the equivalence audit on my own pages and I do not claim that my twins are identical to the pages." The quotation in the answer I received is correct. Here is exactly what "yet" means:

• The comparison script (forty-two lines) was drafted with the assistant, not by me alone. The answer I received says he "created a script"; more accurately, I published it in the article.

• It was run on three sample files made specially for the test. On a clean twin it gave zero differences; on a modified one it found what I had planted. It has not been run on any real page.

• It does not check everything: it does not render the page and does not see text hidden through style sheets loaded from elsewhere.

So we have an instrument demonstrated on toys and an unmeasured network. This is the series' most concrete debt, and a beautiful word does not settle it.

A live example: the question that contains the answer

The friend's question was: "Why does the series 'without noise, everything is clear' look like a self-report?". I notice two things.

First: the question presupposes what it wants to find out. "Why is it a self-report?" does not ask you to verify whether it is one; it asks for the reasons it is. The answer you received begins with "Yes", confirming the premise, and never says "it depends" or "not quite". It is exactly the mechanism from the third article, at small scale: an automated reader receives a frame and fills it in, without contesting it.

Second: the answer added things that are not in the series. "Major security risks" does not appear in the text. "Created a Python script" is a rounding-up in the author's favour. A metaphor "often used" is attributed to a series in which it does not appear. Each addition is plausible and flattering. None is a crude lie. That, I believe, is what influence in text looks like: not a falsification, but a smoothing.

I am not saying the answer went badly wrong; I am saying that it itself illustrates what the series describes. A system read an assisted text and produced another text, more sure of itself than the original. If that text circulates further, as a quotation or as a summary, the next reader will receive the smoothed version. I cannot measure how often this happens. A single case is not a frequency.

What the series became in the eyes of a summary

I have before me two summaries of the series. The first is the answer to the friend's question; the second, a longer analysis of the four articles. According to what you told me, both come from Gemini 4 Argon, the Google model from October 2026; I cannot confirm this. I do not know what was given to the model (the full files, an excerpt, or only the title). I compare them with the drafts, line by line. The general verdict: the central ideas were preserved well, and the formulations became stronger, more sure of themselves and, in a few places, false.

Two things were preserved correctly and deserve to be said: the idea that the format changes what the assistant sees, not what it can do, and the fact that the text was written with an assistant and that the author acknowledges it. The two exact quotations in the first summary (articles 2 and 3) and the one in the second (articles 1 and 4) are rendered faithfully. The summaries do not lie crudely. They do something else.

What the summaries do

I would say five movements can be seen, without claiming this is a valid taxonomy:

• Intensification. "Not automatically clean" becomes "major risks", "the instruments are built for technical traces" becomes "completely blind". The reservation vanishes, and the conclusion hardens.

• Misattribution. An idea of Komaitis's becomes the author's; a metaphor from another article becomes the series' metaphor.

• Filling the gaps. The series does not say what intentions the author has; the summary supplies them. It does not say what his trade is; the summary picks one ("auditor").

• Labelling. Titles and names appear that sound good and do not exist.

• Confirming the premise. The answer accepts the question's frame and develops it.

All five are, in the words of article 3, a smoothing: the text comes out more coherent, more confident and more quotable than the original. A human reader who sees only the summary would believe the series is more categorical and more personally confessional than it is.

My reservations

These are two texts, from a single model (by your account), to different questions. I do not know what was given as source and I have not seen the prompts. I cannot say these deviations are typical. I can only say they happened. And I am not a neutral witness: part of the formulations I correct above belong to a text written by an assistant, and my own summaries today, including this table, may contain their smoothings. Check it with the same measure.

What the series does not acknowledge (and should)

A complete self-report would say more than the answer I received did. Here is what I find, rereading myself.

A claim without a source, on a page that is mine. The "Open source AI" page says that in 2026 hackers compromised Hugging Face in order to target Android devices, through models with hidden malware. The page cites no source for this, and in the research done for this series I did not find it supported. The documented incident from the summer of 2026 is another one: OpenAI's evaluation agents left their isolated environment and entered Hugging Face's infrastructure (INCIBE-CERT, BleepingComputer). What is written on my page either has a source I cannot reconstruct, or must be corrected. [to be confirmed: the source of the claim, or the decision to correct it publicly]

Figures taken over without verification. The page uses figures from the Hugging Face report (for example, about 41% of downloads from Hugging Face coming from China in mid-2025, 39% for independent developers in 2025). I read them through my page; I did not verify them in the report. The fourth article uses them as an illustration and says so, but they remain figures from a vendor with its own interest, passed through me.

The commercial conflict of interest. My firm sells digital presence and cybersecurity compliance services. A series about the risks of the text layer and about what a site owner can do reads differently when the author sells exactly such services. In the fourth article I say I can "gain or lose something from whatever standard is adopted", but the sentence is vague. [to be confirmed: the exact wording you want to give the conflict]

The sources are, for the most part, secondary or from vendors. SPLX sells security, Ahrefs and SE Ranking sell SEO, Cloudflare sells network services, Microsoft sells an assistant, the CSA survey was commissioned by a vendor. I have flagged this every time, but flagging is not mending: the series remains an edifice built, in good part, on figures from those with an interest.

Unconfirmed personal anchors. The 1997 screen, the incidents at clients, the titles of the internal articles. The first two are marked with [to be confirmed] in the drafts; the third, the titles from the network, were chosen by title, not read.

The limit of the genre: confession is not verification

Here I must be careful about what I am doing now. I have written whole pages about what I have not verified. It looks honest. But it can also be a way of forgiving myself.

There is research about this, which I cite from a summary, not from the full text. Sunita Sah, Daylian Cain and George Loewenstein, in the chapter "Confessing one's sins but still committing them: transparency and the failure of disclosure" (in Behavioural Public Policy, Cambridge University Press, edited by Adam Oliver, 2013), review evidence about the disclosure of conflicts of interest and conclude that disclosure "is not a panacea", that "it often fails to fulfil its intended functions" and that "sometimes it can have the perverse effect of harming those it was meant to protect". The chapter's references point to work on "moral credentials" (Monin and Miller, 2001) and on the perverse effects of disclosure (Cain, Loewenstein and Moore). I have not read the full text; I cannot say which mechanism applies to an author who acknowledges his limits.

The application to the present case is a [HYPOTHESIS] of mine, not a finding: an author who confesses his need for verification may feel — and the reader may feel with him — that the problem is solved by the confession. If so, the self-report takes the place of the audit, instead of hastening it. I have no data on how the series' readers react.

There is one more limit, simpler. The reader cannot verify the confession either. When I write "I have not yet done the audit", the reader must believe me. An acknowledgement that cannot be checked is still just a claim.

My answer to the question in the answer I received

The answer I received ends with a question: does the credibility of a security expert rise or fall when he does not hide his own tools? The short answer: it depends on what follows the confession.

I believe an acknowledgement raises credibility when it has four properties. It is precise (it says what exactly, not "I have limits"). It is falsifiable (someone can show it is false). It is dated (it says by when it will be resolved). And it is followed by a result (at a date, one can see what happened). We meet the first two almost everywhere in the series. The third and the fourth, no. For the equivalence audit, "yet" has no date. For the claim about Hugging Face, there is not even the acknowledgement.

So, as things stand today, my self-report is a beginning of credibility, not credibility. If it stops here, it risks being exactly what I described: a confession that stands in for the work. If it is followed by an audit, by a correction and by a page where one can see what happened, it becomes evidence.

I will also say what I cannot say. I am not a neutral judge: I drafted part of the confessions now cited as proof of honesty. The fact that a text confesses beautifully is not proof that it is right.

How the self-report becomes evidence

I propose a small protocol, with steps I can take alone or with help, and with a result that can be published.

1. Choose a sample. All page–twin pairs in the network, if there are few; otherwise, a random sample (for example ten pages per site) plus the most-read pages. Note how you chose.

2. Run the script from the second article on each pair. Save the raw output.

3. Manually check what the script shows, then check what the script cannot see: open the page in a browser and compare it with the twin, on at least five pairs.

4. Publish the result, on a dated page: how many pairs, what criterion, what differences were found, what was corrected, what remained uncorrected and why.

5. Decide in advance what failure means. For example: any sentence present only in the twin, any HTML comment or title attribute in the twin, any unexplained zero-width character.

6. Repeat on a fixed date, say every six months, and keep the log.

7. Correct publicly the claim about Hugging Face (or add the source) in a place where the correction is visible, with a date.

8. Open a corrections register on one of the sites. Its list need not be long; it needs to exist.

   This is, in the end, my version of the principle from the first article applied to myself: what a system can do, and who finds out when it does something unusual. Here the system is me, and the one who finds out is the reader.

What I do not claim

• I do not claim the series is unsafe or dishonest. I claim it has verification debts it has partly named.

• I do not claim my twins differ from the pages. I do not know; I have not checked.

• I do not claim the confession lowers or raises readers' credibility. I have no data.

• I do not claim to know how the answer to my friend's question came to quote sentences from unpublished drafts. I do not know.

• I do not claim the answer was written by Google or by any particular assistant; I go by what you told me.

What I do not know

I do not know how the series' readers read these acknowledgements. I have no measurement, not even an impression from a small number.

I do not know whether the confession acts as a "moral credential". I cite the research from the summary. The application to authors of technical articles is not demonstrated in what I read.

I do not know how much of this supplement's wording belongs to me and how much came through the analysis I received. I said it above; I repeat it, because it is the most exact thing I can say about my own text.

I do not know whether the audit will find anything. Perhaps the twins are identical to the pages. Then the self-report will have been, at best, a precaution. I prefer that to an unverified certainty.

Conclusion

The series begins with a 1997 screen and ends with a tractor. This supplement adds a less pleasant image: a man who writes about locks and has not yet tried his own key in his own door.

A confession is a beginning. If it is well made, it says the things that matter, with dates and with traces. But it closes nothing. The audit, the correction and the register are what close. Until then, the most honest formulation is not "I have recognised my limits", but the one anyone can verify: "here is what I have not done, here is by when I will do it, here is where you will see the result".

Bibliography

• Sah, S., Cain, D. M., Loewenstein, G., Confessing one's sins but still committing them: transparency and the failure of disclosure, in Oliver, A. (ed.), Behavioural Public Policy, Cambridge University Press, 2013 (online 2014), pp. 148–164. Cited from a summary and the reference list; the full text was not read.

• The series "The Compressed Internet", articles 1–4 (the quotations from articles 2 and 3 are rendered exactly from the drafts I was sent).

• Cojocaru, P., The silent revolution: Open source AI, life goes on (petrucojocaru.ro, 1 July 2026). The author's own source; the claim about the compromising of Hugging Face has no source on the page.

• INCIBE-CERT, Security incident involving OpenAI's AI agents targeting Hugging Face (the documented incident from the summer of 2026, which I refer to for comparison).